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         unsigned long i;
 640         int ret;
 641
 642         if (!sev_es_guest(kvm))
 643                 return -ENOTTY;
 644
 645         kvm_for_each_vcpu(i, vcpu, kvm) {
 646                 ret = mutex_lock_killable(&vcpu->mutex);
 647                 if (ret)
 648                         return ret;
 649
 650                 ret = __sev_launch_update_vmsa(kvm, vcpu, &argp->error);
 651
 652                 mutex_unlock(&vcpu->mutex);
 653                 if (ret)
 654                         return ret;
 655         }
 656
 657         return 0;
 658 }
 659
 660 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 661 {
 662         void __user *measure = (void __user *)(uintptr_t)argp->data;
 663         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 664         struct sev_data_launch_measure data;
 665         struct kvm_sev_launch_measure params;
 666         void __user *p = NULL;
 667         void *blob = NULL;
 668         int ret;
 669
 670         if (!sev_guest(kvm))
 671                 return -ENOTTY;
 672
 673         if (copy_from_user(&params, measure, sizeof(params)))
 674                 return -EFAULT;
 675
 676         memset(&data, 0, sizeof(data));
 677
 678         /* User wants to query the blob length */
 679         if (!params.len)
 680                 goto cmd;
 681
 682         p = (void __user *)(uintptr_t)params.uaddr;
 683         if (p) {
 684                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
 685                         return -EINVAL;
 686
 687                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
 688                 if (!blob)
 689                         return -ENOMEM;
 690
 691                 data.address = __psp_pa(blob);
 692                 data.len = params.len;
 693         }
 694
 695 cmd:
 696         data.handle = sev->handle;
 697         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, &data, &argp->error);
 698
 699         /*
 700          * If we query the session length, FW responded with expected data.
 701          */
 702         if (!params.len)
 703                 goto done;
 704
 705         if (ret)
 706                 goto e_free_blob;
 707
 708         if (blob) {
 709                 if (copy_to_user(p, blob, params.len))
 710                         ret = -EFAULT;
 711         }
 712
 713 done:
 714         params.len = data.len;
 715         if (copy_to_user(measure, &params, sizeof(params)))
 716                 ret = -EFAULT;
 717 e_free_blob:
 718         kfree(blob);
 719         return ret;
 720 }
 721
 722 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 723 {
 724         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 725         struct sev_data_launch_finish data;
 726
 727         if (!sev_guest(kvm))
 728                 return -ENOTTY;
 729
 730         data.handle = sev->handle;
 731         return sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, &data, &argp->error);
 732 }
 733
 734 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 735 {
 736         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 737         struct kvm_sev_guest_status params;
 738         struct sev_data_guest_status data;
 739         int ret;
 740
 741         if (!sev_guest(kvm))
 742                 return -ENOTTY;
 743
 744         memset(&data, 0, sizeof(data));
 745
 746         data.handle = sev->handle;
 747         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, &data, &argp->error);
 748         if (ret)
 749                 return ret;
 750
 751         params.policy = data.policy;
 752         params.state = data.state;
 753         params.handle = data.handle;
 754
 755         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 756                 ret = -EFAULT;
 757
 758         return ret;
 759 }
 760
 761 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 762                                unsigned long dst, int size,
 763                                int *error, bool enc)
 764 {
 765         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 766         struct sev_data_dbg data;
 767
 768         data.reserved = 0;
 769         data.handle = sev->handle;
 770         data.dst_addr = dst;
 771         data.src_addr = src;
 772         data.len = size;
 773
 774         return sev_issue_cmd(kvm,
 775                              enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 776                              &data, error);
 777 }
 778
 779 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 780                              unsigned long dst_paddr, int sz, int *err)
 781 {
 782         int offset;
 783
 784         /*
 785          * Its safe to read more than we are asked, caller should ensure that
 786          * destination has enough space.
 787          */
 788         offset = src_paddr & 15;
 789         src_paddr = round_down(src_paddr, 16);
 790         sz = round_up(sz + offset, 16);
 791
 792         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 793 }
 794
 795 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 796                                   void __user *dst_uaddr,
 797                                   unsigned long dst_paddr,
 798                                   int size, int *err)
 799 {
 800         struct page *tpage = NULL;
 801         int ret, offset;
 802
 803         /* if inputs are not 16-byte then use intermediate buffer */
 804         if (!IS_ALIGNED(dst_paddr, 16) ||
 805             !IS_ALIGNED(paddr,     16) ||
 806             !IS_ALIGNED(size,      16)) {
 807                 tpage = (void *)alloc_page(GFP_KERNEL);
 808                 if (!tpage)
 809                         return -ENOMEM;
 810
 811                 dst_paddr = __sme_page_pa(tpage);
 812         }
 813
 814         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 815         if (ret)
 816                 goto e_free;
 817
 818         if (tpage) {
 819                 offset = paddr & 15;
 820                 if (copy_to_user(dst_uaddr, page_address(tpage) + offset, size))
 821                         ret = -EFAULT;
 822         }
 823
 824 e_free:
 825         if (tpage)
 826                 __free_page(tpage);
 827
 828         return ret;
 829 }
 830
 831 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 832                                   void __user *vaddr,
 833                                   unsigned long dst_paddr,
 834                                   void __user *dst_vaddr,
 835                                   int size, int *error)
 836 {
 837         struct page *src_tpage = NULL;
 838         struct page *dst_tpage = NULL;
 839         int ret, len = size;
 840
 841         /* If source buffer is not aligned then use an intermediate buffer */
 842         if (!IS_ALIGNED((unsigned long)vaddr, 16)) {
 843                 src_tpage = alloc_page(GFP_KERNEL);
 844                 if (!src_tpage)
 845                         return -ENOMEM;
 846
 847                 if (copy_from_user(page_address(src_tpage), vaddr, size)) {
 848                         __free_page(src_tpage);
 849                         return -EFAULT;
 850                 }
 851
 852                 paddr = __sme_page_pa(src_tpage);
 853         }
 854
 855         /*
 856          *  If destination buffer or length is not aligned then do read-modify-write:
 857          *   - decrypt destination in an intermediate buffer
 858          *   - copy the source buffer in an intermediate buffer
 859          *   - use the intermediate buffer as source buffer
 860          */
 861         if (!IS_ALIGNED((unsigned long)dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 862                 int dst_offset;
 863
 864                 dst_tpage = alloc_page(GFP_KERNEL);
 865                 if (!dst_tpage) {
 866                         ret = -ENOMEM;
 867                         goto e_free;
 868                 }
 869
 870                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 871                                         __sme_page_pa(dst_tpage), size, error);
 872                 if (ret)
 873                         goto e_free;
 874
 875                 /*
 876                  *  If source is kernel buffer then use memcpy() otherwise
 877                  *  copy_from_user().
 878                  */
 879                 dst_offset = dst_paddr & 15;
 880
 881                 if (src_tpage)
 882                         memcpy(page_address(dst_tpage) + dst_offset,
 883                                page_address(src_tpage), size);
 884                 else {
 885                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 886                                            vaddr, size)) {
 887                                 ret = -EFAULT;
 888                                 goto e_free;
 889                         }
 890                 }
 891
 892                 paddr = __sme_page_pa(dst_tpage);
 893                 dst_paddr = round_down(dst_paddr, 16);
 894                 len = round_up(size, 16);
 895         }
 896
 897         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 898
 899 e_free:
 900         if (src_tpage)
 901                 __free_page(src_tpage);
 902         if (dst_tpage)
 903                 __free_page(dst_tpage);
 904         return ret;
 905 }
 906
 907 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 908 {
 909         unsigned long vaddr, vaddr_end, next_vaddr;
 910         unsigned long dst_vaddr;
 911         struct page **src_p, **dst_p;
 912         struct kvm_sev_dbg debug;
 913         unsigned long n;
 914         unsigned int size;
 915         int ret;
 916
 917         if (!sev_guest(kvm))
 918                 return -ENOTTY;
 919
 920         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 921                 return -EFAULT;
 922
 923         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 924                 return -EINVAL;
 925         if (!debug.dst_uaddr)
 926                 return -EINVAL;
 927
 928         vaddr = debug.src_uaddr;
 929         size = debug.len;
 930         vaddr_end = vaddr + size;
 931         dst_vaddr = debug.dst_uaddr;
 932
 933         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 934                 int len, s_off, d_off;
 935
 936                 /* lock userspace source and destination page */
 937                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 938                 if (IS_ERR(src_p))
 939                         return PTR_ERR(src_p);
 940
 941                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 942                 if (IS_ERR(dst_p)) {
 943                         sev_unpin_memory(kvm, src_p, n);
 944                         return PTR_ERR(dst_p);
 945                 }
 946
 947                 /*
 948                  * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
 949                  * the pages; flush the destination too so that future accesses do not
 950                  * see stale data.
 951                  */
 952                 sev_clflush_pages(src_p, 1);
 953                 sev_clflush_pages(dst_p, 1);
 954
 955                 /*
 956                  * Since user buffer may not be page aligned, calculate the
 957                  * offset within the page.
 958                  */
 959                 s_off = vaddr & ~PAGE_MASK;
 960                 d_off = dst_vaddr & ~PAGE_MASK;
 961                 len = min_t(size_t, (PAGE_SIZE - s_off), size);
 962
 963                 if (dec)
 964                         ret = __sev_dbg_decrypt_user(kvm,
 965                                                      __sme_page_pa(src_p[0]) + s_off,
 966                                                      (void __user *)dst_vaddr,
 967                                                      __sme_page_pa(dst_p[0]) + d_off,
 968                                                      len, &argp->error);
 969                 else
 970                         ret = __sev_dbg_encrypt_user(kvm,
 971                                                      __sme_page_pa(src_p[0]) + s_off,
 972                                                      (void __user *)vaddr,
 973                                                      __sme_page_pa(dst_p[0]) + d_off,
 974                                                      (void __user *)dst_vaddr,
 975                                                      len, &argp->error);
 976
 977                 sev_unpin_memory(kvm, src_p, n);
 978                 sev_unpin_memory(kvm, dst_p, n);
 979
 980                 if (ret)
 981                         goto err;
 982
 983                 next_vaddr = vaddr + len;
 984                 dst_vaddr = dst_vaddr + len;
 985                 size -= len;
 986         }
 987 err:
 988         return ret;
 989 }
 990
 991 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
 992 {
 993         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 994         struct sev_data_launch_secret data;
 995         struct kvm_sev_launch_secret params;
 996         struct page **pages;
 997         void *blob, *hdr;
 998         unsigned long n, i;
 999         int ret, offset;
1000
1001         if (!sev_guest(kvm))
1002                 return -ENOTTY;
1003
1004         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1005                 return -EFAULT;
1006
1007         pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, 1);
1008         if (IS_ERR(pages))
1009                 return PTR_ERR(pages);
1010
1011         /*
1012          * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
1013          * place; the cache may contain the data that was written unencrypted.
