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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kvm_types.h>
  12 #include <linux/kvm_host.h>
  13 #include <linux/kernel.h>
  14 #include <linux/highmem.h>
  15 #include <linux/psp.h>
  16 #include <linux/psp-sev.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/swap.h>
  19 #include <linux/misc_cgroup.h>
  20 #include <linux/processor.h>
  21 #include <linux/trace_events.h>
  22
  23 #include <asm/pkru.h>
  24 #include <asm/trapnr.h>
  25 #include <asm/fpu/xcr.h>
  26 #include <asm/debugreg.h>
  27
  28 #include "mmu.h"
  29 #include "x86.h"
  30 #include "svm.h"
  31 #include "svm_ops.h"
  32 #include "cpuid.h"
  33 #include "trace.h"
  34
  35 #ifndef CONFIG_KVM_AMD_SEV
  36 /*
  37  * When this config is not defined, SEV feature is not supported and APIs in
  38  * this file are not used but this file still gets compiled into the KVM AMD
  39  * module.
  40  *
  41  * We will not have MISC_CG_RES_SEV and MISC_CG_RES_SEV_ES entries in the enum
  42  * misc_res_type {} defined in linux/misc_cgroup.h.
  43  *
  44  * Below macros allow compilation to succeed.
  45  */
  46 #define MISC_CG_RES_SEV MISC_CG_RES_TYPES
  47 #define MISC_CG_RES_SEV_ES MISC_CG_RES_TYPES
  48 #endif
  49
  50 #ifdef CONFIG_KVM_AMD_SEV
  51 /* enable/disable SEV support */
  52 static bool sev_enabled = true;
  53 module_param_named(sev, sev_enabled, bool, 0444);
  54
  55 /* enable/disable SEV-ES support */
  56 static bool sev_es_enabled = true;
  57 module_param_named(sev_es, sev_es_enabled, bool, 0444);
  58
  59 /* enable/disable SEV-ES DebugSwap support */
  60 static bool sev_es_debug_swap_enabled = true;
  61 module_param_named(debug_swap, sev_es_debug_swap_enabled, bool, 0444);
  62 #else
  63 #define sev_enabled false
  64 #define sev_es_enabled false
  65 #define sev_es_debug_swap_enabled false
  66 #endif /* CONFIG_KVM_AMD_SEV */
  67
  68 static u8 sev_enc_bit;
  69 static DECLARE_RWSEM(sev_deactivate_lock);
  70 static DEFINE_MUTEX(sev_bitmap_lock);
  71 unsigned int max_sev_asid;
  72 static unsigned int min_sev_asid;
  73 static unsigned long sev_me_mask;
  74 static unsigned int nr_asids;
  75 static unsigned long *sev_asid_bitmap;
  76 static unsigned long *sev_reclaim_asid_bitmap;
  77
  78 struct enc_region {
  79         struct list_head list;
  80         unsigned long npages;
  81         struct page **pages;
  82         unsigned long uaddr;
  83         unsigned long size;
  84 };
  85
  86 /* Called with the sev_bitmap_lock held, or on shutdown  */
  87 static int sev_flush_asids(int min_asid, int max_asid)
  88 {
  89         int ret, asid, error = 0;
  90
  91         /* Check if there are any ASIDs to reclaim before performing a flush */
  92         asid = find_next_bit(sev_reclaim_asid_bitmap, nr_asids, min_asid);
  93         if (asid > max_asid)
  94                 return -EBUSY;
  95
  96         /*
  97          * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
  98          * so it must be guarded.
  99          */
 100         down_write(&sev_deactivate_lock);
 101
 102         wbinvd_on_all_cpus();
 103         ret = sev_guest_df_flush(&error);
 104
 105         up_write(&sev_deactivate_lock);
 106
 107         if (ret)
 108                 pr_err("SEV: DF_FLUSH failed, ret=%d, error=%#x\n", ret, error);
 109
 110         return ret;
 111 }
 112
 113 static inline bool is_mirroring_enc_context(struct kvm *kvm)
 114 {
 115         return !!to_kvm_svm(kvm)->sev_info.enc_context_owner;
 116 }
 117
 118 /* Must be called with the sev_bitmap_lock held */
 119 static bool __sev_recycle_asids(int min_asid, int max_asid)
 120 {
 121         if (sev_flush_asids(min_asid, max_asid))
 122                 return false;
 123
 124         /* The flush process will flush all reclaimable SEV and SEV-ES ASIDs */
 125         bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
 126                    nr_asids);
 127         bitmap_zero(sev_reclaim_asid_bitmap, nr_asids);
 128
 129         return true;
 130 }
 131
 132 static int sev_misc_cg_try_charge(struct kvm_sev_info *sev)
 133 {
 134         enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 135         return misc_cg_try_charge(type, sev->misc_cg, 1);
 136 }
 137
 138 static void sev_misc_cg_uncharge(struct kvm_sev_info *sev)
 139 {
 140         enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 141         misc_cg_uncharge(type, sev->misc_cg, 1);
 142 }
 143
 144 static int sev_asid_new(struct kvm_sev_info *sev)
 145 {
 146         int asid, min_asid, max_asid, ret;
 147         bool retry = true;
 148
 149         WARN_ON(sev->misc_cg);
 150         sev->misc_cg = get_current_misc_cg();
 151         ret = sev_misc_cg_try_charge(sev);
 152         if (ret) {
 153                 put_misc_cg(sev->misc_cg);
 154                 sev->misc_cg = NULL;
 155                 return ret;
 156         }
 157
 158         mutex_lock(&sev_bitmap_lock);
 159
 160         /*
 161          * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
 162          * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.
 163          */
 164         min_asid = sev->es_active ? 1 : min_sev_asid;
 165         max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
 166 again:
 167         asid = find_next_zero_bit(sev_asid_bitmap, max_asid + 1, min_asid);
 168         if (asid > max_asid) {
 169                 if (retry && __sev_recycle_asids(min_asid, max_asid)) {
 170                         retry = false;
 171                         goto again;
 172                 }
 173                 mutex_unlock(&sev_bitmap_lock);
 174                 ret = -EBUSY;
 175                 goto e_uncharge;
 176         }
 177
 178         __set_bit(asid, sev_asid_bitmap);
 179
 180         mutex_unlock(&sev_bitmap_lock);
 181
 182         return asid;
 183 e_uncharge:
 184         sev_misc_cg_uncharge(sev);
 185         put_misc_cg(sev->misc_cg);
 186         sev->misc_cg = NULL;
 187         return ret;
 188 }
 189
 190 static int sev_get_asid(struct kvm *kvm)
 191 {
 192         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 193
 194         return sev->asid;
 195 }
 196
 197 static void sev_asid_free(struct kvm_sev_info *sev)
 198 {
 199         struct svm_cpu_data *sd;
 200         int cpu;
 201
 202         mutex_lock(&sev_bitmap_lock);
 203
 204         __set_bit(sev->asid, sev_reclaim_asid_bitmap);
 205
 206         for_each_possible_cpu(cpu) {
 207                 sd = per_cpu_ptr(&svm_data, cpu);
 208                 sd->sev_vmcbs[sev->asid] = NULL;
 209         }
 210
 211         mutex_unlock(&sev_bitmap_lock);
 212
 213         sev_misc_cg_uncharge(sev);
 214         put_misc_cg(sev->misc_cg);
 215         sev->misc_cg = NULL;
 216 }
 217
 218 static void sev_decommission(unsigned int handle)
 219 {
 220         struct sev_data_decommission decommission;
 221
 222         if (!handle)
 223                 return;
 224
 225         decommission.handle = handle;
 226         sev_guest_decommission(&decommission, NULL);
 227 }
 228
 229 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
 230 {
 231         struct sev_data_deactivate deactivate;
 232
 233         if (!handle)
 234                 return;
 235
 236         deactivate.handle = handle;
 237
 238         /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
 239         down_read(&sev_deactivate_lock);
 240         sev_guest_deactivate(&deactivate, NULL);
 241         up_read(&sev_deactivate_lock);
 242
 243         sev_decommission(handle);
 244 }
 245
 246 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 247 {
 248         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 249         int asid, ret;
 250
 251         if (kvm->created_vcpus)
 252                 return -EINVAL;
 253
 254         ret = -EBUSY;
 255         if (unlikely(sev->active))
 256                 return ret;
 257
 258         sev->active = true;
 259         sev->es_active = argp->id == KVM_SEV_ES_INIT;
 260         asid = sev_asid_new(sev);
 261         if (asid < 0)
 262                 goto e_no_asid;
 263         sev->asid = asid;
 264
 265         ret = sev_platform_init(&argp->error);
 266         if (ret)
 267                 goto e_free;
 268
 269         INIT_LIST_HEAD(&sev->regions_list);
 270         INIT_LIST_HEAD(&sev->mirror_vms);
 271
 272         kvm_set_apicv_inhibit(kvm, APICV_INHIBIT_REASON_SEV);
 273
 274         return 0;
 275
 276 e_free:
 277         sev_asid_free(sev);
 278         sev->asid = 0;
 279 e_no_asid:
 280         sev->es_active = false;
 281         sev->active = false;
 282         return ret;
 283 }
 284
 285 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
 286 {
 287         struct sev_data_activate activate;
 288         int asid = sev_get_asid(kvm);
 289         int ret;
 290
 291         /* activate ASID on the given handle */
 292         activate.handle = handle;
 293         activate.asid   = asid;
 294         ret = sev_guest_activate(&activate, error);
 295
 296         return ret;
 297 }
 298
 299 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
 300 {
 301         struct fd f;
 302         int ret;
 303
 304         f = fdget(fd);
 305         if (!f.file)
 306                 return -EBADF;
 307
 308         ret = sev_issue_cmd_external_user(f.file, id, data, error);
 309
 310         fdput(f);
 311         return ret;
 312 }
 313
 314 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
 315 {
 316         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 317
 318         return __sev_issue_cmd(sev->fd, id, data, error);
 319 }
 320
 321 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
 322 {
 323         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 324         struct sev_data_launch_start start;
 325         struct kvm_sev_launch_start params;
 326         void *dh_blob, *session_blob;
 327         int *error = &argp->error;
 328         int ret;
 329
 330         if (!sev_guest(kvm))
 331                 return -ENOTTY;
 332
 333         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 334                 return -EFAULT;
 335
 336         memset(&start, 0, sizeof(start));
 337
 338         dh_blob = NULL;
 339         if (params.dh_uaddr) {
 340                 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
 341                 if (IS_ERR(dh_blob))
 342                         return PTR_ERR(dh_blob);
 343
 344                 start.dh_cert_address = __sme_set(__pa(dh_blob));
 345                 start.dh_cert_len = params.dh_len;
 346         }
 347
 348         session_blob = NULL;
 349         if (params.session_uaddr) {
 350                 session_blob = psp_copy_user_blob(params.session_uaddr, params.session_len);
 351                 if (IS_ERR(session_blob)) {
 352                         ret = PTR_ERR(session_blob);
 353                         goto e_free_dh;
 354                 }
 355
 356                 start.session_address = __sme_set(__pa(session_blob));