1014          */
1015         sev_clflush_pages(pages, n);
1016
1017         /*
1018          * The secret must be copied into contiguous memory region, lets verify
1019          * that userspace memory pages are contiguous before we issue command.
1020          */
1021         if (get_num_contig_pages(0, pages, n) != n) {
1022                 ret = -EINVAL;
1023                 goto e_unpin_memory;
1024         }
1025
1026         memset(&data, 0, sizeof(data));
1027
1028         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1029         data.guest_address = __sme_page_pa(pages[0]) + offset;
1030         data.guest_len = params.guest_len;
1031
1032         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1033         if (IS_ERR(blob)) {
1034                 ret = PTR_ERR(blob);
1035                 goto e_unpin_memory;
1036         }
1037
1038         data.trans_address = __psp_pa(blob);
1039         data.trans_len = params.trans_len;
1040
1041         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1042         if (IS_ERR(hdr)) {
1043                 ret = PTR_ERR(hdr);
1044                 goto e_free_blob;
1045         }
1046         data.hdr_address = __psp_pa(hdr);
1047         data.hdr_len = params.hdr_len;
1048
1049         data.handle = sev->handle;
1050         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, &data, &argp->error);
1051
1052         kfree(hdr);
1053
1054 e_free_blob:
1055         kfree(blob);
1056 e_unpin_memory:
1057         /* content of memory is updated, mark pages dirty */
1058         for (i = 0; i < n; i++) {
1059                 set_page_dirty_lock(pages[i]);
1060                 mark_page_accessed(pages[i]);
1061         }
1062         sev_unpin_memory(kvm, pages, n);
1063         return ret;
1064 }
1065
1066 static int sev_get_attestation_report(struct kvm *kvm, struct kvm_sev_cmd *argp)
1067 {
1068         void __user *report = (void __user *)(uintptr_t)argp->data;
1069         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1070         struct sev_data_attestation_report data;
1071         struct kvm_sev_attestation_report params;
1072         void __user *p;
1073         void *blob = NULL;
1074         int ret;
1075
1076         if (!sev_guest(kvm))
1077                 return -ENOTTY;
1078
1079         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1080                 return -EFAULT;
1081
1082         memset(&data, 0, sizeof(data));
1083
1084         /* User wants to query the blob length */
1085         if (!params.len)
1086                 goto cmd;
1087
1088         p = (void __user *)(uintptr_t)params.uaddr;
1089         if (p) {
1090                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
1091                         return -EINVAL;
1092
1093                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
1094                 if (!blob)
1095                         return -ENOMEM;
1096
1097                 data.address = __psp_pa(blob);
1098                 data.len = params.len;
1099                 memcpy(data.mnonce, params.mnonce, sizeof(params.mnonce));
1100         }
1101 cmd:
1102         data.handle = sev->handle;
1103         ret = sev_issue_cmd(kvm, SEV_CMD_ATTESTATION_REPORT, &data, &argp->error);
1104         /*
1105          * If we query the session length, FW responded with expected data.
1106          */
1107         if (!params.len)
1108                 goto done;
1109
1110         if (ret)
1111                 goto e_free_blob;
1112
1113         if (blob) {
1114                 if (copy_to_user(p, blob, params.len))
1115                         ret = -EFAULT;
1116         }
1117
1118 done:
1119         params.len = data.len;
1120         if (copy_to_user(report, &params, sizeof(params)))
1121                 ret = -EFAULT;
1122 e_free_blob:
1123         kfree(blob);
1124         return ret;
1125 }
1126
1127 /* Userspace wants to query session length. */
1128 static int
1129 __sev_send_start_query_session_length(struct kvm *kvm, struct kvm_sev_cmd *argp,
1130                                       struct kvm_sev_send_start *params)
1131 {
1132         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1133         struct sev_data_send_start data;
1134         int ret;
1135
1136         memset(&data, 0, sizeof(data));
1137         data.handle = sev->handle;
1138         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1139
1140         params->session_len = data.session_len;
1141         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1142                                 sizeof(struct kvm_sev_send_start)))
1143                 ret = -EFAULT;
1144
1145         return ret;
1146 }
1147
1148 static int sev_send_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1149 {
1150         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1151         struct sev_data_send_start data;
1152         struct kvm_sev_send_start params;
1153         void *amd_certs, *session_data;
1154         void *pdh_cert, *plat_certs;
1155         int ret;
1156
1157         if (!sev_guest(kvm))
1158                 return -ENOTTY;
1159
1160         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1161                                 sizeof(struct kvm_sev_send_start)))
1162                 return -EFAULT;
1163
1164         /* if session_len is zero, userspace wants to query the session length */
1165         if (!params.session_len)
1166                 return __sev_send_start_query_session_length(kvm, argp,
1167                                 &params);
1168
1169         /* some sanity checks */
1170         if (!params.pdh_cert_uaddr || !params.pdh_cert_len ||
1171             !params.session_uaddr || params.session_len > SEV_FW_BLOB_MAX_SIZE)
1172                 return -EINVAL;
1173
1174         /* allocate the memory to hold the session data blob */
1175         session_data = kmalloc(params.session_len, GFP_KERNEL_ACCOUNT);
1176         if (!session_data)
1177                 return -ENOMEM;
1178
1179         /* copy the certificate blobs from userspace */
1180         pdh_cert = psp_copy_user_blob(params.pdh_cert_uaddr,
1181                                 params.pdh_cert_len);
1182         if (IS_ERR(pdh_cert)) {
1183                 ret = PTR_ERR(pdh_cert);
1184                 goto e_free_session;
1185         }
1186
1187         plat_certs = psp_copy_user_blob(params.plat_certs_uaddr,
1188                                 params.plat_certs_len);
1189         if (IS_ERR(plat_certs)) {
1190                 ret = PTR_ERR(plat_certs);
1191                 goto e_free_pdh;
1192         }
1193
1194         amd_certs = psp_copy_user_blob(params.amd_certs_uaddr,
1195                                 params.amd_certs_len);
1196         if (IS_ERR(amd_certs)) {
1197                 ret = PTR_ERR(amd_certs);
1198                 goto e_free_plat_cert;
1199         }
1200
1201         /* populate the FW SEND_START field with system physical address */
1202         memset(&data, 0, sizeof(data));
1203         data.pdh_cert_address = __psp_pa(pdh_cert);
1204         data.pdh_cert_len = params.pdh_cert_len;
1205         data.plat_certs_address = __psp_pa(plat_certs);
1206         data.plat_certs_len = params.plat_certs_len;
1207         data.amd_certs_address = __psp_pa(amd_certs);
1208         data.amd_certs_len = params.amd_certs_len;
1209         data.session_address = __psp_pa(session_data);
1210         data.session_len = params.session_len;
1211         data.handle = sev->handle;
1212
1213         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1214
1215         if (!ret && copy_to_user((void __user *)(uintptr_t)params.session_uaddr,
1216                         session_data, params.session_len)) {
1217                 ret = -EFAULT;
1218                 goto e_free_amd_cert;
1219         }
1220
1221         params.policy = data.policy;
1222         params.session_len = data.session_len;
1223         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params,
1224                                 sizeof(struct kvm_sev_send_start)))
1225                 ret = -EFAULT;
1226
1227 e_free_amd_cert:
1228         kfree(amd_certs);
1229 e_free_plat_cert:
1230         kfree(plat_certs);
1231 e_free_pdh:
1232         kfree(pdh_cert);
1233 e_free_session:
1234         kfree(session_data);
1235         return ret;
1236 }
1237
1238 /* Userspace wants to query either header or trans length. */
1239 static int
1240 __sev_send_update_data_query_lengths(struct kvm *kvm, struct kvm_sev_cmd *argp,
1241                                      struct kvm_sev_send_update_data *params)
1242 {
1243         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1244         struct sev_data_send_update_data data;
1245         int ret;
1246
1247         memset(&data, 0, sizeof(data));
1248         data.handle = sev->handle;
1249         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1250
1251         params->hdr_len = data.hdr_len;
1252         params->trans_len = data.trans_len;
1253
1254         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1255                          sizeof(struct kvm_sev_send_update_data)))
1256                 ret = -EFAULT;
1257
1258         return ret;
1259 }
1260
1261 static int sev_send_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1262 {
1263         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1264         struct sev_data_send_update_data data;
1265         struct kvm_sev_send_update_data params;
1266         void *hdr, *trans_data;
1267         struct page **guest_page;
1268         unsigned long n;
1269         int ret, offset;
1270
1271         if (!sev_guest(kvm))
1272                 return -ENOTTY;
1273
1274         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1275                         sizeof(struct kvm_sev_send_update_data)))
1276                 return -EFAULT;
1277
1278         /* userspace wants to query either header or trans length */
1279         if (!params.trans_len || !params.hdr_len)
1280                 return __sev_send_update_data_query_lengths(kvm, argp, &params);
1281
1282         if (!params.trans_uaddr || !params.guest_uaddr ||
1283             !params.guest_len || !params.hdr_uaddr)
1284                 return -EINVAL;
1285
1286         /* Check if we are crossing the page boundary */
1287         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1288         if ((params.guest_len + offset > PAGE_SIZE))
1289                 return -EINVAL;
1290
1291         /* Pin guest memory */
1292         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1293                                     PAGE_SIZE, &n, 0);
1294         if (IS_ERR(guest_page))
1295                 return PTR_ERR(guest_page);
1296
1297         /* allocate memory for header and transport buffer */
1298         ret = -ENOMEM;
1299         hdr = kmalloc(params.hdr_len, GFP_KERNEL_ACCOUNT);
1300         if (!hdr)
1301                 goto e_unpin;
1302
1303         trans_data = kmalloc(params.trans_len, GFP_KERNEL_ACCOUNT);
1304         if (!trans_data)
1305                 goto e_free_hdr;
1306
1307         memset(&data, 0, sizeof(data));
1308         data.hdr_address = __psp_pa(hdr);
1309         data.hdr_len = params.hdr_len;
1310         data.trans_address = __psp_pa(trans_data);
1311         data.trans_len = params.trans_len;
1312
1313         /* The SEND_UPDATE_DATA command requires C-bit to be always set. */
1314         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1315         data.guest_address |= sev_me_mask;
1316         data.guest_len = params.guest_len;
1317         data.handle = sev->handle;
1318
1319         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1320
1321         if (ret)
1322                 goto e_free_trans_data;
1323
1324         /* copy transport buffer to user space */
1325         if (copy_to_user((void __user *)(uintptr_t)params.trans_uaddr,
1326                          trans_data, params.trans_len)) {
1327                 ret = -EFAULT;
1328                 goto e_free_trans_data;
1329         }
1330
1331         /* Copy packet header to userspace. */
1332         if (copy_to_user((void __user *)(uintptr_t)params.hdr_uaddr, hdr,
1333                          params.hdr_len))
1334                 ret = -EFAULT;
1335
1336 e_free_trans_data:
1337         kfree(trans_data);
1338 e_free_hdr:
1339         kfree(hdr);
1340 e_unpin:
1341         sev_unpin_memory(kvm, guest_page, n);
1342
1343         return ret;
1344 }
1345
1346 static int sev_send_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1347 {
1348         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1349         struct sev_data_send_finish data;
1350
1351         if (!sev_guest(kvm))
1352                 return -ENOTTY;
1353
1354         data.handle = sev->handle;
1355         return sev_issue_cmd(kvm, SEV_CMD_SEND_FINISH, &data, &argp->error);
1356 }
1357
1358 static int sev_send_cancel(struct kvm *kvm, struct kvm_sev_cmd *argp)
1359 {
1360         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1361         struct sev_data_send_cancel data;
1362
1363         if (!sev_guest(kvm))
1364                 return -ENOTTY;
1365
1366         data.handle = sev->handle;
1367         return sev_issue_cmd(kvm, SEV_CMD_SEND_CANCEL, &data, &argp->error);
1368 }
1369
1370 static int sev_receive_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1371 {
1372         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1373         struct sev_data_receive_start start;
1374         struct kvm_sev_receive_start params;
1375         int *error = &argp->error;
1376         void *session_data;
1377         void *pdh_data;
1378         int ret;
1379
1380         if (!sev_guest(kvm))
1381                 return -ENOTTY;
1382
1383         /* Get parameter from the userspace */
1384         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1385                         sizeof(struct kvm_sev_receive_start)))
1386                 return -EFAULT;
1387
1388         /* some sanity checks */
1389         if (!params.pdh_uaddr || !params.pdh_len ||
1390             !params.session_uaddr || !params.session_len)
1391                 return -EINVAL;
1392
1393         pdh_data = psp_copy_user_blob(params.pdh_uaddr, params.pdh_len);
1394         if (IS_ERR(pdh_data))
1395                 return PTR_ERR(pdh_data);
1396
1397         session_data = psp_copy_user_blob(params.session_uaddr,
1398                         params.session_len);
1399         if (IS_ERR(session_data)) {
1400                 ret = PTR_ERR(session_data);
1401                 goto e_free_pdh;
1402         }
1403
1404         memset(&start, 0, sizeof(start));
1405         start.handle = params.handle;
1406         start.policy = params.policy;
1407         start.pdh_cert_address = __psp_pa(pdh_data);
1408         start.pdh_cert_len = params.pdh_len;
1409         start.session_address = __psp_pa(session_data);
1410         start.session_len = params.session_len;
1411
1412         /* create memory encryption context */
1413         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_RECEIVE_START, &start,
1414                                 error);
1415         if (ret)
1416                 goto e_free_session;
1417
1418         /* Bind ASID to this guest */
1419         ret = sev_bind_asid(kvm, start.handle, error);
1420         if (ret) {
1421                 sev_decommission(start.handle);
1422                 goto e_free_session;
1423         }
1424
1425         params.handle = start.handle;
1426         if (copy_to_user((void __user *)(uintptr_t)argp->data,
1427                          &params, sizeof(struct kvm_sev_receive_start))) {
1428                 ret = -EFAULT;
1429                 sev_unbind_asid(kvm, start.handle);
1430                 goto e_free_session;
1431         }
1432
1433         sev->handle = start.handle;
1434         sev->fd = argp->sev_fd;
1435
1436 e_free_session:
1437         kfree(session_data);
1438 e_free_pdh:
1439         kfree(pdh_data);
1440
1441         return ret;
1442 }
1443
1444 static int sev_receive_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1445 {
1446         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1447         struct kvm_sev_receive_update_data params;
1448         struct sev_data_receive_update_data data;
1449         void *hdr = NULL, *trans = NULL;
1450         struct page **guest_page;
1451         unsigned long n;
1452         int ret, offset;
1453
1454         if (!sev_guest(kvm))
1455                 return -EINVAL;
1456
1457         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1458                         sizeof(struct kvm_sev_receive_update_data)))
1459                 return -EFAULT;
1460
1461         if (!params.hdr_uaddr || !params.hdr_len ||
1462             !params.guest_uaddr || !params.guest_len ||
1463             !params.trans_uaddr || !params.trans_len)
1464                 return -EINVAL;
1465
1466         /* Check if we are crossing the page boundary */
1467         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1468         if ((params.guest_len + offset > PAGE_SIZE))
1469                 return -EINVAL;
1470
1471         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1472         if (IS_ERR(hdr))
1473                 return PTR_ERR(hdr);
1474
1475         trans = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1476         if (IS_ERR(trans)) {
1477                 ret = PTR_ERR(trans);
1478                 goto e_free_hdr;
1479         }
1480
1481         memset(&data, 0, sizeof(data));
1482         data.hdr_address = __psp_pa(hdr);
1483         data.hdr_len = params.hdr_len;
1484         data.trans_address = __psp_pa(trans);
1485         data.trans_len = params.trans_len;
1486
1487         /* Pin guest memory */
1488         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1489                                     PAGE_SIZE, &n, 1);
1490         if (IS_ERR(guest_page)) {
1491                 ret = PTR_ERR(guest_page);
1492                 goto e_free_trans;
1493         }
1494
1495         /*
1496          * Flush (on non-coherent CPUs) before RECEIVE_UPDATE_DATA, the PSP
1497          * encrypts the written data with the guest's key, and the cache may
1498          * contain dirty, unencrypted data.
1499          */
1500         sev_clflush_pages(guest_page, n);
1501
1502         /* The RECEIVE_UPDATE_DATA command requires C-bit to be always set. */
1503         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1504         data.guest_address |= sev_me_mask;
1505         data.guest_len = params.guest_len;
1506         data.handle = sev->handle;
1507
1508         ret = sev_issue_cmd(kvm, SEV_CMD_RECEIVE_UPDATE_DATA, &data,
1509                                 &argp->error);
1510
1511         sev_unpin_memory(kvm, guest_page, n);
1512
1513 e_free_trans:
1514         kfree(trans);
1515 e_free_hdr:
1516         kfree(hdr);
1517
1518         return ret;
1519 }
1520
1521 static int sev_receive_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1522 {
1523         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1524         struct sev_data_receive_finish data;
1525
1526         if (!sev_guest(kvm))
1527                 return -ENOTTY;
1528
1529         data.handle = sev->handle;
1530         return sev_issue_cmd(kvm, SEV_CMD_RECEIVE_FINISH, &data, &argp->error);
1531 }
1532
1533 static bool is_cmd_allowed_from_mirror(u32 cmd_id)
1534 {
1535         /*
1536          * Allow mirrors VM to call KVM_SEV_LAUNCH_UPDATE_VMSA to enable SEV-ES
1537          * active mirror VMs. Also allow the debugging and status commands.
1538          */
1539         if (cmd_id == KVM_SEV_LAUNCH_UPDATE_VMSA ||
1540             cmd_id == KVM_SEV_GUEST_STATUS || cmd_id == KVM_SEV_DBG_DECRYPT ||
1541             cmd_id == KVM_SEV_DBG_ENCRYPT)
1542                 return true;
1543
1544         return false;
1545 }
1546
1547 static int sev_lock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1548 {
1549         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1550         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1551         int r = -EBUSY;
1552
1553         if (dst_kvm == src_kvm)
1554                 return -EINVAL;
1555
1556         /*
1557          * Bail if these VMs are already involved in a migration to avoid
1558          * deadlock between two VMs trying to migrate to/from each other.
1559          */
1560         if (atomic_cmpxchg_acquire(&dst_sev->migration_in_progress, 0, 1))
1561                 return -EBUSY;
1562
1563         if (atomic_cmpxchg_acquire(&src_sev->migration_in_progress, 0, 1))
1564                 goto release_dst;
1565
1566         r = -EINTR;
1567         if (mutex_lock_killable(&dst_kvm->lock))
1568                 goto release_src;
1569         if (mutex_lock_killable_nested(&src_kvm->lock, SINGLE_DEPTH_NESTING))
1570                 goto unlock_dst;
1571         return 0;
1572
1573 unlock_dst:
1574         mutex_unlock(&dst_kvm->lock);
1575 release_src:
1576         atomic_set_release(&src_sev->migration_in_progress, 0);
1577 release_dst:
1578         atomic_set_release(&dst_sev->migration_in_progress, 0);
1579         return r;
1580 }
1581
1582 static void sev_unlock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1583 {
1584         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1585         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1586
1587         mutex_unlock(&dst_kvm->lock);
1588         mutex_unlock(&src_kvm->lock);
1589         atomic_set_release(&dst_sev->migration_in_progress, 0);
1590         atomic_set_release(&src_sev->migration_in_progress, 0);
1591 }
1592
1593
1594 static int sev_lock_vcpus_for_migration(struct kvm *kvm)
1595 {
1596         struct kvm_vcpu *vcpu;
1597         unsigned long i, j;
1598
1599         kvm_for_each_vcpu(i, vcpu, kvm) {
1600                 if (mutex_lock_killable(&vcpu->mutex))
1601                         goto out_unlock;
1602         }
1603
1604         return 0;
1605
1606 out_unlock:
1607         kvm_for_each_vcpu(j, vcpu, kvm) {
1608                 if (i == j)
1609                         break;
1610
1611                 mutex_unlock(&vcpu->mutex);
1612         }
1613         return -EINTR;
1614 }
1615
1616 static void sev_unlock_vcpus_for_migration(struct kvm *kvm)
1617 {
1618         struct kvm_vcpu *vcpu;
1619         unsigned long i;
1620
1621         kvm_for_each_vcpu(i, vcpu, kvm) {
1622                 mutex_unlock(&vcpu->mutex);
1623         }
1624 }
1625
1626 static void sev_migrate_from(struct kvm_sev_info *dst,
1627                               struct kvm_sev_info *src)
1628 {
1629         dst->active = true;
1630         dst->asid = src->asid;
1631         dst->handle = src->handle;
1632         dst->pages_locked = src->pages_locked;
1633         dst->enc_context_owner = src->enc_context_owner;
1634
1635         src->asid = 0;
1636         src->active = false;
1637         src->handle = 0;
1638         src->pages_locked = 0;
1639         src->enc_context_owner = NULL;
1640
1641         list_cut_before(&dst->regions_list, &src->regions_list, &src->regions_list);
1642 }
1643
1644 static int sev_es_migrate_from(struct kvm *dst, struct kvm *src)
1645 {
1646         unsigned long i;
1647         struct kvm_vcpu *dst_vcpu, *src_vcpu;
1648         struct vcpu_svm *dst_svm, *src_svm;
1649
1650         if (atomic_read(&src->online_vcpus) != atomic_read(&dst->online_vcpus))
1651                 return -EINVAL;
1652
1653         kvm_for_each_vcpu(i, src_vcpu, src) {
1654                 if (!src_vcpu->arch.guest_state_protected)
1655                         return -EINVAL;
1656         }
1657
1658         kvm_for_each_vcpu(i, src_vcpu, src) {
1659                 src_svm = to_svm(src_vcpu);
1660                 dst_vcpu = kvm_get_vcpu(dst, i);
1661                 dst_svm = to_svm(dst_vcpu);
1662
1663                 /*
1664                  * Transfer VMSA and GHCB state to the destination.  Nullify and
1665                  * clear source fields as appropriate, the state now belongs to
1666                  * the destination.