 357                 start.session_len = params.session_len;
 358         }
 359
 360         start.handle = params.handle;
 361         start.policy = params.policy;
 362
 363         /* create memory encryption context */
 364         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_LAUNCH_START, &start, error);
 365         if (ret)
 366                 goto e_free_session;
 367
 368         /* Bind ASID to this guest */
 369         ret = sev_bind_asid(kvm, start.handle, error);
 370         if (ret) {
 371                 sev_decommission(start.handle);
 372                 goto e_free_session;
 373         }
 374
 375         /* return handle to userspace */
 376         params.handle = start.handle;
 377         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params))) {
 378                 sev_unbind_asid(kvm, start.handle);
 379                 ret = -EFAULT;
 380                 goto e_free_session;
 381         }
 382
 383         sev->handle = start.handle;
 384         sev->fd = argp->sev_fd;
 385
 386 e_free_session:
 387         kfree(session_blob);
 388 e_free_dh:
 389         kfree(dh_blob);
 390         return ret;
 391 }
 392
 393 static struct page **sev_pin_memory(struct kvm *kvm, unsigned long uaddr,
 394                                     unsigned long ulen, unsigned long *n,
 395                                     int write)
 396 {
 397         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 398         unsigned long npages, size;
 399         int npinned;
 400         unsigned long locked, lock_limit;
 401         struct page **pages;
 402         unsigned long first, last;
 403         int ret;
 404
 405         lockdep_assert_held(&kvm->lock);
 406
 407         if (ulen == 0 || uaddr + ulen < uaddr)
 408                 return ERR_PTR(-EINVAL);
 409
 410         /* Calculate number of pages. */
 411         first = (uaddr & PAGE_MASK) >> PAGE_SHIFT;
 412         last = ((uaddr + ulen - 1) & PAGE_MASK) >> PAGE_SHIFT;
 413         npages = (last - first + 1);
 414
 415         locked = sev->pages_locked + npages;
 416         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 417         if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
 418                 pr_err("SEV: %lu locked pages exceed the lock limit of %lu.\n", locked, lock_limit);
 419                 return ERR_PTR(-ENOMEM);
 420         }
 421
 422         if (WARN_ON_ONCE(npages > INT_MAX))
 423                 return ERR_PTR(-EINVAL);
 424
 425         /* Avoid using vmalloc for smaller buffers. */
 426         size = npages * sizeof(struct page *);
 427         if (size > PAGE_SIZE)
 428                 pages = __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 429         else
 430                 pages = kmalloc(size, GFP_KERNEL_ACCOUNT);
 431
 432         if (!pages)
 433                 return ERR_PTR(-ENOMEM);
 434
 435         /* Pin the user virtual address. */
 436         npinned = pin_user_pages_fast(uaddr, npages, write ? FOLL_WRITE : 0, pages);
 437         if (npinned != npages) {
 438                 pr_err("SEV: Failure locking %lu pages.\n", npages);
 439                 ret = -ENOMEM;
 440                 goto err;
 441         }
 442
 443         *n = npages;
 444         sev->pages_locked = locked;
 445
 446         return pages;
 447
 448 err:
 449         if (npinned > 0)
 450                 unpin_user_pages(pages, npinned);
 451
 452         kvfree(pages);
 453         return ERR_PTR(ret);
 454 }
 455
 456 static void sev_unpin_memory(struct kvm *kvm, struct page **pages,
 457                              unsigned long npages)
 458 {
 459         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 460
 461         unpin_user_pages(pages, npages);
 462         kvfree(pages);
 463         sev->pages_locked -= npages;
 464 }
 465
 466 static void sev_clflush_pages(struct page *pages[], unsigned long npages)
 467 {
 468         uint8_t *page_virtual;
 469         unsigned long i;
 470
 471         if (this_cpu_has(X86_FEATURE_SME_COHERENT) || npages == 0 ||
 472             pages == NULL)
 473                 return;
 474
 475         for (i = 0; i < npages; i++) {
 476                 page_virtual = kmap_local_page(pages[i]);
 477                 clflush_cache_range(page_virtual, PAGE_SIZE);
 478                 kunmap_local(page_virtual);
 479                 cond_resched();
 480         }
 481 }
 482
 483 static unsigned long get_num_contig_pages(unsigned long idx,
 484                                 struct page **inpages, unsigned long npages)
 485 {
 486         unsigned long paddr, next_paddr;
 487         unsigned long i = idx + 1, pages = 1;
 488
 489         /* find the number of contiguous pages starting from idx */
 490         paddr = __sme_page_pa(inpages[idx]);
 491         while (i < npages) {
 492                 next_paddr = __sme_page_pa(inpages[i++]);
 493                 if ((paddr + PAGE_SIZE) == next_paddr) {
 494                         pages++;
 495                         paddr = next_paddr;
 496                         continue;
 497                 }
 498                 break;
 499         }
 500
 501         return pages;
 502 }
 503
 504 static int sev_launch_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
 505 {
 506         unsigned long vaddr, vaddr_end, next_vaddr, npages, pages, size, i;
 507         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 508         struct kvm_sev_launch_update_data params;
 509         struct sev_data_launch_update_data data;
 510         struct page **inpages;
 511         int ret;
 512
 513         if (!sev_guest(kvm))
 514                 return -ENOTTY;
 515
 516         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 517                 return -EFAULT;
 518
 519         vaddr = params.uaddr;
 520         size = params.len;
 521         vaddr_end = vaddr + size;
 522
 523         /* Lock the user memory. */
 524         inpages = sev_pin_memory(kvm, vaddr, size, &npages, 1);
 525         if (IS_ERR(inpages))
 526                 return PTR_ERR(inpages);
 527
 528         /*
 529          * Flush (on non-coherent CPUs) before LAUNCH_UPDATE encrypts pages in
 530          * place; the cache may contain the data that was written unencrypted.
 531          */
 532         sev_clflush_pages(inpages, npages);
 533
 534         data.reserved = 0;
 535         data.handle = sev->handle;
 536
 537         for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
 538                 int offset, len;
 539
 540                 /*
 541                  * If the user buffer is not page-aligned, calculate the offset
 542                  * within the page.
 543                  */
 544                 offset = vaddr & (PAGE_SIZE - 1);
 545
 546                 /* Calculate the number of pages that can be encrypted in one go. */
 547                 pages = get_num_contig_pages(i, inpages, npages);
 548
 549                 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
 550
 551                 data.len = len;
 552                 data.address = __sme_page_pa(inpages[i]) + offset;
 553                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, &data, &argp->error);
 554                 if (ret)
 555                         goto e_unpin;
 556
 557                 size -= len;
 558                 next_vaddr = vaddr + len;
 559         }
 560
 561 e_unpin:
 562         /* content of memory is updated, mark pages dirty */
 563         for (i = 0; i < npages; i++) {
 564                 set_page_dirty_lock(inpages[i]);
 565                 mark_page_accessed(inpages[i]);
 566         }
 567         /* unlock the user pages */
 568         sev_unpin_memory(kvm, inpages, npages);
 569         return ret;
 570 }
 571
 572 static int sev_es_sync_vmsa(struct vcpu_svm *svm)
 573 {
 574         struct sev_es_save_area *save = svm->sev_es.vmsa;
 575
 576         /* Check some debug related fields before encrypting the VMSA */
 577         if (svm->vcpu.guest_debug || (svm->vmcb->save.dr7 & ~DR7_FIXED_1))
 578                 return -EINVAL;
 579
 580         /*
 581          * SEV-ES will use a VMSA that is pointed to by the VMCB, not
 582          * the traditional VMSA that is part of the VMCB. Copy the
 583          * traditional VMSA as it has been built so far (in prep
 584          * for LAUNCH_UPDATE_VMSA) to be the initial SEV-ES state.
 585          */
 586         memcpy(save, &svm->vmcb->save, sizeof(svm->vmcb->save));
 587
 588         /* Sync registgers */
 589         save->rax = svm->vcpu.arch.regs[VCPU_REGS_RAX];
 590         save->rbx = svm->vcpu.arch.regs[VCPU_REGS_RBX];
 591         save->rcx = svm->vcpu.arch.regs[VCPU_REGS_RCX];
 592         save->rdx = svm->vcpu.arch.regs[VCPU_REGS_RDX];
 593         save->rsp = svm->vcpu.arch.regs[VCPU_REGS_RSP];
 594         save->rbp = svm->vcpu.arch.regs[VCPU_REGS_RBP];
 595         save->rsi = svm->vcpu.arch.regs[VCPU_REGS_RSI];
 596         save->rdi = svm->vcpu.arch.regs[VCPU_REGS_RDI];
 597 #ifdef CONFIG_X86_64
 598         save->r8  = svm->vcpu.arch.regs[VCPU_REGS_R8];
 599         save->r9  = svm->vcpu.arch.regs[VCPU_REGS_R9];
 600         save->r10 = svm->vcpu.arch.regs[VCPU_REGS_R10];
 601         save->r11 = svm->vcpu.arch.regs[VCPU_REGS_R11];
 602         save->r12 = svm->vcpu.arch.regs[VCPU_REGS_R12];
 603         save->r13 = svm->vcpu.arch.regs[VCPU_REGS_R13];
 604         save->r14 = svm->vcpu.arch.regs[VCPU_REGS_R14];
 605         save->r15 = svm->vcpu.arch.regs[VCPU_REGS_R15];
 606 #endif
 607         save->rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
 608
 609         /* Sync some non-GPR registers before encrypting */
 610         save->xcr0 = svm->vcpu.arch.xcr0;
 611         save->pkru = svm->vcpu.arch.pkru;
 612         save->xss  = svm->vcpu.arch.ia32_xss;
 613         save->dr6  = svm->vcpu.arch.dr6;
 614
 615         if (sev_es_debug_swap_enabled)
 616                 save->sev_features |= SVM_SEV_FEAT_DEBUG_SWAP;
 617
 618         pr_debug("Virtual Machine Save Area (VMSA):\n");
 619         print_hex_dump_debug("", DUMP_PREFIX_NONE, 16, 1, save, sizeof(*save), false);
 620
 621         return 0;
 622 }
 623
 624 static int __sev_launch_update_vmsa(struct kvm *kvm, struct kvm_vcpu *vcpu,
 625                                     int *error)
 626 {
 627         struct sev_data_launch_update_vmsa vmsa;
 628         struct vcpu_svm *svm = to_svm(vcpu);
 629         int ret;
 630
 631         if (vcpu->guest_debug) {
 632                 pr_warn_once("KVM_SET_GUEST_DEBUG for SEV-ES guest is not supported");
 633                 return -EINVAL;
 634         }
 635
 636         /* Perform some pre-encryption checks against the VMSA */
 637         ret = sev_es_sync_vmsa(svm);
 638         if (ret)
 639                 return ret;
 640
 641         /*
 642          * The LAUNCH_UPDATE_VMSA command will perform in-place encryption of
 643          * the VMSA memory content (i.e it will write the same memory region
 644          * with the guest's key), so invalidate it first.