1667                  */
1668                 memcpy(&dst_svm->sev_es, &src_svm->sev_es, sizeof(src_svm->sev_es));
1669                 dst_svm->vmcb->control.ghcb_gpa = src_svm->vmcb->control.ghcb_gpa;
1670                 dst_svm->vmcb->control.vmsa_pa = src_svm->vmcb->control.vmsa_pa;
1671                 dst_vcpu->arch.guest_state_protected = true;
1672
1673                 memset(&src_svm->sev_es, 0, sizeof(src_svm->sev_es));
1674                 src_svm->vmcb->control.ghcb_gpa = INVALID_PAGE;
1675                 src_svm->vmcb->control.vmsa_pa = INVALID_PAGE;
1676                 src_vcpu->arch.guest_state_protected = false;
1677         }
1678         to_kvm_svm(src)->sev_info.es_active = false;
1679         to_kvm_svm(dst)->sev_info.es_active = true;
1680
1681         return 0;
1682 }
1683
1684 int svm_vm_migrate_from(struct kvm *kvm, unsigned int source_fd)
1685 {
1686         struct kvm_sev_info *dst_sev = &to_kvm_svm(kvm)->sev_info;
1687         struct kvm_sev_info *src_sev, *cg_cleanup_sev;
1688         struct file *source_kvm_file;
1689         struct kvm *source_kvm;
1690         bool charged = false;
1691         int ret;
1692
1693         source_kvm_file = fget(source_fd);
1694         if (!file_is_kvm(source_kvm_file)) {
1695                 ret = -EBADF;
1696                 goto out_fput;
1697         }
1698
1699         source_kvm = source_kvm_file->private_data;
1700         ret = sev_lock_two_vms(kvm, source_kvm);
1701         if (ret)
1702                 goto out_fput;
1703
1704         if (sev_guest(kvm) || !sev_guest(source_kvm)) {
1705                 ret = -EINVAL;
1706                 goto out_unlock;
1707         }
1708
1709         src_sev = &to_kvm_svm(source_kvm)->sev_info;
1710
1711         /*
1712          * VMs mirroring src's encryption context rely on it to keep the
1713          * ASID allocated, but below we are clearing src_sev->asid.
1714          */
1715         if (src_sev->num_mirrored_vms) {
1716                 ret = -EBUSY;
1717                 goto out_unlock;
1718         }
1719
1720         dst_sev->misc_cg = get_current_misc_cg();
1721         cg_cleanup_sev = dst_sev;
1722         if (dst_sev->misc_cg != src_sev->misc_cg) {
1723                 ret = sev_misc_cg_try_charge(dst_sev);
1724                 if (ret)
1725                         goto out_dst_cgroup;
1726                 charged = true;
1727         }
1728
1729         ret = sev_lock_vcpus_for_migration(kvm);
1730         if (ret)
1731                 goto out_dst_cgroup;
1732         ret = sev_lock_vcpus_for_migration(source_kvm);
1733         if (ret)
1734                 goto out_dst_vcpu;
1735
1736         if (sev_es_guest(source_kvm)) {
1737                 ret = sev_es_migrate_from(kvm, source_kvm);
1738                 if (ret)
1739                         goto out_source_vcpu;
1740         }
1741         sev_migrate_from(dst_sev, src_sev);
1742         kvm_vm_dead(source_kvm);
1743         cg_cleanup_sev = src_sev;
1744         ret = 0;
1745
1746 out_source_vcpu:
1747         sev_unlock_vcpus_for_migration(source_kvm);
1748 out_dst_vcpu:
1749         sev_unlock_vcpus_for_migration(kvm);
1750 out_dst_cgroup:
1751         /* Operates on the source on success, on the destination on failure.  */
1752         if (charged)
1753                 sev_misc_cg_uncharge(cg_cleanup_sev);
1754         put_misc_cg(cg_cleanup_sev->misc_cg);
1755         cg_cleanup_sev->misc_cg = NULL;
1756 out_unlock:
1757         sev_unlock_two_vms(kvm, source_kvm);
1758 out_fput:
1759         if (source_kvm_file)
1760                 fput(source_kvm_file);
1761         return ret;
1762 }
1763
1764 int svm_mem_enc_op(struct kvm *kvm, void __user *argp)
1765 {
1766         struct kvm_sev_cmd sev_cmd;
1767         int r;
1768
1769         if (!sev_enabled)
1770                 return -ENOTTY;
1771
1772         if (!argp)
1773                 return 0;
1774
1775         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
1776                 return -EFAULT;
1777
1778         mutex_lock(&kvm->lock);
1779
1780         /* Only the enc_context_owner handles some memory enc operations. */
1781         if (is_mirroring_enc_context(kvm) &&
1782             !is_cmd_allowed_from_mirror(sev_cmd.id)) {
1783                 r = -EINVAL;
1784                 goto out;
1785         }
1786
1787         switch (sev_cmd.id) {
1788         case KVM_SEV_ES_INIT:
1789                 if (!sev_es_enabled) {
1790                         r = -ENOTTY;
1791                         goto out;
1792                 }
1793                 fallthrough;
1794         case KVM_SEV_INIT:
1795                 r = sev_guest_init(kvm, &sev_cmd);
1796                 break;
1797         case KVM_SEV_LAUNCH_START:
1798                 r = sev_launch_start(kvm, &sev_cmd);
1799                 break;
1800         case KVM_SEV_LAUNCH_UPDATE_DATA:
1801                 r = sev_launch_update_data(kvm, &sev_cmd);
1802                 break;
1803         case KVM_SEV_LAUNCH_UPDATE_VMSA:
1804                 r = sev_launch_update_vmsa(kvm, &sev_cmd);
1805                 break;
1806         case KVM_SEV_LAUNCH_MEASURE:
1807                 r = sev_launch_measure(kvm, &sev_cmd);
1808                 break;
1809         case KVM_SEV_LAUNCH_FINISH:
1810                 r = sev_launch_finish(kvm, &sev_cmd);
1811                 break;
1812         case KVM_SEV_GUEST_STATUS:
1813                 r = sev_guest_status(kvm, &sev_cmd);
1814                 break;
1815         case KVM_SEV_DBG_DECRYPT:
1816                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
1817                 break;
1818         case KVM_SEV_DBG_ENCRYPT:
1819                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
1820                 break;
1821         case KVM_SEV_LAUNCH_SECRET:
1822                 r = sev_launch_secret(kvm, &sev_cmd);
1823                 break;
1824         case KVM_SEV_GET_ATTESTATION_REPORT:
1825                 r = sev_get_attestation_report(kvm, &sev_cmd);
1826                 break;
1827         case KVM_SEV_SEND_START:
1828                 r = sev_send_start(kvm, &sev_cmd);
1829                 break;
1830         case KVM_SEV_SEND_UPDATE_DATA:
1831                 r = sev_send_update_data(kvm, &sev_cmd);
1832                 break;
1833         case KVM_SEV_SEND_FINISH:
1834                 r = sev_send_finish(kvm, &sev_cmd);
1835                 break;
1836         case KVM_SEV_SEND_CANCEL:
1837                 r = sev_send_cancel(kvm, &sev_cmd);
1838                 break;
1839         case KVM_SEV_RECEIVE_START:
1840                 r = sev_receive_start(kvm, &sev_cmd);
1841                 break;
1842         case KVM_SEV_RECEIVE_UPDATE_DATA:
1843                 r = sev_receive_update_data(kvm, &sev_cmd);
1844                 break;
1845         case KVM_SEV_RECEIVE_FINISH:
1846                 r = sev_receive_finish(kvm, &sev_cmd);
1847                 break;
1848         default:
1849                 r = -EINVAL;
1850                 goto out;
1851         }
1852
1853         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
1854                 r = -EFAULT;
1855
1856 out:
1857         mutex_unlock(&kvm->lock);
1858         return r;
1859 }
1860
1861 int svm_register_enc_region(struct kvm *kvm,
1862                             struct kvm_enc_region *range)
1863 {
1864         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1865         struct enc_region *region;
1866         int ret = 0;
1867
1868         if (!sev_guest(kvm))
1869                 return -ENOTTY;
1870
1871         /* If kvm is mirroring encryption context it isn't responsible for it */
1872         if (is_mirroring_enc_context(kvm))
1873                 return -EINVAL;
1874
1875         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
1876                 return -EINVAL;
1877
1878         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
1879         if (!region)
1880                 return -ENOMEM;
1881
1882         mutex_lock(&kvm->lock);
1883         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
1884         if (IS_ERR(region->pages)) {
1885                 ret = PTR_ERR(region->pages);
1886                 mutex_unlock(&kvm->lock);
1887                 goto e_free;
1888         }
1889
1890         region->uaddr = range->addr;
1891         region->size = range->size;
1892
1893         list_add_tail(&region->list, &sev->regions_list);
1894         mutex_unlock(&kvm->lock);
1895
1896         /*
1897          * The guest may change the memory encryption attribute from C=0 -> C=1
1898          * or vice versa for this memory range. Lets make sure caches are
1899          * flushed to ensure that guest data gets written into memory with
1900          * correct C-bit.