 645          */
 646         clflush_cache_range(svm->sev_es.vmsa, PAGE_SIZE);
 647
 648         vmsa.reserved = 0;
 649         vmsa.handle = to_kvm_svm(kvm)->sev_info.handle;
 650         vmsa.address = __sme_pa(svm->sev_es.vmsa);
 651         vmsa.len = PAGE_SIZE;
 652         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_VMSA, &vmsa, error);
 653         if (ret)
 654           return ret;
 655
 656         vcpu->arch.guest_state_protected = true;
 657         return 0;
 658 }
 659
 660 static int sev_launch_update_vmsa(struct kvm *kvm, struct kvm_sev_cmd *argp)
 661 {
 662         struct kvm_vcpu *vcpu;
 663         unsigned long i;
 664         int ret;
 665
 666         if (!sev_es_guest(kvm))
 667                 return -ENOTTY;
 668
 669         kvm_for_each_vcpu(i, vcpu, kvm) {
 670                 ret = mutex_lock_killable(&vcpu->mutex);
 671                 if (ret)
 672                         return ret;
 673
 674                 ret = __sev_launch_update_vmsa(kvm, vcpu, &argp->error);
 675
 676                 mutex_unlock(&vcpu->mutex);
 677                 if (ret)
 678                         return ret;
 679         }
 680
 681         return 0;
 682 }
 683
 684 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 685 {
 686         void __user *measure = (void __user *)(uintptr_t)argp->data;
 687         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 688         struct sev_data_launch_measure data;
 689         struct kvm_sev_launch_measure params;
 690         void __user *p = NULL;
 691         void *blob = NULL;
 692         int ret;
 693
 694         if (!sev_guest(kvm))
 695                 return -ENOTTY;
 696
 697         if (copy_from_user(&params, measure, sizeof(params)))
 698                 return -EFAULT;
 699
 700         memset(&data, 0, sizeof(data));
 701
 702         /* User wants to query the blob length */
 703         if (!params.len)
 704                 goto cmd;
 705
 706         p = (void __user *)(uintptr_t)params.uaddr;
 707         if (p) {
 708                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
 709                         return -EINVAL;
 710
 711                 blob = kzalloc(params.len, GFP_KERNEL_ACCOUNT);
 712                 if (!blob)
 713                         return -ENOMEM;
 714
 715                 data.address = __psp_pa(blob);
 716                 data.len = params.len;
 717         }
 718
 719 cmd:
 720         data.handle = sev->handle;
 721         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, &data, &argp->error);
 722
 723         /*
 724          * If we query the session length, FW responded with expected data.
 725          */
 726         if (!params.len)
 727                 goto done;
 728
 729         if (ret)
 730                 goto e_free_blob;
 731
 732         if (blob) {
 733                 if (copy_to_user(p, blob, params.len))
 734                         ret = -EFAULT;
 735         }
 736
 737 done:
 738         params.len = data.len;
 739         if (copy_to_user(measure, &params, sizeof(params)))
 740                 ret = -EFAULT;
 741 e_free_blob:
 742         kfree(blob);
 743         return ret;
 744 }
 745
 746 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 747 {
 748         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 749         struct sev_data_launch_finish data;
 750
 751         if (!sev_guest(kvm))
 752                 return -ENOTTY;
 753
 754         data.handle = sev->handle;
 755         return sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, &data, &argp->error);
 756 }
 757
 758 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 759 {
 760         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 761         struct kvm_sev_guest_status params;
 762         struct sev_data_guest_status data;
 763         int ret;
 764
 765         if (!sev_guest(kvm))
 766                 return -ENOTTY;
 767
 768         memset(&data, 0, sizeof(data));
 769
 770         data.handle = sev->handle;
 771         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, &data, &argp->error);
 772         if (ret)
 773                 return ret;
 774
 775         params.policy = data.policy;
 776         params.state = data.state;
 777         params.handle = data.handle;
 778
 779         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 780                 ret = -EFAULT;
 781
 782         return ret;
 783 }
 784
 785 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 786                                unsigned long dst, int size,
 787                                int *error, bool enc)
 788 {
 789         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 790         struct sev_data_dbg data;
 791
 792         data.reserved = 0;
 793         data.handle = sev->handle;
 794         data.dst_addr = dst;
 795         data.src_addr = src;
 796         data.len = size;
 797
 798         return sev_issue_cmd(kvm,
 799                              enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 800                              &data, error);
 801 }
 802
 803 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 804                              unsigned long dst_paddr, int sz, int *err)
 805 {
 806         int offset;
 807
 808         /*
 809          * Its safe to read more than we are asked, caller should ensure that
 810          * destination has enough space.
 811          */
 812         offset = src_paddr & 15;
 813         src_paddr = round_down(src_paddr, 16);
 814         sz = round_up(sz + offset, 16);
 815
 816         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 817 }
 818
 819 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 820                                   void __user *dst_uaddr,
 821                                   unsigned long dst_paddr,
 822                                   int size, int *err)
 823 {
 824         struct page *tpage = NULL;
 825         int ret, offset;
 826
 827         /* if inputs are not 16-byte then use intermediate buffer */
 828         if (!IS_ALIGNED(dst_paddr, 16) ||
 829             !IS_ALIGNED(paddr,     16) ||
 830             !IS_ALIGNED(size,      16)) {
 831                 tpage = (void *)alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 832                 if (!tpage)
 833                         return -ENOMEM;
 834
 835                 dst_paddr = __sme_page_pa(tpage);
 836         }
 837
 838         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 839         if (ret)
 840                 goto e_free;
 841
 842         if (tpage) {
 843                 offset = paddr & 15;
 844                 if (copy_to_user(dst_uaddr, page_address(tpage) + offset, size))
 845                         ret = -EFAULT;
 846         }
 847
 848 e_free:
 849         if (tpage)
 850                 __free_page(tpage);
 851
 852         return ret;
 853 }
 854
 855 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 856                                   void __user *vaddr,
 857                                   unsigned long dst_paddr,
 858                                   void __user *dst_vaddr,
 859                                   int size, int *error)
 860 {
 861         struct page *src_tpage = NULL;
 862         struct page *dst_tpage = NULL;
 863         int ret, len = size;
 864
 865         /* If source buffer is not aligned then use an intermediate buffer */
 866         if (!IS_ALIGNED((unsigned long)vaddr, 16)) {
 867                 src_tpage = alloc_page(GFP_KERNEL_ACCOUNT);
 868                 if (!src_tpage)
 869                         return -ENOMEM;
 870
 871                 if (copy_from_user(page_address(src_tpage), vaddr, size)) {
 872                         __free_page(src_tpage);
 873                         return -EFAULT;
 874                 }
 875
 876                 paddr = __sme_page_pa(src_tpage);
 877         }
 878
 879         /*
 880          *  If destination buffer or length is not aligned then do read-modify-write:
 881          *   - decrypt destination in an intermediate buffer
 882          *   - copy the source buffer in an intermediate buffer
 883          *   - use the intermediate buffer as source buffer
 884          */
 885         if (!IS_ALIGNED((unsigned long)dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 886                 int dst_offset;
 887
 888                 dst_tpage = alloc_page(GFP_KERNEL_ACCOUNT);
 889                 if (!dst_tpage) {
 890                         ret = -ENOMEM;
 891                         goto e_free;
 892                 }
 893
 894                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 895                                         __sme_page_pa(dst_tpage), size, error);
 896                 if (ret)
 897                         goto e_free;
 898
 899                 /*
 900                  *  If source is kernel buffer then use memcpy() otherwise
 901                  *  copy_from_user().
 902                  */
 903                 dst_offset = dst_paddr & 15;
 904
 905                 if (src_tpage)
 906                         memcpy(page_address(dst_tpage) + dst_offset,
 907                                page_address(src_tpage), size);
 908                 else {
 909                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 910                                            vaddr, size)) {
 911                                 ret = -EFAULT;
 912                                 goto e_free;
 913                         }
 914                 }
 915
 916                 paddr = __sme_page_pa(dst_tpage);
 917                 dst_paddr = round_down(dst_paddr, 16);
 918                 len = round_up(size, 16);
 919         }
 920
 921         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 922
 923 e_free:
 924         if (src_tpage)
 925                 __free_page(src_tpage);
 926         if (dst_tpage)
 927                 __free_page(dst_tpage);
 928         return ret;
 929 }
 930
 931 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 932 {
 933         unsigned long vaddr, vaddr_end, next_vaddr;
 934         unsigned long dst_vaddr;
 935         struct page **src_p, **dst_p;
 936         struct kvm_sev_dbg debug;
 937         unsigned long n;
 938         unsigned int size;
 939         int ret;
 940
 941         if (!sev_guest(kvm))
 942                 return -ENOTTY;
 943
 944         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 945                 return -EFAULT;
 946
 947         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 948                 return -EINVAL;
 949         if (!debug.dst_uaddr)
 950                 return -EINVAL;
 951
 952         vaddr = debug.src_uaddr;
 953         size = debug.len;
 954         vaddr_end = vaddr + size;
 955         dst_vaddr = debug.dst_uaddr;
 956
 957         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 958                 int len, s_off, d_off;
 959
 960                 /* lock userspace source and destination page */
 961                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 962                 if (IS_ERR(src_p))
 963                         return PTR_ERR(src_p);
 964
 965                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 966                 if (IS_ERR(dst_p)) {
 967                         sev_unpin_memory(kvm, src_p, n);
 968                         return PTR_ERR(dst_p);
 969                 }
 970
 971                 /*
 972                  * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
 973                  * the pages; flush the destination too so that future accesses do not
 974                  * see stale data.
 975                  */
 976                 sev_clflush_pages(src_p, 1);
 977                 sev_clflush_pages(dst_p, 1);
 978
 979                 /*
 980                  * Since user buffer may not be page aligned, calculate the
 981                  * offset within the page.