1901          */
1902         sev_clflush_pages(region->pages, region->npages);
1903
1904         return ret;
1905
1906 e_free:
1907         kfree(region);
1908         return ret;
1909 }
1910
1911 static struct enc_region *
1912 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
1913 {
1914         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1915         struct list_head *head = &sev->regions_list;
1916         struct enc_region *i;
1917
1918         list_for_each_entry(i, head, list) {
1919                 if (i->uaddr == range->addr &&
1920                     i->size == range->size)
1921                         return i;
1922         }
1923
1924         return NULL;
1925 }
1926
1927 static void __unregister_enc_region_locked(struct kvm *kvm,
1928                                            struct enc_region *region)
1929 {
1930         sev_unpin_memory(kvm, region->pages, region->npages);
1931         list_del(&region->list);
1932         kfree(region);
1933 }
1934
1935 int svm_unregister_enc_region(struct kvm *kvm,
1936                               struct kvm_enc_region *range)
1937 {
1938         struct enc_region *region;
1939         int ret;
1940
1941         /* If kvm is mirroring encryption context it isn't responsible for it */
1942         if (is_mirroring_enc_context(kvm))
1943                 return -EINVAL;
1944
1945         mutex_lock(&kvm->lock);
1946
1947         if (!sev_guest(kvm)) {
1948                 ret = -ENOTTY;
1949                 goto failed;
1950         }
1951
1952         region = find_enc_region(kvm, range);
1953         if (!region) {
1954                 ret = -EINVAL;
1955                 goto failed;
1956         }
1957
1958         /*
1959          * Ensure that all guest tagged cache entries are flushed before
1960          * releasing the pages back to the system for use. CLFLUSH will
1961          * not do this, so issue a WBINVD.
1962          */
1963         wbinvd_on_all_cpus();
1964
1965         __unregister_enc_region_locked(kvm, region);
1966
1967         mutex_unlock(&kvm->lock);
1968         return 0;
1969
1970 failed:
1971         mutex_unlock(&kvm->lock);
1972         return ret;
1973 }
1974
1975 int svm_vm_copy_asid_from(struct kvm *kvm, unsigned int source_fd)
1976 {
1977         struct file *source_kvm_file;
1978         struct kvm *source_kvm;
1979         struct kvm_sev_info *source_sev, *mirror_sev;
1980         int ret;
1981
1982         source_kvm_file = fget(source_fd);
1983         if (!file_is_kvm(source_kvm_file)) {
1984                 ret = -EBADF;
1985                 goto e_source_fput;
1986         }
1987
1988         source_kvm = source_kvm_file->private_data;
1989         ret = sev_lock_two_vms(kvm, source_kvm);
1990         if (ret)
1991                 goto e_source_fput;
1992
1993         /*
1994          * Mirrors of mirrors should work, but let's not get silly.  Also
1995          * disallow out-of-band SEV/SEV-ES init if the target is already an
1996          * SEV guest, or if vCPUs have been created.  KVM relies on vCPUs being
1997          * created after SEV/SEV-ES initialization, e.g. to init intercepts.
1998          */
1999         if (sev_guest(kvm) || !sev_guest(source_kvm) ||
2000             is_mirroring_enc_context(source_kvm) || kvm->created_vcpus) {
2001                 ret = -EINVAL;
2002                 goto e_unlock;
2003         }
2004
2005         /*
2006          * The mirror kvm holds an enc_context_owner ref so its asid can't
2007          * disappear until we're done with it
2008          */
2009         source_sev = &to_kvm_svm(source_kvm)->sev_info;
2010         kvm_get_kvm(source_kvm);
2011         source_sev->num_mirrored_vms++;
2012
2013         /* Set enc_context_owner and copy its encryption context over */
2014         mirror_sev = &to_kvm_svm(kvm)->sev_info;
2015         mirror_sev->enc_context_owner = source_kvm;
2016         mirror_sev->active = true;
2017         mirror_sev->asid = source_sev->asid;
2018         mirror_sev->fd = source_sev->fd;
2019         mirror_sev->es_active = source_sev->es_active;
2020         mirror_sev->handle = source_sev->handle;
2021         INIT_LIST_HEAD(&mirror_sev->regions_list);
2022         ret = 0;
2023
2024         /*
2025          * Do not copy ap_jump_table. Since the mirror does not share the same
2026          * KVM contexts as the original, and they may have different
2027          * memory-views.
2028          */
2029
2030 e_unlock:
2031         sev_unlock_two_vms(kvm, source_kvm);
2032 e_source_fput:
2033         if (source_kvm_file)
2034                 fput(source_kvm_file);
2035         return ret;
2036 }
2037
2038 void sev_vm_destroy(struct kvm *kvm)
2039 {
2040         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
2041         struct list_head *head = &sev->regions_list;
2042         struct list_head *pos, *q;
2043
2044         WARN_ON(sev->num_mirrored_vms);
2045
2046         if (!sev_guest(kvm))
2047                 return;
2048
2049         /* If this is a mirror_kvm release the enc_context_owner and skip sev cleanup */
2050         if (is_mirroring_enc_context(kvm)) {
2051                 struct kvm *owner_kvm = sev->enc_context_owner;
2052                 struct kvm_sev_info *owner_sev = &to_kvm_svm(owner_kvm)->sev_info;
2053
2054                 mutex_lock(&owner_kvm->lock);
2055                 if (!WARN_ON(!owner_sev->num_mirrored_vms))
2056                         owner_sev->num_mirrored_vms--;
2057                 mutex_unlock(&owner_kvm->lock);
2058                 kvm_put_kvm(owner_kvm);
2059                 return;
2060         }
2061
2062         /*
2063          * Ensure that all guest tagged cache entries are flushed before
2064          * releasing the pages back to the system for use. CLFLUSH will
2065          * not do this, so issue a WBINVD.
2066          */
2067         wbinvd_on_all_cpus();
2068
2069         /*
2070          * if userspace was terminated before unregistering the memory regions
2071          * then lets unpin all the registered memory.
2072          */
2073         if (!list_empty(head)) {
2074                 list_for_each_safe(pos, q, head) {
2075                         __unregister_enc_region_locked(kvm,
2076                                 list_entry(pos, struct enc_region, list));
2077                         cond_resched();
2078                 }
2079         }
2080
2081         sev_unbind_asid(kvm, sev->handle);
2082         sev_asid_free(sev);
2083 }
2084
2085 void __init sev_set_cpu_caps(void)
2086 {
2087         if (!sev_enabled)
2088                 kvm_cpu_cap_clear(X86_FEATURE_SEV);
2089         if (!sev_es_enabled)
2090                 kvm_cpu_cap_clear(X86_FEATURE_SEV_ES);
2091 }
2092
2093 void __init sev_hardware_setup(void)
2094 {
2095 #ifdef CONFIG_KVM_AMD_SEV
2096         unsigned int eax, ebx, ecx, edx, sev_asid_count, sev_es_asid_count;
2097         bool sev_es_supported = false;
2098         bool sev_supported = false;
2099
2100         if (!sev_enabled || !npt_enabled)
2101                 goto out;
2102
2103         /*
2104          * SEV must obviously be supported in hardware.  Sanity check that the
2105          * CPU supports decode assists, which is mandatory for SEV guests to
2106          * support instruction emulation.
2107          */
2108         if (!boot_cpu_has(X86_FEATURE_SEV) ||
2109             WARN_ON_ONCE(!boot_cpu_has(X86_FEATURE_DECODEASSISTS)))
2110                 goto out;
2111
2112         /* Retrieve SEV CPUID information */
2113         cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
2114
2115         /* Set encryption bit location for SEV-ES guests */
2116         sev_enc_bit = ebx & 0x3f;
2117
2118         /* Maximum number of encrypted guests supported simultaneously */
2119         max_sev_asid = ecx;
2120         if (!max_sev_asid)
2121                 goto out;
2122
2123         /* Minimum ASID value that should be used for SEV guest */
2124         min_sev_asid = edx;
2125         sev_me_mask = 1UL << (ebx & 0x3f);
2126
2127         /*
2128          * Initialize SEV ASID bitmaps. Allocate space for ASID 0 in the bitmap,
2129          * even though it's never used, so that the bitmap is indexed by the
2130          * actual ASID.
2131          */
2132         nr_asids = max_sev_asid + 1;
2133         sev_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2134         if (!sev_asid_bitmap)
2135                 goto out;
2136
2137         sev_reclaim_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2138         if (!sev_reclaim_asid_bitmap) {
2139                 bitmap_free(sev_asid_bitmap);
2140                 sev_asid_bitmap = NULL;
2141                 goto out;
2142         }
2143
2144         sev_asid_count = max_sev_asid - min_sev_asid + 1;
2145         if (misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count))
2146                 goto out;
2147
2148         pr_info("SEV supported: %u ASIDs\n", sev_asid_count);
2149         sev_supported = true;
2150
2151         /* SEV-ES support requested? */
2152         if (!sev_es_enabled)
2153                 goto out;
2154
2155         /* Does the CPU support SEV-ES? */
2156         if (!boot_cpu_has(X86_FEATURE_SEV_ES))
2157                 goto out;
2158
2159         /* Has the system been allocated ASIDs for SEV-ES? */
2160         if (min_sev_asid == 1)
2161                 goto out;
2162
2163         sev_es_asid_count = min_sev_asid - 1;
2164         if (misc_cg_set_capacity(MISC_CG_RES_SEV_ES, sev_es_asid_count))
2165                 goto out;
2166
2167         pr_info("SEV-ES supported: %u ASIDs\n", sev_es_asid_count);
2168         sev_es_supported = true;
2169
2170 out:
2171         sev_enabled = sev_supported;
2172         sev_es_enabled = sev_es_supported;
2173 #endif
2174 }
2175
2176 void sev_hardware_teardown(void)
2177 {
2178         if (!sev_enabled)
2179                 return;
2180
2181         /* No need to take sev_bitmap_lock, all VMs have been destroyed. */
2182         sev_flush_asids(1, max_sev_asid);
2183
2184         bitmap_free(sev_asid_bitmap);
2185         bitmap_free(sev_reclaim_asid_bitmap);
2186
2187         misc_cg_set_capacity(MISC_CG_RES_SEV, 0);
2188         misc_cg_set_capacity(MISC_CG_RES_SEV_ES, 0);
2189 }
2190
2191 int sev_cpu_init(struct svm_cpu_data *sd)
2192 {
2193         if (!sev_enabled)
2194                 return 0;
2195
2196         sd->sev_vmcbs = kcalloc(nr_asids, sizeof(void *), GFP_KERNEL);
2197         if (!sd->sev_vmcbs)
2198                 return -ENOMEM;
2199
2200         return 0;
2201 }
2202
2203 /*
2204  * Pages used by hardware to hold guest encrypted state must be flushed before
2205  * returning them to the system.