 982                  */
 983                 s_off = vaddr & ~PAGE_MASK;
 984                 d_off = dst_vaddr & ~PAGE_MASK;
 985                 len = min_t(size_t, (PAGE_SIZE - s_off), size);
 986
 987                 if (dec)
 988                         ret = __sev_dbg_decrypt_user(kvm,
 989                                                      __sme_page_pa(src_p[0]) + s_off,
 990                                                      (void __user *)dst_vaddr,
 991                                                      __sme_page_pa(dst_p[0]) + d_off,
 992                                                      len, &argp->error);
 993                 else
 994                         ret = __sev_dbg_encrypt_user(kvm,
 995                                                      __sme_page_pa(src_p[0]) + s_off,
 996                                                      (void __user *)vaddr,
 997                                                      __sme_page_pa(dst_p[0]) + d_off,
 998                                                      (void __user *)dst_vaddr,
 999                                                      len, &argp->error);
1000
1001                 sev_unpin_memory(kvm, src_p, n);
1002                 sev_unpin_memory(kvm, dst_p, n);
1003
1004                 if (ret)
1005                         goto err;
1006
1007                 next_vaddr = vaddr + len;
1008                 dst_vaddr = dst_vaddr + len;
1009                 size -= len;
1010         }
1011 err:
1012         return ret;
1013 }
1014
1015 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
1016 {
1017         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1018         struct sev_data_launch_secret data;
1019         struct kvm_sev_launch_secret params;
1020         struct page **pages;
1021         void *blob, *hdr;
1022         unsigned long n, i;
1023         int ret, offset;
1024
1025         if (!sev_guest(kvm))
1026                 return -ENOTTY;
1027
1028         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1029                 return -EFAULT;
1030
1031         pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, 1);
1032         if (IS_ERR(pages))
1033                 return PTR_ERR(pages);
1034
1035         /*
1036          * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
1037          * place; the cache may contain the data that was written unencrypted.
1038          */
1039         sev_clflush_pages(pages, n);
1040
1041         /*
1042          * The secret must be copied into contiguous memory region, lets verify
1043          * that userspace memory pages are contiguous before we issue command.
1044          */
1045         if (get_num_contig_pages(0, pages, n) != n) {
1046                 ret = -EINVAL;
1047                 goto e_unpin_memory;
1048         }
1049
1050         memset(&data, 0, sizeof(data));
1051
1052         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1053         data.guest_address = __sme_page_pa(pages[0]) + offset;
1054         data.guest_len = params.guest_len;
1055
1056         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1057         if (IS_ERR(blob)) {
1058                 ret = PTR_ERR(blob);
1059                 goto e_unpin_memory;
1060         }
1061
1062         data.trans_address = __psp_pa(blob);
1063         data.trans_len = params.trans_len;
1064
1065         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1066         if (IS_ERR(hdr)) {
1067                 ret = PTR_ERR(hdr);
1068                 goto e_free_blob;
1069         }
1070         data.hdr_address = __psp_pa(hdr);
1071         data.hdr_len = params.hdr_len;
1072
1073         data.handle = sev->handle;
1074         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, &data, &argp->error);
1075
1076         kfree(hdr);
1077
1078 e_free_blob:
1079         kfree(blob);
1080 e_unpin_memory:
1081         /* content of memory is updated, mark pages dirty */
1082         for (i = 0; i < n; i++) {
1083                 set_page_dirty_lock(pages[i]);
1084                 mark_page_accessed(pages[i]);
1085         }
1086         sev_unpin_memory(kvm, pages, n);
1087         return ret;
1088 }
1089
1090 static int sev_get_attestation_report(struct kvm *kvm, struct kvm_sev_cmd *argp)
1091 {
1092         void __user *report = (void __user *)(uintptr_t)argp->data;
1093         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1094         struct sev_data_attestation_report data;
1095         struct kvm_sev_attestation_report params;
1096         void __user *p;
1097         void *blob = NULL;
1098         int ret;
1099
1100         if (!sev_guest(kvm))
1101                 return -ENOTTY;
1102
1103         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1104                 return -EFAULT;
1105
1106         memset(&data, 0, sizeof(data));
1107
1108         /* User wants to query the blob length */
1109         if (!params.len)
1110                 goto cmd;
1111
1112         p = (void __user *)(uintptr_t)params.uaddr;
1113         if (p) {
1114                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
1115                         return -EINVAL;
1116
1117                 blob = kzalloc(params.len, GFP_KERNEL_ACCOUNT);
1118                 if (!blob)
1119                         return -ENOMEM;
1120
1121                 data.address = __psp_pa(blob);
1122                 data.len = params.len;
1123                 memcpy(data.mnonce, params.mnonce, sizeof(params.mnonce));
1124         }
1125 cmd:
1126         data.handle = sev->handle;
1127         ret = sev_issue_cmd(kvm, SEV_CMD_ATTESTATION_REPORT, &data, &argp->error);
1128         /*
1129          * If we query the session length, FW responded with expected data.
1130          */
1131         if (!params.len)
1132                 goto done;
1133
1134         if (ret)
1135                 goto e_free_blob;
1136
1137         if (blob) {
1138                 if (copy_to_user(p, blob, params.len))
1139                         ret = -EFAULT;
1140         }
1141
1142 done:
1143         params.len = data.len;
1144         if (copy_to_user(report, &params, sizeof(params)))
1145                 ret = -EFAULT;
1146 e_free_blob:
1147         kfree(blob);
1148         return ret;
1149 }
1150
1151 /* Userspace wants to query session length. */
1152 static int
1153 __sev_send_start_query_session_length(struct kvm *kvm, struct kvm_sev_cmd *argp,
1154                                       struct kvm_sev_send_start *params)
1155 {
1156         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1157         struct sev_data_send_start data;
1158         int ret;
1159
1160         memset(&data, 0, sizeof(data));
1161         data.handle = sev->handle;
1162         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1163
1164         params->session_len = data.session_len;
1165         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1166                                 sizeof(struct kvm_sev_send_start)))
1167                 ret = -EFAULT;
1168
1169         return ret;
1170 }
1171
1172 static int sev_send_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1173 {
1174         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1175         struct sev_data_send_start data;
1176         struct kvm_sev_send_start params;
1177         void *amd_certs, *session_data;
1178         void *pdh_cert, *plat_certs;
1179         int ret;
1180
1181         if (!sev_guest(kvm))
1182                 return -ENOTTY;
1183
1184         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1185                                 sizeof(struct kvm_sev_send_start)))
1186                 return -EFAULT;
1187
1188         /* if session_len is zero, userspace wants to query the session length */
1189         if (!params.session_len)
1190                 return __sev_send_start_query_session_length(kvm, argp,
1191                                 &params);
1192
1193         /* some sanity checks */
1194         if (!params.pdh_cert_uaddr || !params.pdh_cert_len ||
1195             !params.session_uaddr || params.session_len > SEV_FW_BLOB_MAX_SIZE)
1196                 return -EINVAL;
1197
1198         /* allocate the memory to hold the session data blob */
1199         session_data = kzalloc(params.session_len, GFP_KERNEL_ACCOUNT);
1200         if (!session_data)
1201                 return -ENOMEM;
1202
1203         /* copy the certificate blobs from userspace */
1204         pdh_cert = psp_copy_user_blob(params.pdh_cert_uaddr,
1205                                 params.pdh_cert_len);
1206         if (IS_ERR(pdh_cert)) {
1207                 ret = PTR_ERR(pdh_cert);
1208                 goto e_free_session;
1209         }
1210
1211         plat_certs = psp_copy_user_blob(params.plat_certs_uaddr,
1212                                 params.plat_certs_len);
1213         if (IS_ERR(plat_certs)) {
1214                 ret = PTR_ERR(plat_certs);
1215                 goto e_free_pdh;
1216         }
1217
1218         amd_certs = psp_copy_user_blob(params.amd_certs_uaddr,
1219                                 params.amd_certs_len);
1220         if (IS_ERR(amd_certs)) {
1221                 ret = PTR_ERR(amd_certs);
1222                 goto e_free_plat_cert;
1223         }
1224
1225         /* populate the FW SEND_START field with system physical address */
1226         memset(&data, 0, sizeof(data));
1227         data.pdh_cert_address = __psp_pa(pdh_cert);
1228         data.pdh_cert_len = params.pdh_cert_len;
1229         data.plat_certs_address = __psp_pa(plat_certs);
1230         data.plat_certs_len = params.plat_certs_len;
1231         data.amd_certs_address = __psp_pa(amd_certs);
1232         data.amd_certs_len = params.amd_certs_len;
1233         data.session_address = __psp_pa(session_data);
1234         data.session_len = params.session_len;
1235         data.handle = sev->handle;
1236
1237         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1238
1239         if (!ret && copy_to_user((void __user *)(uintptr_t)params.session_uaddr,
1240                         session_data, params.session_len)) {
1241                 ret = -EFAULT;
1242                 goto e_free_amd_cert;
1243         }
1244
1245         params.policy = data.policy;
1246         params.session_len = data.session_len;
1247         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params,
1248                                 sizeof(struct kvm_sev_send_start)))
1249                 ret = -EFAULT;
1250
1251 e_free_amd_cert:
1252         kfree(amd_certs);
1253 e_free_plat_cert:
1254         kfree(plat_certs);
1255 e_free_pdh:
1256         kfree(pdh_cert);
1257 e_free_session:
1258         kfree(session_data);
1259         return ret;
1260 }
1261
1262 /* Userspace wants to query either header or trans length. */
1263 static int
1264 __sev_send_update_data_query_lengths(struct kvm *kvm, struct kvm_sev_cmd *argp,
1265                                      struct kvm_sev_send_update_data *params)
1266 {
1267         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1268         struct sev_data_send_update_data data;
1269         int ret;
1270
1271         memset(&data, 0, sizeof(data));
1272         data.handle = sev->handle;
1273         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1274
1275         params->hdr_len = data.hdr_len;
1276         params->trans_len = data.trans_len;
1277
1278         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1279                          sizeof(struct kvm_sev_send_update_data)))
1280                 ret = -EFAULT;
1281
1282         return ret;
1283 }
1284
1285 static int sev_send_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1286 {
1287         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1288         struct sev_data_send_update_data data;
1289         struct kvm_sev_send_update_data params;
1290         void *hdr, *trans_data;
1291         struct page **guest_page;
1292         unsigned long n;
1293         int ret, offset;
1294
1295         if (!sev_guest(kvm))
1296                 return -ENOTTY;
1297
1298         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1299                         sizeof(struct kvm_sev_send_update_data)))
1300                 return -EFAULT;
1301
1302         /* userspace wants to query either header or trans length */
1303         if (!params.trans_len || !params.hdr_len)
1304                 return __sev_send_update_data_query_lengths(kvm, argp, &params);
1305
1306         if (!params.trans_uaddr || !params.guest_uaddr ||
1307             !params.guest_len || !params.hdr_uaddr)
1308                 return -EINVAL;
1309
1310         /* Check if we are crossing the page boundary */
1311         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1312         if (params.guest_len > PAGE_SIZE || (params.guest_len + offset) > PAGE_SIZE)
1313                 return -EINVAL;
1314
1315         /* Pin guest memory */
1316         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1317                                     PAGE_SIZE, &n, 0);
1318         if (IS_ERR(guest_page))
1319                 return PTR_ERR(guest_page);
1320
1321         /* allocate memory for header and transport buffer */
1322         ret = -ENOMEM;
1323         hdr = kzalloc(params.hdr_len, GFP_KERNEL_ACCOUNT);
1324         if (!hdr)
1325                 goto e_unpin;
1326
1327         trans_data = kzalloc(params.trans_len, GFP_KERNEL_ACCOUNT);
1328         if (!trans_data)
1329                 goto e_free_hdr;
1330
1331         memset(&data, 0, sizeof(data));
1332         data.hdr_address = __psp_pa(hdr);
1333         data.hdr_len = params.hdr_len;
1334         data.trans_address = __psp_pa(trans_data);
1335         data.trans_len = params.trans_len;
1336
1337         /* The SEND_UPDATE_DATA command requires C-bit to be always set. */
1338         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1339         data.guest_address |= sev_me_mask;
1340         data.guest_len = params.guest_len;
1341         data.handle = sev->handle;
1342
1343         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1344
1345         if (ret)
1346                 goto e_free_trans_data;
1347
1348         /* copy transport buffer to user space */
1349         if (copy_to_user((void __user *)(uintptr_t)params.trans_uaddr,
1350                          trans_data, params.trans_len)) {
1351                 ret = -EFAULT;
1352                 goto e_free_trans_data;
1353         }
1354
1355         /* Copy packet header to userspace. */
1356         if (copy_to_user((void __user *)(uintptr_t)params.hdr_uaddr, hdr,
1357                          params.hdr_len))
1358                 ret = -EFAULT;
1359
1360 e_free_trans_data:
1361         kfree(trans_data);
1362 e_free_hdr:
1363         kfree(hdr);
1364 e_unpin:
1365         sev_unpin_memory(kvm, guest_page, n);
1366
1367         return ret;
1368 }
1369
1370 static int sev_send_finish(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_send_finish data;
1374
1375         if (!sev_guest(kvm))
1376                 return -ENOTTY;
1377
1378         data.handle = sev->handle;
1379         return sev_issue_cmd(kvm, SEV_CMD_SEND_FINISH, &data, &argp->error);
1380 }
1381
1382 static int sev_send_cancel(struct kvm *kvm, struct kvm_sev_cmd *argp)
1383 {
1384         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1385         struct sev_data_send_cancel data;
1386
1387         if (!sev_guest(kvm))
1388                 return -ENOTTY;
1389
1390         data.handle = sev->handle;
1391         return sev_issue_cmd(kvm, SEV_CMD_SEND_CANCEL, &data, &argp->error);
1392 }
1393
1394 static int sev_receive_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1395 {
1396         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1397         struct sev_data_receive_start start;
1398         struct kvm_sev_receive_start params;
1399         int *error = &argp->error;
1400         void *session_data;
1401         void *pdh_data;
1402         int ret;
1403
1404         if (!sev_guest(kvm))
1405                 return -ENOTTY;
1406
1407         /* Get parameter from the userspace */
1408         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1409                         sizeof(struct kvm_sev_receive_start)))
1410                 return -EFAULT;
1411
1412         /* some sanity checks */
1413         if (!params.pdh_uaddr || !params.pdh_len ||
1414             !params.session_uaddr || !params.session_len)
1415                 return -EINVAL;
1416
1417         pdh_data = psp_copy_user_blob(params.pdh_uaddr, params.pdh_len);
1418         if (IS_ERR(pdh_data))
1419                 return PTR_ERR(pdh_data);
1420
1421         session_data = psp_copy_user_blob(params.session_uaddr,
1422                         params.session_len);
1423         if (IS_ERR(session_data)) {
1424                 ret = PTR_ERR(session_data);
1425                 goto e_free_pdh;
1426         }
1427
1428         memset(&start, 0, sizeof(start));
1429         start.handle = params.handle;
1430         start.policy = params.policy;
1431         start.pdh_cert_address = __psp_pa(pdh_data);
1432         start.pdh_cert_len = params.pdh_len;
1433         start.session_address = __psp_pa(session_data);
1434         start.session_len = params.session_len;
1435
1436         /* create memory encryption context */
1437         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_RECEIVE_START, &start,
1438                                 error);
1439         if (ret)
1440                 goto e_free_session;
1441
1442         /* Bind ASID to this guest */
1443         ret = sev_bind_asid(kvm, start.handle, error);
1444         if (ret) {
1445                 sev_decommission(start.handle);
1446                 goto e_free_session;
1447         }
1448
1449         params.handle = start.handle;
1450         if (copy_to_user((void __user *)(uintptr_t)argp->data,
1451                          &params, sizeof(struct kvm_sev_receive_start))) {
1452                 ret = -EFAULT;
1453                 sev_unbind_asid(kvm, start.handle);
1454                 goto e_free_session;
1455         }
1456
1457         sev->handle = start.handle;
1458         sev->fd = argp->sev_fd;
1459
1460 e_free_session:
1461         kfree(session_data);
1462 e_free_pdh:
1463         kfree(pdh_data);
1464
1465         return ret;
1466 }
1467
1468 static int sev_receive_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1469 {
1470         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1471         struct kvm_sev_receive_update_data params;
1472         struct sev_data_receive_update_data data;
1473         void *hdr = NULL, *trans = NULL;
1474         struct page **guest_page;
1475         unsigned long n;
1476         int ret, offset;
1477
1478         if (!sev_guest(kvm))
1479                 return -EINVAL;
1480
1481         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1482                         sizeof(struct kvm_sev_receive_update_data)))
1483                 return -EFAULT;
1484
1485         if (!params.hdr_uaddr || !params.hdr_len ||
1486             !params.guest_uaddr || !params.guest_len ||
1487             !params.trans_uaddr || !params.trans_len)
1488                 return -EINVAL;
1489
1490         /* Check if we are crossing the page boundary */
1491         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1492         if (params.guest_len > PAGE_SIZE || (params.guest_len + offset) > PAGE_SIZE)
1493                 return -EINVAL;
1494
1495         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1496         if (IS_ERR(hdr))
1497                 return PTR_ERR(hdr);
1498
1499         trans = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1500         if (IS_ERR(trans)) {
1501                 ret = PTR_ERR(trans);
1502                 goto e_free_hdr;
1503         }
1504
1505         memset(&data, 0, sizeof(data));
1506         data.hdr_address = __psp_pa(hdr);
1507         data.hdr_len = params.hdr_len;
1508         data.trans_address = __psp_pa(trans);
1509         data.trans_len = params.trans_len;
1510
1511         /* Pin guest memory */
1512         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1513                                     PAGE_SIZE, &n, 1);
1514         if (IS_ERR(guest_page)) {
1515                 ret = PTR_ERR(guest_page);
1516                 goto e_free_trans;
1517         }
1518
1519         /*
1520          * Flush (on non-coherent CPUs) before RECEIVE_UPDATE_DATA, the PSP
1521          * encrypts the written data with the guest's key, and the cache may
1522          * contain dirty, unencrypted data.
1523          */
1524         sev_clflush_pages(guest_page, n);
1525
1526         /* The RECEIVE_UPDATE_DATA command requires C-bit to be always set. */
1527         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1528         data.guest_address |= sev_me_mask;
1529         data.guest_len = params.guest_len;
1530         data.handle = sev->handle;
1531
1532         ret = sev_issue_cmd(kvm, SEV_CMD_RECEIVE_UPDATE_DATA, &data,
1533                                 &argp->error);
1534
1535         sev_unpin_memory(kvm, guest_page, n);
1536
1537 e_free_trans:
1538         kfree(trans);
1539 e_free_hdr:
1540         kfree(hdr);
1541
1542         return ret;
1543 }
1544
1545 static int sev_receive_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1546 {
1547         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1548         struct sev_data_receive_finish data;
1549
1550         if (!sev_guest(kvm))
1551                 return -ENOTTY;
1552
1553         data.handle = sev->handle;
1554         return sev_issue_cmd(kvm, SEV_CMD_RECEIVE_FINISH, &data, &argp->error);
1555 }
1556
1557 static bool is_cmd_allowed_from_mirror(u32 cmd_id)
1558 {
1559         /*
1560          * Allow mirrors VM to call KVM_SEV_LAUNCH_UPDATE_VMSA to enable SEV-ES
1561          * active mirror VMs. Also allow the debugging and status commands.
1562          */
1563         if (cmd_id == KVM_SEV_LAUNCH_UPDATE_VMSA ||
1564             cmd_id == KVM_SEV_GUEST_STATUS || cmd_id == KVM_SEV_DBG_DECRYPT ||
1565             cmd_id == KVM_SEV_DBG_ENCRYPT)
1566                 return true;
1567
1568         return false;
1569 }
1570
1571 static int sev_lock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1572 {
1573         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1574         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1575         int r = -EBUSY;
1576
1577         if (dst_kvm == src_kvm)
1578                 return -EINVAL;
1579
1580         /*
1581          * Bail if these VMs are already involved in a migration to avoid
1582          * deadlock between two VMs trying to migrate to/from each other.
1583          */
1584         if (atomic_cmpxchg_acquire(&dst_sev->migration_in_progress, 0, 1))
1585                 return -EBUSY;
1586
1587         if (atomic_cmpxchg_acquire(&src_sev->migration_in_progress, 0, 1))
1588                 goto release_dst;
1589
1590         r = -EINTR;
1591         if (mutex_lock_killable(&dst_kvm->lock))
1592                 goto release_src;
1593         if (mutex_lock_killable_nested(&src_kvm->lock, SINGLE_DEPTH_NESTING))
1594                 goto unlock_dst;
1595         return 0;
1596
1597 unlock_dst:
1598         mutex_unlock(&dst_kvm->lock);
1599 release_src:
1600         atomic_set_release(&src_sev->migration_in_progress, 0);
1601 release_dst:
1602         atomic_set_release(&dst_sev->migration_in_progress, 0);
1603         return r;
1604 }
1605
1606 static void sev_unlock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1607 {
1608         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1609         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1610
1611         mutex_unlock(&dst_kvm->lock);
1612         mutex_unlock(&src_kvm->lock);
1613         atomic_set_release(&dst_sev->migration_in_progress, 0);
1614         atomic_set_release(&src_sev->migration_in_progress, 0);
1615 }
1616
1617 /* vCPU mutex subclasses.  */
1618 enum sev_migration_role {
1619         SEV_MIGRATION_SOURCE = 0,
1620         SEV_MIGRATION_TARGET,
1621         SEV_NR_MIGRATION_ROLES,
1622 };
1623
1624 static int sev_lock_vcpus_for_migration(struct kvm *kvm,
1625                                         enum sev_migration_role role)
1626 {
1627         struct kvm_vcpu *vcpu;
1628         unsigned long i, j;
1629
1630         kvm_for_each_vcpu(i, vcpu, kvm) {
1631                 if (mutex_lock_killable_nested(&vcpu->mutex, role))
1632                         goto out_unlock;
1633
1634 #ifdef CONFIG_PROVE_LOCKING
1635                 if (!i)
1636                         /*
1637                          * Reset the role to one that avoids colliding with
1638                          * the role used for the first vcpu mutex.