2206  */
2207 static void sev_flush_guest_memory(struct vcpu_svm *svm, void *va,
2208                                    unsigned long len)
2209 {
2210         /*
2211          * If hardware enforced cache coherency for encrypted mappings of the
2212          * same physical page is supported, nothing to do.
2213          */
2214         if (boot_cpu_has(X86_FEATURE_SME_COHERENT))
2215                 return;
2216
2217         /*
2218          * If the VM Page Flush MSR is supported, use it to flush the page
2219          * (using the page virtual address and the guest ASID).
2220          */
2221         if (boot_cpu_has(X86_FEATURE_VM_PAGE_FLUSH)) {
2222                 struct kvm_sev_info *sev;
2223                 unsigned long va_start;
2224                 u64 start, stop;
2225
2226                 /* Align start and stop to page boundaries. */
2227                 va_start = (unsigned long)va;
2228                 start = (u64)va_start & PAGE_MASK;
2229                 stop = PAGE_ALIGN((u64)va_start + len);
2230
2231                 if (start < stop) {
2232                         sev = &to_kvm_svm(svm->vcpu.kvm)->sev_info;
2233
2234                         while (start < stop) {
2235                                 wrmsrl(MSR_AMD64_VM_PAGE_FLUSH,
2236                                        start | sev->asid);
2237
2238                                 start += PAGE_SIZE;
2239                         }
2240
2241                         return;
2242                 }
2243
2244                 WARN(1, "Address overflow, using WBINVD\n");
2245         }
2246
2247         /*
2248          * Hardware should always have one of the above features,
2249          * but if not, use WBINVD and issue a warning.
2250          */
2251         WARN_ONCE(1, "Using WBINVD to flush guest memory\n");
2252         wbinvd_on_all_cpus();
2253 }
2254
2255 void sev_free_vcpu(struct kvm_vcpu *vcpu)
2256 {
2257         struct vcpu_svm *svm;
2258
2259         if (!sev_es_guest(vcpu->kvm))
2260                 return;
2261
2262         svm = to_svm(vcpu);
2263
2264         if (vcpu->arch.guest_state_protected)
2265                 sev_flush_guest_memory(svm, svm->sev_es.vmsa, PAGE_SIZE);
2266         __free_page(virt_to_page(svm->sev_es.vmsa));
2267
2268         if (svm->sev_es.ghcb_sa_free)
2269                 kvfree(svm->sev_es.ghcb_sa);
2270 }
2271
2272 static void dump_ghcb(struct vcpu_svm *svm)
2273 {
2274         struct ghcb *ghcb = svm->sev_es.ghcb;
2275         unsigned int nbits;
2276
2277         /* Re-use the dump_invalid_vmcb module parameter */
2278         if (!dump_invalid_vmcb) {
2279                 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
2280                 return;
2281         }
2282
2283         nbits = sizeof(ghcb->save.valid_bitmap) * 8;
2284
2285         pr_err("GHCB (GPA=%016llx):\n", svm->vmcb->control.ghcb_gpa);
2286         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
2287                ghcb->save.sw_exit_code, ghcb_sw_exit_code_is_valid(ghcb));
2288         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
2289                ghcb->save.sw_exit_info_1, ghcb_sw_exit_info_1_is_valid(ghcb));
2290         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
2291                ghcb->save.sw_exit_info_2, ghcb_sw_exit_info_2_is_valid(ghcb));
2292         pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
2293                ghcb->save.sw_scratch, ghcb_sw_scratch_is_valid(ghcb));
2294         pr_err("%-20s%*pb\n", "valid_bitmap", nbits, ghcb->save.valid_bitmap);
2295 }
2296
2297 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
2298 {
2299         struct kvm_vcpu *vcpu = &svm->vcpu;
2300         struct ghcb *ghcb = svm->sev_es.ghcb;
2301
2302         /*
2303          * The GHCB protocol so far allows for the following data
2304          * to be returned:
2305          *   GPRs RAX, RBX, RCX, RDX
2306          *
2307          * Copy their values, even if they may not have been written during the
2308          * VM-Exit.  It's the guest's responsibility to not consume random data.
2309          */
2310         ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
2311         ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
2312         ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
2313         ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
2314 }
2315
2316 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
2317 {
2318         struct vmcb_control_area *control = &svm->vmcb->control;
2319         struct kvm_vcpu *vcpu = &svm->vcpu;
2320         struct ghcb *ghcb = svm->sev_es.ghcb;
2321         u64 exit_code;
2322
2323         /*
2324          * The GHCB protocol so far allows for the following data
2325          * to be supplied:
2326          *   GPRs RAX, RBX, RCX, RDX
2327          *   XCR0
2328          *   CPL
2329          *
2330          * VMMCALL allows the guest to provide extra registers. KVM also
2331          * expects RSI for hypercalls, so include that, too.
2332          *
2333          * Copy their values to the appropriate location if supplied.
2334          */
2335         memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
2336
2337         vcpu->arch.regs[VCPU_REGS_RAX] = ghcb_get_rax_if_valid(ghcb);
2338         vcpu->arch.regs[VCPU_REGS_RBX] = ghcb_get_rbx_if_valid(ghcb);
2339         vcpu->arch.regs[VCPU_REGS_RCX] = ghcb_get_rcx_if_valid(ghcb);
2340         vcpu->arch.regs[VCPU_REGS_RDX] = ghcb_get_rdx_if_valid(ghcb);
2341         vcpu->arch.regs[VCPU_REGS_RSI] = ghcb_get_rsi_if_valid(ghcb);
2342
2343         svm->vmcb->save.cpl = ghcb_get_cpl_if_valid(ghcb);
2344
2345         if (ghcb_xcr0_is_valid(ghcb)) {
2346                 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
2347                 kvm_update_cpuid_runtime(vcpu);
2348         }
2349
2350         /* Copy the GHCB exit information into the VMCB fields */
2351         exit_code = ghcb_get_sw_exit_code(ghcb);
2352         control->exit_code = lower_32_bits(exit_code);
2353         control->exit_code_hi = upper_32_bits(exit_code);
2354         control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
2355         control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
2356
2357         /* Clear the valid entries fields */
2358         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
2359 }
2360
2361 static bool sev_es_validate_vmgexit(struct vcpu_svm *svm)
2362 {
2363         struct kvm_vcpu *vcpu;
2364         struct ghcb *ghcb;
2365         u64 exit_code;
2366         u64 reason;
2367
2368         ghcb = svm->sev_es.ghcb;
2369
2370         /*
2371          * Retrieve the exit code now even though it may not be marked valid
2372          * as it could help with debugging.
2373          */
2374         exit_code = ghcb_get_sw_exit_code(ghcb);
2375
2376         /* Only GHCB Usage code 0 is supported */
2377         if (ghcb->ghcb_usage) {
2378                 reason = GHCB_ERR_INVALID_USAGE;
2379                 goto vmgexit_err;
2380         }
2381
2382         reason = GHCB_ERR_MISSING_INPUT;
2383
2384         if (!ghcb_sw_exit_code_is_valid(ghcb) ||
2385             !ghcb_sw_exit_info_1_is_valid(ghcb) ||
2386             !ghcb_sw_exit_info_2_is_valid(ghcb))
2387                 goto vmgexit_err;
2388
2389         switch (ghcb_get_sw_exit_code(ghcb)) {
2390         case SVM_EXIT_READ_DR7:
2391                 break;
2392         case SVM_EXIT_WRITE_DR7:
2393                 if (!ghcb_rax_is_valid(ghcb))
2394                         goto vmgexit_err;
2395                 break;
2396         case SVM_EXIT_RDTSC:
2397                 break;
2398         case SVM_EXIT_RDPMC:
2399                 if (!ghcb_rcx_is_valid(ghcb))
2400                         goto vmgexit_err;
2401                 break;
2402         case SVM_EXIT_CPUID:
2403                 if (!ghcb_rax_is_valid(ghcb) ||
2404                     !ghcb_rcx_is_valid(ghcb))
2405                         goto vmgexit_err;
2406                 if (ghcb_get_rax(ghcb) == 0xd)
2407                         if (!ghcb_xcr0_is_valid(ghcb))
2408                                 goto vmgexit_err;
2409                 break;
2410         case SVM_EXIT_INVD:
2411                 break;
2412         case SVM_EXIT_IOIO:
2413                 if (ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_STR_MASK) {
2414                         if (!ghcb_sw_scratch_is_valid(ghcb))
2415                                 goto vmgexit_err;
2416                 } else {
2417                         if (!(ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_TYPE_MASK))
2418                                 if (!ghcb_rax_is_valid(ghcb))
2419                                         goto vmgexit_err;
2420                 }
2421                 break;
2422         case SVM_EXIT_MSR:
2423                 if (!ghcb_rcx_is_valid(ghcb))
2424                         goto vmgexit_err;
2425                 if (ghcb_get_sw_exit_info_1(ghcb)) {
2426                         if (!ghcb_rax_is_valid(ghcb) ||
2427                             !ghcb_rdx_is_valid(ghcb))
2428                                 goto vmgexit_err;
2429                 }
2430                 break;
2431         case SVM_EXIT_VMMCALL:
2432                 if (!ghcb_rax_is_valid(ghcb) ||
2433                     !ghcb_cpl_is_valid(ghcb))
2434                         goto vmgexit_err;
2435                 break;
2436         case SVM_EXIT_RDTSCP:
2437                 break;
2438         case SVM_EXIT_WBINVD:
2439                 break;
2440         case SVM_EXIT_MONITOR:
2441                 if (!ghcb_rax_is_valid(ghcb) ||
2442                     !ghcb_rcx_is_valid(ghcb) ||
2443                     !ghcb_rdx_is_valid(ghcb))
2444                         goto vmgexit_err;
2445                 break;
2446         case SVM_EXIT_MWAIT:
2447                 if (!ghcb_rax_is_valid(ghcb) ||
2448                     !ghcb_rcx_is_valid(ghcb))
2449                         goto vmgexit_err;
2450                 break;
2451         case SVM_VMGEXIT_MMIO_READ:
2452         case SVM_VMGEXIT_MMIO_WRITE:
2453                 if (!ghcb_sw_scratch_is_valid(ghcb))
2454                         goto vmgexit_err;
2455                 break;
2456         case SVM_VMGEXIT_NMI_COMPLETE:
2457         case SVM_VMGEXIT_AP_HLT_LOOP:
2458         case SVM_VMGEXIT_AP_JUMP_TABLE:
2459         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2460                 break;
2461         default:
2462                 reason = GHCB_ERR_INVALID_EVENT;
2463                 goto vmgexit_err;
2464         }
2465
2466         return true;
2467
2468 vmgexit_err:
2469         vcpu = &svm->vcpu;
2470
2471         if (reason == GHCB_ERR_INVALID_USAGE) {
2472                 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
2473                             ghcb->ghcb_usage);
2474         } else if (reason == GHCB_ERR_INVALID_EVENT) {
2475                 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx is not valid\n",
2476                             exit_code);
2477         } else {
2478                 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx input is not valid\n",
2479                             exit_code);
2480                 dump_ghcb(svm);
2481         }
2482
2483         /* Clear the valid entries fields */
2484         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
2485
2486         ghcb_set_sw_exit_info_1(ghcb, 2);
2487         ghcb_set_sw_exit_info_2(ghcb, reason);
2488
2489         return false;
2490 }
2491
2492 void sev_es_unmap_ghcb(struct vcpu_svm *svm)
2493 {
2494         if (!svm->sev_es.ghcb)
2495                 return;
2496
2497         if (svm->sev_es.ghcb_sa_free) {
2498                 /*
2499                  * The scratch area lives outside the GHCB, so there is a
2500                  * buffer that, depending on the operation performed, may
2501                  * need to be synced, then freed.