1639                          */
1640                         role = SEV_NR_MIGRATION_ROLES;
1641                 else
1642                         mutex_release(&vcpu->mutex.dep_map, _THIS_IP_);
1643 #endif
1644         }
1645
1646         return 0;
1647
1648 out_unlock:
1649
1650         kvm_for_each_vcpu(j, vcpu, kvm) {
1651                 if (i == j)
1652                         break;
1653
1654 #ifdef CONFIG_PROVE_LOCKING
1655                 if (j)
1656                         mutex_acquire(&vcpu->mutex.dep_map, role, 0, _THIS_IP_);
1657 #endif
1658
1659                 mutex_unlock(&vcpu->mutex);
1660         }
1661         return -EINTR;
1662 }
1663
1664 static void sev_unlock_vcpus_for_migration(struct kvm *kvm)
1665 {
1666         struct kvm_vcpu *vcpu;
1667         unsigned long i;
1668         bool first = true;
1669
1670         kvm_for_each_vcpu(i, vcpu, kvm) {
1671                 if (first)
1672                         first = false;
1673                 else
1674                         mutex_acquire(&vcpu->mutex.dep_map,
1675                                       SEV_NR_MIGRATION_ROLES, 0, _THIS_IP_);
1676
1677                 mutex_unlock(&vcpu->mutex);
1678         }
1679 }
1680
1681 static void sev_migrate_from(struct kvm *dst_kvm, struct kvm *src_kvm)
1682 {
1683         struct kvm_sev_info *dst = &to_kvm_svm(dst_kvm)->sev_info;
1684         struct kvm_sev_info *src = &to_kvm_svm(src_kvm)->sev_info;
1685         struct kvm_vcpu *dst_vcpu, *src_vcpu;
1686         struct vcpu_svm *dst_svm, *src_svm;
1687         struct kvm_sev_info *mirror;
1688         unsigned long i;
1689
1690         dst->active = true;
1691         dst->asid = src->asid;
1692         dst->handle = src->handle;
1693         dst->pages_locked = src->pages_locked;
1694         dst->enc_context_owner = src->enc_context_owner;
1695         dst->es_active = src->es_active;
1696
1697         src->asid = 0;
1698         src->active = false;
1699         src->handle = 0;
1700         src->pages_locked = 0;
1701         src->enc_context_owner = NULL;
1702         src->es_active = false;
1703
1704         list_cut_before(&dst->regions_list, &src->regions_list, &src->regions_list);
1705
1706         /*
1707          * If this VM has mirrors, "transfer" each mirror's refcount of the
1708          * source to the destination (this KVM).  The caller holds a reference
1709          * to the source, so there's no danger of use-after-free.
1710          */
1711         list_cut_before(&dst->mirror_vms, &src->mirror_vms, &src->mirror_vms);
1712         list_for_each_entry(mirror, &dst->mirror_vms, mirror_entry) {
1713                 kvm_get_kvm(dst_kvm);
1714                 kvm_put_kvm(src_kvm);
1715                 mirror->enc_context_owner = dst_kvm;
1716         }
1717
1718         /*
1719          * If this VM is a mirror, remove the old mirror from the owners list
1720          * and add the new mirror to the list.
1721          */
1722         if (is_mirroring_enc_context(dst_kvm)) {
1723                 struct kvm_sev_info *owner_sev_info =
1724                         &to_kvm_svm(dst->enc_context_owner)->sev_info;
1725
1726                 list_del(&src->mirror_entry);
1727                 list_add_tail(&dst->mirror_entry, &owner_sev_info->mirror_vms);
1728         }
1729
1730         kvm_for_each_vcpu(i, dst_vcpu, dst_kvm) {
1731                 dst_svm = to_svm(dst_vcpu);
1732
1733                 sev_init_vmcb(dst_svm);
1734
1735                 if (!dst->es_active)
1736                         continue;
1737
1738                 /*
1739                  * Note, the source is not required to have the same number of
1740                  * vCPUs as the destination when migrating a vanilla SEV VM.
1741                  */
1742                 src_vcpu = kvm_get_vcpu(src_kvm, i);
1743                 src_svm = to_svm(src_vcpu);
1744
1745                 /*
1746                  * Transfer VMSA and GHCB state to the destination.  Nullify and
1747                  * clear source fields as appropriate, the state now belongs to
1748                  * the destination.
1749                  */
1750                 memcpy(&dst_svm->sev_es, &src_svm->sev_es, sizeof(src_svm->sev_es));
1751                 dst_svm->vmcb->control.ghcb_gpa = src_svm->vmcb->control.ghcb_gpa;
1752                 dst_svm->vmcb->control.vmsa_pa = src_svm->vmcb->control.vmsa_pa;
1753                 dst_vcpu->arch.guest_state_protected = true;
1754
1755                 memset(&src_svm->sev_es, 0, sizeof(src_svm->sev_es));
1756                 src_svm->vmcb->control.ghcb_gpa = INVALID_PAGE;
1757                 src_svm->vmcb->control.vmsa_pa = INVALID_PAGE;
1758                 src_vcpu->arch.guest_state_protected = false;
1759         }
1760 }
1761
1762 static int sev_check_source_vcpus(struct kvm *dst, struct kvm *src)
1763 {
1764         struct kvm_vcpu *src_vcpu;
1765         unsigned long i;
1766
1767         if (!sev_es_guest(src))
1768                 return 0;
1769
1770         if (atomic_read(&src->online_vcpus) != atomic_read(&dst->online_vcpus))
1771                 return -EINVAL;
1772
1773         kvm_for_each_vcpu(i, src_vcpu, src) {
1774                 if (!src_vcpu->arch.guest_state_protected)
1775                         return -EINVAL;
1776         }
1777
1778         return 0;
1779 }
1780
1781 int sev_vm_move_enc_context_from(struct kvm *kvm, unsigned int source_fd)
1782 {
1783         struct kvm_sev_info *dst_sev = &to_kvm_svm(kvm)->sev_info;
1784         struct kvm_sev_info *src_sev, *cg_cleanup_sev;
1785         struct fd f = fdget(source_fd);
1786         struct kvm *source_kvm;
1787         bool charged = false;
1788         int ret;
1789
1790         if (!f.file)
1791                 return -EBADF;
1792
1793         if (!file_is_kvm(f.file)) {
1794                 ret = -EBADF;
1795                 goto out_fput;
1796         }
1797
1798         source_kvm = f.file->private_data;
1799         ret = sev_lock_two_vms(kvm, source_kvm);
1800         if (ret)
1801                 goto out_fput;
1802
1803         if (sev_guest(kvm) || !sev_guest(source_kvm)) {
1804                 ret = -EINVAL;
1805                 goto out_unlock;
1806         }
1807
1808         src_sev = &to_kvm_svm(source_kvm)->sev_info;
1809
1810         dst_sev->misc_cg = get_current_misc_cg();
1811         cg_cleanup_sev = dst_sev;
1812         if (dst_sev->misc_cg != src_sev->misc_cg) {
1813                 ret = sev_misc_cg_try_charge(dst_sev);
1814                 if (ret)
1815                         goto out_dst_cgroup;
1816                 charged = true;
1817         }
1818
1819         ret = sev_lock_vcpus_for_migration(kvm, SEV_MIGRATION_SOURCE);
1820         if (ret)
1821                 goto out_dst_cgroup;
1822         ret = sev_lock_vcpus_for_migration(source_kvm, SEV_MIGRATION_TARGET);
1823         if (ret)
1824                 goto out_dst_vcpu;
1825
1826         ret = sev_check_source_vcpus(kvm, source_kvm);
1827         if (ret)
1828                 goto out_source_vcpu;
1829
1830         sev_migrate_from(kvm, source_kvm);
1831         kvm_vm_dead(source_kvm);
1832         cg_cleanup_sev = src_sev;
1833         ret = 0;
1834
1835 out_source_vcpu:
1836         sev_unlock_vcpus_for_migration(source_kvm);
1837 out_dst_vcpu:
1838         sev_unlock_vcpus_for_migration(kvm);
1839 out_dst_cgroup:
1840         /* Operates on the source on success, on the destination on failure.  */
1841         if (charged)
1842                 sev_misc_cg_uncharge(cg_cleanup_sev);
1843         put_misc_cg(cg_cleanup_sev->misc_cg);
1844         cg_cleanup_sev->misc_cg = NULL;
1845 out_unlock:
1846         sev_unlock_two_vms(kvm, source_kvm);
1847 out_fput:
1848         fdput(f);
1849         return ret;
1850 }
1851
1852 int sev_mem_enc_ioctl(struct kvm *kvm, void __user *argp)
1853 {
1854         struct kvm_sev_cmd sev_cmd;
1855         int r;
1856
1857         if (!sev_enabled)
1858                 return -ENOTTY;
1859
1860         if (!argp)
1861                 return 0;
1862
1863         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
1864                 return -EFAULT;
1865
1866         mutex_lock(&kvm->lock);
1867
1868         /* Only the enc_context_owner handles some memory enc operations. */
1869         if (is_mirroring_enc_context(kvm) &&
1870             !is_cmd_allowed_from_mirror(sev_cmd.id)) {
1871                 r = -EINVAL;
1872                 goto out;
1873         }
1874
1875         switch (sev_cmd.id) {
1876         case KVM_SEV_ES_INIT:
1877                 if (!sev_es_enabled) {
1878                         r = -ENOTTY;
1879                         goto out;
1880                 }
1881                 fallthrough;
1882         case KVM_SEV_INIT:
1883                 r = sev_guest_init(kvm, &sev_cmd);
1884                 break;
1885         case KVM_SEV_LAUNCH_START:
1886                 r = sev_launch_start(kvm, &sev_cmd);
1887                 break;
1888         case KVM_SEV_LAUNCH_UPDATE_DATA:
1889                 r = sev_launch_update_data(kvm, &sev_cmd);
1890                 break;
1891         case KVM_SEV_LAUNCH_UPDATE_VMSA:
1892                 r = sev_launch_update_vmsa(kvm, &sev_cmd);
1893                 break;
1894         case KVM_SEV_LAUNCH_MEASURE:
1895                 r = sev_launch_measure(kvm, &sev_cmd);
1896                 break;
1897         case KVM_SEV_LAUNCH_FINISH:
1898                 r = sev_launch_finish(kvm, &sev_cmd);
1899                 break;
1900         case KVM_SEV_GUEST_STATUS:
1901                 r = sev_guest_status(kvm, &sev_cmd);
1902                 break;
1903         case KVM_SEV_DBG_DECRYPT:
1904                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
1905                 break;
1906         case KVM_SEV_DBG_ENCRYPT:
1907                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
1908                 break;
1909         case KVM_SEV_LAUNCH_SECRET:
1910                 r = sev_launch_secret(kvm, &sev_cmd);
1911                 break;
1912         case KVM_SEV_GET_ATTESTATION_REPORT:
1913                 r = sev_get_attestation_report(kvm, &sev_cmd);
1914                 break;
1915         case KVM_SEV_SEND_START:
1916                 r = sev_send_start(kvm, &sev_cmd);
1917                 break;
1918         case KVM_SEV_SEND_UPDATE_DATA:
1919                 r = sev_send_update_data(kvm, &sev_cmd);
1920                 break;
1921         case KVM_SEV_SEND_FINISH:
1922                 r = sev_send_finish(kvm, &sev_cmd);
1923                 break;
1924         case KVM_SEV_SEND_CANCEL:
1925                 r = sev_send_cancel(kvm, &sev_cmd);
1926                 break;
1927         case KVM_SEV_RECEIVE_START:
1928                 r = sev_receive_start(kvm, &sev_cmd);
1929                 break;
1930         case KVM_SEV_RECEIVE_UPDATE_DATA:
1931                 r = sev_receive_update_data(kvm, &sev_cmd);
1932                 break;
1933         case KVM_SEV_RECEIVE_FINISH:
1934                 r = sev_receive_finish(kvm, &sev_cmd);
1935                 break;
1936         default:
1937                 r = -EINVAL;
1938                 goto out;
1939         }
1940
1941         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
1942                 r = -EFAULT;
1943
1944 out:
1945         mutex_unlock(&kvm->lock);
1946         return r;
1947 }
1948
1949 int sev_mem_enc_register_region(struct kvm *kvm,
1950                                 struct kvm_enc_region *range)
1951 {
1952         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1953         struct enc_region *region;
1954         int ret = 0;
1955
1956         if (!sev_guest(kvm))
1957                 return -ENOTTY;
1958
1959         /* If kvm is mirroring encryption context it isn't responsible for it */
1960         if (is_mirroring_enc_context(kvm))
1961                 return -EINVAL;
1962
1963         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
1964                 return -EINVAL;
1965
1966         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
1967         if (!region)
1968                 return -ENOMEM;
1969
1970         mutex_lock(&kvm->lock);
1971         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
1972         if (IS_ERR(region->pages)) {
1973                 ret = PTR_ERR(region->pages);
1974                 mutex_unlock(&kvm->lock);
1975                 goto e_free;
1976         }
1977
1978         region->uaddr = range->addr;
1979         region->size = range->size;
1980
1981         list_add_tail(&region->list, &sev->regions_list);
1982         mutex_unlock(&kvm->lock);
1983
1984         /*
1985          * The guest may change the memory encryption attribute from C=0 -> C=1
1986          * or vice versa for this memory range. Lets make sure caches are
1987          * flushed to ensure that guest data gets written into memory with
1988          * correct C-bit.