2502                  */
2503                 if (svm->sev_es.ghcb_sa_sync) {
2504                         kvm_write_guest(svm->vcpu.kvm,
2505                                         ghcb_get_sw_scratch(svm->sev_es.ghcb),
2506                                         svm->sev_es.ghcb_sa,
2507                                         svm->sev_es.ghcb_sa_len);
2508                         svm->sev_es.ghcb_sa_sync = false;
2509                 }
2510
2511                 kvfree(svm->sev_es.ghcb_sa);
2512                 svm->sev_es.ghcb_sa = NULL;
2513                 svm->sev_es.ghcb_sa_free = false;
2514         }
2515
2516         trace_kvm_vmgexit_exit(svm->vcpu.vcpu_id, svm->sev_es.ghcb);
2517
2518         sev_es_sync_to_ghcb(svm);
2519
2520         kvm_vcpu_unmap(&svm->vcpu, &svm->sev_es.ghcb_map, true);
2521         svm->sev_es.ghcb = NULL;
2522 }
2523
2524 void pre_sev_run(struct vcpu_svm *svm, int cpu)
2525 {
2526         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
2527         int asid = sev_get_asid(svm->vcpu.kvm);
2528
2529         /* Assign the asid allocated with this SEV guest */
2530         svm->asid = asid;
2531
2532         /*
2533          * Flush guest TLB:
2534          *
2535          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
2536          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
2537          */
2538         if (sd->sev_vmcbs[asid] == svm->vmcb &&
2539             svm->vcpu.arch.last_vmentry_cpu == cpu)
2540                 return;
2541
2542         sd->sev_vmcbs[asid] = svm->vmcb;
2543         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
2544         vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
2545 }
2546
2547 #define GHCB_SCRATCH_AREA_LIMIT         (16ULL * PAGE_SIZE)
2548 static bool setup_vmgexit_scratch(struct vcpu_svm *svm, bool sync, u64 len)
2549 {
2550         struct vmcb_control_area *control = &svm->vmcb->control;
2551         struct ghcb *ghcb = svm->sev_es.ghcb;
2552         u64 ghcb_scratch_beg, ghcb_scratch_end;
2553         u64 scratch_gpa_beg, scratch_gpa_end;
2554         void *scratch_va;
2555
2556         scratch_gpa_beg = ghcb_get_sw_scratch(ghcb);
2557         if (!scratch_gpa_beg) {
2558                 pr_err("vmgexit: scratch gpa not provided\n");
2559                 goto e_scratch;
2560         }
2561
2562         scratch_gpa_end = scratch_gpa_beg + len;
2563         if (scratch_gpa_end < scratch_gpa_beg) {
2564                 pr_err("vmgexit: scratch length (%#llx) not valid for scratch address (%#llx)\n",
2565                        len, scratch_gpa_beg);
2566                 goto e_scratch;
2567         }
2568
2569         if ((scratch_gpa_beg & PAGE_MASK) == control->ghcb_gpa) {
2570                 /* Scratch area begins within GHCB */
2571                 ghcb_scratch_beg = control->ghcb_gpa +
2572                                    offsetof(struct ghcb, shared_buffer);
2573                 ghcb_scratch_end = control->ghcb_gpa +
2574                                    offsetof(struct ghcb, reserved_1);
2575
2576                 /*
2577                  * If the scratch area begins within the GHCB, it must be
2578                  * completely contained in the GHCB shared buffer area.
2579                  */
2580                 if (scratch_gpa_beg < ghcb_scratch_beg ||
2581                     scratch_gpa_end > ghcb_scratch_end) {
2582                         pr_err("vmgexit: scratch area is outside of GHCB shared buffer area (%#llx - %#llx)\n",
2583                                scratch_gpa_beg, scratch_gpa_end);
2584                         goto e_scratch;
2585                 }
2586
2587                 scratch_va = (void *)svm->sev_es.ghcb;
2588                 scratch_va += (scratch_gpa_beg - control->ghcb_gpa);
2589         } else {
2590                 /*
2591                  * The guest memory must be read into a kernel buffer, so
2592                  * limit the size
2593                  */
2594                 if (len > GHCB_SCRATCH_AREA_LIMIT) {
2595                         pr_err("vmgexit: scratch area exceeds KVM limits (%#llx requested, %#llx limit)\n",
2596                                len, GHCB_SCRATCH_AREA_LIMIT);
2597                         goto e_scratch;
2598                 }
2599                 scratch_va = kvzalloc(len, GFP_KERNEL_ACCOUNT);
2600                 if (!scratch_va)
2601                         goto e_scratch;
2602
2603                 if (kvm_read_guest(svm->vcpu.kvm, scratch_gpa_beg, scratch_va, len)) {
2604                         /* Unable to copy scratch area from guest */
2605                         pr_err("vmgexit: kvm_read_guest for scratch area failed\n");
2606
2607                         kvfree(scratch_va);
2608                         goto e_scratch;
2609                 }
2610
2611                 /*
2612                  * The scratch area is outside the GHCB. The operation will
2613                  * dictate whether the buffer needs to be synced before running
2614                  * the vCPU next time (i.e. a read was requested so the data
2615                  * must be written back to the guest memory).
2616                  */
2617                 svm->sev_es.ghcb_sa_sync = sync;
2618                 svm->sev_es.ghcb_sa_free = true;
2619         }
2620
2621         svm->sev_es.ghcb_sa = scratch_va;
2622         svm->sev_es.ghcb_sa_len = len;
2623
2624         return true;
2625
2626 e_scratch:
2627         ghcb_set_sw_exit_info_1(ghcb, 2);
2628         ghcb_set_sw_exit_info_2(ghcb, GHCB_ERR_INVALID_SCRATCH_AREA);
2629
2630         return false;
2631 }
2632
2633 static void set_ghcb_msr_bits(struct vcpu_svm *svm, u64 value, u64 mask,
2634                               unsigned int pos)
2635 {
2636         svm->vmcb->control.ghcb_gpa &= ~(mask << pos);
2637         svm->vmcb->control.ghcb_gpa |= (value & mask) << pos;
2638 }
2639
2640 static u64 get_ghcb_msr_bits(struct vcpu_svm *svm, u64 mask, unsigned int pos)
2641 {
2642         return (svm->vmcb->control.ghcb_gpa >> pos) & mask;
2643 }
2644
2645 static void set_ghcb_msr(struct vcpu_svm *svm, u64 value)
2646 {
2647         svm->vmcb->control.ghcb_gpa = value;
2648 }
2649
2650 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
2651 {
2652         struct vmcb_control_area *control = &svm->vmcb->control;
2653         struct kvm_vcpu *vcpu = &svm->vcpu;
2654         u64 ghcb_info;
2655         int ret = 1;
2656
2657         ghcb_info = control->ghcb_gpa & GHCB_MSR_INFO_MASK;
2658
2659         trace_kvm_vmgexit_msr_protocol_enter(svm->vcpu.vcpu_id,
2660                                              control->ghcb_gpa);
2661
2662         switch (ghcb_info) {
2663         case GHCB_MSR_SEV_INFO_REQ:
2664                 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2665                                                     GHCB_VERSION_MIN,
2666                                                     sev_enc_bit));
2667                 break;
2668         case GHCB_MSR_CPUID_REQ: {
2669                 u64 cpuid_fn, cpuid_reg, cpuid_value;
2670
2671                 cpuid_fn = get_ghcb_msr_bits(svm,
2672                                              GHCB_MSR_CPUID_FUNC_MASK,
2673                                              GHCB_MSR_CPUID_FUNC_POS);
2674
2675                 /* Initialize the registers needed by the CPUID intercept */
2676                 vcpu->arch.regs[VCPU_REGS_RAX] = cpuid_fn;
2677                 vcpu->arch.regs[VCPU_REGS_RCX] = 0;
2678
2679                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_CPUID);
2680                 if (!ret) {
2681                         /* Error, keep GHCB MSR value as-is */
2682                         break;
2683                 }
2684
2685                 cpuid_reg = get_ghcb_msr_bits(svm,
2686                                               GHCB_MSR_CPUID_REG_MASK,
2687                                               GHCB_MSR_CPUID_REG_POS);
2688                 if (cpuid_reg == 0)
2689                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RAX];
2690                 else if (cpuid_reg == 1)
2691                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RBX];
2692                 else if (cpuid_reg == 2)
2693                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RCX];
2694                 else
2695                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RDX];
2696
2697                 set_ghcb_msr_bits(svm, cpuid_value,
2698                                   GHCB_MSR_CPUID_VALUE_MASK,
2699                                   GHCB_MSR_CPUID_VALUE_POS);
2700
2701                 set_ghcb_msr_bits(svm, GHCB_MSR_CPUID_RESP,
2702                                   GHCB_MSR_INFO_MASK,
2703                                   GHCB_MSR_INFO_POS);
2704                 break;
2705         }
2706         case GHCB_MSR_TERM_REQ: {
2707                 u64 reason_set, reason_code;
2708
2709                 reason_set = get_ghcb_msr_bits(svm,
2710                                                GHCB_MSR_TERM_REASON_SET_MASK,
2711                                                GHCB_MSR_TERM_REASON_SET_POS);
2712                 reason_code = get_ghcb_msr_bits(svm,
2713                                                 GHCB_MSR_TERM_REASON_MASK,
2714                                                 GHCB_MSR_TERM_REASON_POS);
2715                 pr_info("SEV-ES guest requested termination: %#llx:%#llx\n",
2716                         reason_set, reason_code);
2717
2718                 ret = -EINVAL;
2719                 break;
2720         }