1989          */
1990         sev_clflush_pages(region->pages, region->npages);
1991
1992         return ret;
1993
1994 e_free:
1995         kfree(region);
1996         return ret;
1997 }
1998
1999 static struct enc_region *
2000 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
2001 {
2002         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
2003         struct list_head *head = &sev->regions_list;
2004         struct enc_region *i;
2005
2006         list_for_each_entry(i, head, list) {
2007                 if (i->uaddr == range->addr &&
2008                     i->size == range->size)
2009                         return i;
2010         }
2011
2012         return NULL;
2013 }
2014
2015 static void __unregister_enc_region_locked(struct kvm *kvm,
2016                                            struct enc_region *region)
2017 {
2018         sev_unpin_memory(kvm, region->pages, region->npages);
2019         list_del(&region->list);
2020         kfree(region);
2021 }
2022
2023 int sev_mem_enc_unregister_region(struct kvm *kvm,
2024                                   struct kvm_enc_region *range)
2025 {
2026         struct enc_region *region;
2027         int ret;
2028
2029         /* If kvm is mirroring encryption context it isn't responsible for it */
2030         if (is_mirroring_enc_context(kvm))
2031                 return -EINVAL;
2032
2033         mutex_lock(&kvm->lock);
2034
2035         if (!sev_guest(kvm)) {
2036                 ret = -ENOTTY;
2037                 goto failed;
2038         }
2039
2040         region = find_enc_region(kvm, range);
2041         if (!region) {
2042                 ret = -EINVAL;
2043                 goto failed;
2044         }
2045
2046         /*
2047          * Ensure that all guest tagged cache entries are flushed before
2048          * releasing the pages back to the system for use. CLFLUSH will
2049          * not do this, so issue a WBINVD.
2050          */
2051         wbinvd_on_all_cpus();
2052
2053         __unregister_enc_region_locked(kvm, region);
2054
2055         mutex_unlock(&kvm->lock);
2056         return 0;
2057
2058 failed:
2059         mutex_unlock(&kvm->lock);
2060         return ret;
2061 }
2062
2063 int sev_vm_copy_enc_context_from(struct kvm *kvm, unsigned int source_fd)
2064 {
2065         struct fd f = fdget(source_fd);
2066         struct kvm *source_kvm;
2067         struct kvm_sev_info *source_sev, *mirror_sev;
2068         int ret;
2069
2070         if (!f.file)
2071                 return -EBADF;
2072
2073         if (!file_is_kvm(f.file)) {
2074                 ret = -EBADF;
2075                 goto e_source_fput;
2076         }
2077
2078         source_kvm = f.file->private_data;
2079         ret = sev_lock_two_vms(kvm, source_kvm);
2080         if (ret)
2081                 goto e_source_fput;
2082
2083         /*
2084          * Mirrors of mirrors should work, but let's not get silly.  Also
2085          * disallow out-of-band SEV/SEV-ES init if the target is already an
2086          * SEV guest, or if vCPUs have been created.  KVM relies on vCPUs being
2087          * created after SEV/SEV-ES initialization, e.g. to init intercepts.
2088          */
2089         if (sev_guest(kvm) || !sev_guest(source_kvm) ||
2090             is_mirroring_enc_context(source_kvm) || kvm->created_vcpus) {
2091                 ret = -EINVAL;
2092                 goto e_unlock;
2093         }
2094
2095         /*
2096          * The mirror kvm holds an enc_context_owner ref so its asid can't
2097          * disappear until we're done with it
2098          */
2099         source_sev = &to_kvm_svm(source_kvm)->sev_info;
2100         kvm_get_kvm(source_kvm);
2101         mirror_sev = &to_kvm_svm(kvm)->sev_info;
2102         list_add_tail(&mirror_sev->mirror_entry, &source_sev->mirror_vms);
2103
2104         /* Set enc_context_owner and copy its encryption context over */
2105         mirror_sev->enc_context_owner = source_kvm;
2106         mirror_sev->active = true;
2107         mirror_sev->asid = source_sev->asid;
2108         mirror_sev->fd = source_sev->fd;
2109         mirror_sev->es_active = source_sev->es_active;
2110         mirror_sev->handle = source_sev->handle;
2111         INIT_LIST_HEAD(&mirror_sev->regions_list);
2112         INIT_LIST_HEAD(&mirror_sev->mirror_vms);
2113         ret = 0;
2114
2115         /*
2116          * Do not copy ap_jump_table. Since the mirror does not share the same
2117          * KVM contexts as the original, and they may have different
2118          * memory-views.
2119          */
2120
2121 e_unlock:
2122         sev_unlock_two_vms(kvm, source_kvm);
2123 e_source_fput:
2124         fdput(f);
2125         return ret;
2126 }
2127
2128 void sev_vm_destroy(struct kvm *kvm)
2129 {
2130         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
2131         struct list_head *head = &sev->regions_list;
2132         struct list_head *pos, *q;
2133
2134         if (!sev_guest(kvm))
2135                 return;
2136
2137         WARN_ON(!list_empty(&sev->mirror_vms));
2138
2139         /* If this is a mirror_kvm release the enc_context_owner and skip sev cleanup */
2140         if (is_mirroring_enc_context(kvm)) {
2141                 struct kvm *owner_kvm = sev->enc_context_owner;
2142
2143                 mutex_lock(&owner_kvm->lock);
2144                 list_del(&sev->mirror_entry);
2145                 mutex_unlock(&owner_kvm->lock);
2146                 kvm_put_kvm(owner_kvm);
2147                 return;
2148         }
2149
2150         /*
2151          * Ensure that all guest tagged cache entries are flushed before
2152          * releasing the pages back to the system for use. CLFLUSH will
2153          * not do this, so issue a WBINVD.
2154          */
2155         wbinvd_on_all_cpus();
2156
2157         /*
2158          * if userspace was terminated before unregistering the memory regions
2159          * then lets unpin all the registered memory.
2160          */
2161         if (!list_empty(head)) {
2162                 list_for_each_safe(pos, q, head) {
2163                         __unregister_enc_region_locked(kvm,
2164                                 list_entry(pos, struct enc_region, list));
2165                         cond_resched();
2166                 }
2167         }
2168
2169         sev_unbind_asid(kvm, sev->handle);
2170         sev_asid_free(sev);
2171 }
2172
2173 void __init sev_set_cpu_caps(void)
2174 {
2175         if (!sev_enabled)
2176                 kvm_cpu_cap_clear(X86_FEATURE_SEV);
2177         if (!sev_es_enabled)
2178                 kvm_cpu_cap_clear(X86_FEATURE_SEV_ES);
2179 }
2180
2181 void __init sev_hardware_setup(void)
2182 {
2183 #ifdef CONFIG_KVM_AMD_SEV
2184         unsigned int eax, ebx, ecx, edx, sev_asid_count, sev_es_asid_count;
2185         bool sev_es_supported = false;
2186         bool sev_supported = false;
2187
2188         if (!sev_enabled || !npt_enabled || !nrips)
2189                 goto out;
2190
2191         /*
2192          * SEV must obviously be supported in hardware.  Sanity check that the
2193          * CPU supports decode assists, which is mandatory for SEV guests to
2194          * support instruction emulation.
2195          */
2196         if (!boot_cpu_has(X86_FEATURE_SEV) ||
2197             WARN_ON_ONCE(!boot_cpu_has(X86_FEATURE_DECODEASSISTS)))
2198                 goto out;
2199
2200         /* Retrieve SEV CPUID information */
2201         cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
2202
2203         /* Set encryption bit location for SEV-ES guests */
2204         sev_enc_bit = ebx & 0x3f;
2205
2206         /* Maximum number of encrypted guests supported simultaneously */
2207         max_sev_asid = ecx;
2208         if (!max_sev_asid)
2209                 goto out;
2210
2211         /* Minimum ASID value that should be used for SEV guest */
2212         min_sev_asid = edx;
2213         sev_me_mask = 1UL << (ebx & 0x3f);
2214
2215         /*
2216          * Initialize SEV ASID bitmaps. Allocate space for ASID 0 in the bitmap,
2217          * even though it's never used, so that the bitmap is indexed by the
2218          * actual ASID.