2721         default:
2722                 /* Error, keep GHCB MSR value as-is */
2723                 break;
2724         }
2725
2726         trace_kvm_vmgexit_msr_protocol_exit(svm->vcpu.vcpu_id,
2727                                             control->ghcb_gpa, ret);
2728
2729         return ret;
2730 }
2731
2732 int sev_handle_vmgexit(struct kvm_vcpu *vcpu)
2733 {
2734         struct vcpu_svm *svm = to_svm(vcpu);
2735         struct vmcb_control_area *control = &svm->vmcb->control;
2736         u64 ghcb_gpa, exit_code;
2737         struct ghcb *ghcb;
2738         int ret;
2739
2740         /* Validate the GHCB */
2741         ghcb_gpa = control->ghcb_gpa;
2742         if (ghcb_gpa & GHCB_MSR_INFO_MASK)
2743                 return sev_handle_vmgexit_msr_protocol(svm);
2744
2745         if (!ghcb_gpa) {
2746                 vcpu_unimpl(vcpu, "vmgexit: GHCB gpa is not set\n");
2747
2748                 /* Without a GHCB, just return right back to the guest */
2749                 return 1;
2750         }
2751
2752         if (kvm_vcpu_map(vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->sev_es.ghcb_map)) {
2753                 /* Unable to map GHCB from guest */
2754                 vcpu_unimpl(vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
2755                             ghcb_gpa);
2756
2757                 /* Without a GHCB, just return right back to the guest */
2758                 return 1;
2759         }
2760
2761         svm->sev_es.ghcb = svm->sev_es.ghcb_map.hva;
2762         ghcb = svm->sev_es.ghcb_map.hva;
2763
2764         trace_kvm_vmgexit_enter(vcpu->vcpu_id, ghcb);
2765
2766         exit_code = ghcb_get_sw_exit_code(ghcb);
2767
2768         if (!sev_es_validate_vmgexit(svm))
2769                 return 1;
2770
2771         sev_es_sync_from_ghcb(svm);
2772         ghcb_set_sw_exit_info_1(ghcb, 0);
2773         ghcb_set_sw_exit_info_2(ghcb, 0);
2774
2775         ret = 1;
2776         switch (exit_code) {
2777         case SVM_VMGEXIT_MMIO_READ:
2778                 if (!setup_vmgexit_scratch(svm, true, control->exit_info_2))
2779                         break;
2780
2781                 ret = kvm_sev_es_mmio_read(vcpu,
2782                                            control->exit_info_1,
2783                                            control->exit_info_2,
2784                                            svm->sev_es.ghcb_sa);
2785                 break;
2786         case SVM_VMGEXIT_MMIO_WRITE:
2787                 if (!setup_vmgexit_scratch(svm, false, control->exit_info_2))
2788                         break;
2789
2790                 ret = kvm_sev_es_mmio_write(vcpu,
2791                                             control->exit_info_1,
2792                                             control->exit_info_2,
2793                                             svm->sev_es.ghcb_sa);
2794                 break;
2795         case SVM_VMGEXIT_NMI_COMPLETE:
2796                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_IRET);
2797                 break;
2798         case SVM_VMGEXIT_AP_HLT_LOOP:
2799                 ret = kvm_emulate_ap_reset_hold(vcpu);
2800                 break;
2801         case SVM_VMGEXIT_AP_JUMP_TABLE: {
2802                 struct kvm_sev_info *sev = &to_kvm_svm(vcpu->kvm)->sev_info;
2803
2804                 switch (control->exit_info_1) {
2805                 case 0:
2806                         /* Set AP jump table address */
2807                         sev->ap_jump_table = control->exit_info_2;
2808                         break;
2809                 case 1:
2810                         /* Get AP jump table address */
2811                         ghcb_set_sw_exit_info_2(ghcb, sev->ap_jump_table);
2812                         break;
2813                 default:
2814                         pr_err("svm: vmgexit: unsupported AP jump table request - exit_info_1=%#llx\n",
2815                                control->exit_info_1);
2816                         ghcb_set_sw_exit_info_1(ghcb, 2);
2817                         ghcb_set_sw_exit_info_2(ghcb, GHCB_ERR_INVALID_INPUT);
2818                 }
2819
2820                 break;
2821         }
2822         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2823                 vcpu_unimpl(vcpu,
2824                             "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
2825                             control->exit_info_1, control->exit_info_2);
2826                 ret = -EINVAL;
2827                 break;
2828         default:
2829                 ret = svm_invoke_exit_handler(vcpu, exit_code);
2830         }
2831
2832         return ret;
2833 }
2834
2835 int sev_es_string_io(struct vcpu_svm *svm, int size, unsigned int port, int in)
2836 {
2837         int count;
2838         int bytes;
2839
2840         if (svm->vmcb->control.exit_info_2 > INT_MAX)
2841                 return -EINVAL;
2842
2843         count = svm->vmcb->control.exit_info_2;
2844         if (unlikely(check_mul_overflow(count, size, &bytes)))
2845                 return -EINVAL;
2846
2847         if (!setup_vmgexit_scratch(svm, in, bytes))
2848                 return 1;
2849
2850         return kvm_sev_es_string_io(&svm->vcpu, size, port, svm->sev_es.ghcb_sa,
2851                                     count, in);
2852 }
2853
2854 void sev_es_init_vmcb(struct vcpu_svm *svm)
2855 {
2856         struct kvm_vcpu *vcpu = &svm->vcpu;
2857
2858         svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ES_ENABLE;
2859         svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
2860
2861         /*
2862          * An SEV-ES guest requires a VMSA area that is a separate from the
2863          * VMCB page. Do not include the encryption mask on the VMSA physical
2864          * address since hardware will access it using the guest key.
2865          */
2866         svm->vmcb->control.vmsa_pa = __pa(svm->sev_es.vmsa);
2867
2868         /* Can't intercept CR register access, HV can't modify CR registers */
2869         svm_clr_intercept(svm, INTERCEPT_CR0_READ);
2870         svm_clr_intercept(svm, INTERCEPT_CR4_READ);
2871         svm_clr_intercept(svm, INTERCEPT_CR8_READ);
2872         svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
2873         svm_clr_intercept(svm, INTERCEPT_CR4_WRITE);
2874         svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
2875
2876         svm_clr_intercept(svm, INTERCEPT_SELECTIVE_CR0);
2877
2878         /* Track EFER/CR register changes */
2879         svm_set_intercept(svm, TRAP_EFER_WRITE);
2880         svm_set_intercept(svm, TRAP_CR0_WRITE);
2881         svm_set_intercept(svm, TRAP_CR4_WRITE);
2882         svm_set_intercept(svm, TRAP_CR8_WRITE);
2883
2884         /* No support for enable_vmware_backdoor */
2885         clr_exception_intercept(svm, GP_VECTOR);
2886
2887         /* Can't intercept XSETBV, HV can't modify XCR0 directly */
2888         svm_clr_intercept(svm, INTERCEPT_XSETBV);
2889
2890         /* Clear intercepts on selected MSRs */
2891         set_msr_interception(vcpu, svm->msrpm, MSR_EFER, 1, 1);
2892         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_CR_PAT, 1, 1);
2893         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
2894         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
2895         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
2896         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
2897 }
2898
2899 void sev_es_vcpu_reset(struct vcpu_svm *svm)
2900 {
2901         /*
2902          * Set the GHCB MSR value as per the GHCB specification when emulating
2903          * vCPU RESET for an SEV-ES guest.
2904          */
2905         set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2906                                             GHCB_VERSION_MIN,
2907                                             sev_enc_bit));
2908 }
2909
2910 void sev_es_prepare_guest_switch(struct vcpu_svm *svm, unsigned int cpu)
2911 {
2912         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
2913         struct vmcb_save_area *hostsa;
2914
2915         /*
2916          * As an SEV-ES guest, hardware will restore the host state on VMEXIT,
2917          * of which one step is to perform a VMLOAD. Since hardware does not
2918          * perform a VMSAVE on VMRUN, the host savearea must be updated.
2919          */
2920         vmsave(__sme_page_pa(sd->save_area));
2921
2922         /* XCR0 is restored on VMEXIT, save the current host value */
2923         hostsa = (struct vmcb_save_area *)(page_address(sd->save_area) + 0x400);
2924         hostsa->xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
2925
2926         /* PKRU is restored on VMEXIT, save the current host value */
2927         hostsa->pkru = read_pkru();
2928
2929         /* MSR_IA32_XSS is restored on VMEXIT, save the currnet host value */
2930         hostsa->xss = host_xss;
2931 }
2932
2933 void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
2934 {
2935         struct vcpu_svm *svm = to_svm(vcpu);
2936
2937         /* First SIPI: Use the values as initially set by the VMM */
2938         if (!svm->sev_es.received_first_sipi) {
2939                 svm->sev_es.received_first_sipi = true;
2940                 return;
2941         }
2942
2943         /*
2944          * Subsequent SIPI: Return from an AP Reset Hold VMGEXIT, where
2945          * the guest will set the CS and RIP. Set SW_EXIT_INFO_2 to a
2946          * non-zero value.
2947          */
2948         if (!svm->sev_es.ghcb)
2949                 return;
2950
2951         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, 1);
2952 }