2219          */
2220         nr_asids = max_sev_asid + 1;
2221         sev_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2222         if (!sev_asid_bitmap)
2223                 goto out;
2224
2225         sev_reclaim_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2226         if (!sev_reclaim_asid_bitmap) {
2227                 bitmap_free(sev_asid_bitmap);
2228                 sev_asid_bitmap = NULL;
2229                 goto out;
2230         }
2231
2232         sev_asid_count = max_sev_asid - min_sev_asid + 1;
2233         WARN_ON_ONCE(misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count));
2234         sev_supported = true;
2235
2236         /* SEV-ES support requested? */
2237         if (!sev_es_enabled)
2238                 goto out;
2239
2240         /*
2241          * SEV-ES requires MMIO caching as KVM doesn't have access to the guest
2242          * instruction stream, i.e. can't emulate in response to a #NPF and
2243          * instead relies on #NPF(RSVD) being reflected into the guest as #VC
2244          * (the guest can then do a #VMGEXIT to request MMIO emulation).
2245          */
2246         if (!enable_mmio_caching)
2247                 goto out;
2248
2249         /* Does the CPU support SEV-ES? */
2250         if (!boot_cpu_has(X86_FEATURE_SEV_ES))
2251                 goto out;
2252
2253         /* Has the system been allocated ASIDs for SEV-ES? */
2254         if (min_sev_asid == 1)
2255                 goto out;
2256
2257         sev_es_asid_count = min_sev_asid - 1;
2258         WARN_ON_ONCE(misc_cg_set_capacity(MISC_CG_RES_SEV_ES, sev_es_asid_count));
2259         sev_es_supported = true;
2260
2261 out:
2262         if (boot_cpu_has(X86_FEATURE_SEV))
2263                 pr_info("SEV %s (ASIDs %u - %u)\n",
2264                         sev_supported ? "enabled" : "disabled",
2265                         min_sev_asid, max_sev_asid);
2266         if (boot_cpu_has(X86_FEATURE_SEV_ES))
2267                 pr_info("SEV-ES %s (ASIDs %u - %u)\n",
2268                         sev_es_supported ? "enabled" : "disabled",
2269                         min_sev_asid > 1 ? 1 : 0, min_sev_asid - 1);
2270
2271         sev_enabled = sev_supported;
2272         sev_es_enabled = sev_es_supported;
2273         if (!sev_es_enabled || !cpu_feature_enabled(X86_FEATURE_DEBUG_SWAP) ||
2274             !cpu_feature_enabled(X86_FEATURE_NO_NESTED_DATA_BP))
2275                 sev_es_debug_swap_enabled = false;
2276 #endif
2277 }
2278
2279 void sev_hardware_unsetup(void)
2280 {
2281         if (!sev_enabled)
2282                 return;
2283
2284         /* No need to take sev_bitmap_lock, all VMs have been destroyed. */
2285         sev_flush_asids(1, max_sev_asid);
2286
2287         bitmap_free(sev_asid_bitmap);
2288         bitmap_free(sev_reclaim_asid_bitmap);
2289
2290         misc_cg_set_capacity(MISC_CG_RES_SEV, 0);
2291         misc_cg_set_capacity(MISC_CG_RES_SEV_ES, 0);
2292 }
2293
2294 int sev_cpu_init(struct svm_cpu_data *sd)
2295 {
2296         if (!sev_enabled)
2297                 return 0;
2298
2299         sd->sev_vmcbs = kcalloc(nr_asids, sizeof(void *), GFP_KERNEL);
2300         if (!sd->sev_vmcbs)
2301                 return -ENOMEM;
2302
2303         return 0;
2304 }
2305
2306 /*
2307  * Pages used by hardware to hold guest encrypted state must be flushed before
2308  * returning them to the system.
2309  */
2310 static void sev_flush_encrypted_page(struct kvm_vcpu *vcpu, void *va)
2311 {
2312         int asid = to_kvm_svm(vcpu->kvm)->sev_info.asid;
2313
2314         /*
2315          * Note!  The address must be a kernel address, as regular page walk
2316          * checks are performed by VM_PAGE_FLUSH, i.e. operating on a user
2317          * address is non-deterministic and unsafe.  This function deliberately
2318          * takes a pointer to deter passing in a user address.
2319          */
2320         unsigned long addr = (unsigned long)va;
2321
2322         /*
2323          * If CPU enforced cache coherency for encrypted mappings of the
2324          * same physical page is supported, use CLFLUSHOPT instead. NOTE: cache
2325          * flush is still needed in order to work properly with DMA devices.
2326          */
2327         if (boot_cpu_has(X86_FEATURE_SME_COHERENT)) {
2328                 clflush_cache_range(va, PAGE_SIZE);
2329                 return;
2330         }
2331
2332         /*
2333          * VM Page Flush takes a host virtual address and a guest ASID.  Fall
2334          * back to WBINVD if this faults so as not to make any problems worse
2335          * by leaving stale encrypted data in the cache.
2336          */
2337         if (WARN_ON_ONCE(wrmsrl_safe(MSR_AMD64_VM_PAGE_FLUSH, addr | asid)))
2338                 goto do_wbinvd;
2339
2340         return;
2341
2342 do_wbinvd:
2343         wbinvd_on_all_cpus();
2344 }
2345
2346 void sev_guest_memory_reclaimed(struct kvm *kvm)
2347 {
2348         if (!sev_guest(kvm))
2349                 return;
2350
2351         wbinvd_on_all_cpus();
2352 }
2353
2354 void sev_free_vcpu(struct kvm_vcpu *vcpu)
2355 {
2356         struct vcpu_svm *svm;
2357
2358         if (!sev_es_guest(vcpu->kvm))
2359                 return;
2360
2361         svm = to_svm(vcpu);
2362
2363         if (vcpu->arch.guest_state_protected)
2364                 sev_flush_encrypted_page(vcpu, svm->sev_es.vmsa);
2365
2366         __free_page(virt_to_page(svm->sev_es.vmsa));
2367
2368         if (svm->sev_es.ghcb_sa_free)
2369                 kvfree(svm->sev_es.ghcb_sa);
2370 }
2371
2372 static void dump_ghcb(struct vcpu_svm *svm)
2373 {
2374         struct ghcb *ghcb = svm->sev_es.ghcb;
2375         unsigned int nbits;
2376
2377         /* Re-use the dump_invalid_vmcb module parameter */
2378         if (!dump_invalid_vmcb) {
2379                 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
2380                 return;
2381         }
2382
2383         nbits = sizeof(ghcb->save.valid_bitmap) * 8;
2384
2385         pr_err("GHCB (GPA=%016llx):\n", svm->vmcb->control.ghcb_gpa);
2386         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
2387                ghcb->save.sw_exit_code, ghcb_sw_exit_code_is_valid(ghcb));
2388         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
2389                ghcb->save.sw_exit_info_1, ghcb_sw_exit_info_1_is_valid(ghcb));
2390         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
2391                ghcb->save.sw_exit_info_2, ghcb_sw_exit_info_2_is_valid(ghcb));
2392         pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
2393                ghcb->save.sw_scratch, ghcb_sw_scratch_is_valid(ghcb));
2394         pr_err("%-20s%*pb\n", "valid_bitmap", nbits, ghcb->save.valid_bitmap);
2395 }
2396
2397 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
2398 {
2399         struct kvm_vcpu *vcpu = &svm->vcpu;
2400         struct ghcb *ghcb = svm->sev_es.ghcb;
2401
2402         /*
2403          * The GHCB protocol so far allows for the following data
2404          * to be returned:
2405          *   GPRs RAX, RBX, RCX, RDX
2406          *
2407          * Copy their values, even if they may not have been written during the
2408          * VM-Exit.  It's the guest's responsibility to not consume random data.
2409          */
2410         ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
2411         ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
2412         ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
2413         ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
2414 }
2415
2416 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
2417 {
2418         struct vmcb_control_area *control = &svm->vmcb->control;
2419         struct kvm_vcpu *vcpu = &svm->vcpu;
2420         struct ghcb *ghcb = svm->sev_es.ghcb;
2421         u64 exit_code;
2422
2423         /*
2424          * The GHCB protocol so far allows for the following data
2425          * to be supplied:
2426          *   GPRs RAX, RBX, RCX, RDX
2427          *   XCR0
2428          *   CPL
2429          *
2430          * VMMCALL allows the guest to provide extra registers. KVM also
2431          * expects RSI for hypercalls, so include that, too.
2432          *
2433          * Copy their values to the appropriate location if supplied.
2434          */
2435         memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
2436
2437         BUILD_BUG_ON(sizeof(svm->sev_es.valid_bitmap) != sizeof(ghcb->save.valid_bitmap));
2438         memcpy(&svm->sev_es.valid_bitmap, &ghcb->save.valid_bitmap, sizeof(ghcb->save.valid_bitmap));
2439
2440         vcpu->arch.regs[VCPU_REGS_RAX] = kvm_ghcb_get_rax_if_valid(svm, ghcb);
2441         vcpu->arch.regs[VCPU_REGS_RBX] = kvm_ghcb_get_rbx_if_valid(svm, ghcb);
2442         vcpu->arch.regs[VCPU_REGS_RCX] = kvm_ghcb_get_rcx_if_valid(svm, ghcb);
2443         vcpu->arch.regs[VCPU_REGS_RDX] = kvm_ghcb_get_rdx_if_valid(svm, ghcb);
2444         vcpu->arch.regs[VCPU_REGS_RSI] = kvm_ghcb_get_rsi_if_valid(svm, ghcb);
2445
2446         svm->vmcb->save.cpl = kvm_ghcb_get_cpl_if_valid(svm, ghcb);
2447
2448         if (kvm_ghcb_xcr0_is_valid(svm)) {
2449                 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
2450                 kvm_update_cpuid_runtime(vcpu);
2451         }
2452
2453         /* Copy the GHCB exit information into the VMCB fields */
2454         exit_code = ghcb_get_sw_exit_code(ghcb);
2455         control->exit_code = lower_32_bits(exit_code);
2456         control->exit_code_hi = upper_32_bits(exit_code);
2457         control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
2458         control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
2459         svm->sev_es.sw_scratch = kvm_ghcb_get_sw_scratch_if_valid(svm, ghcb);
2460
2461         /* Clear the valid entries fields */
2462         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
2463 }
2464
2465 static u64 kvm_ghcb_get_sw_exit_code(struct vmcb_control_area *control)
2466 {
2467         return (((u64)control->exit_code_hi) << 32) | control->exit_code;
2468 }
2469
2470 static int sev_es_validate_vmgexit(struct vcpu_svm *svm)
2471 {
2472         struct vmcb_control_area *control = &svm->vmcb->control;
2473         struct kvm_vcpu *vcpu = &svm->vcpu;
2474         u64 exit_code;
2475         u64 reason;
2476
2477         /*
2478          * Retrieve the exit code now even though it may not be marked valid
2479          * as it could help with debugging.
2480          */
2481         exit_code = kvm_ghcb_get_sw_exit_code(control);
2482
2483         /* Only GHCB Usage code 0 is supported */
2484         if (svm->sev_es.ghcb->ghcb_usage) {
2485                 reason = GHCB_ERR_INVALID_USAGE;
2486                 goto vmgexit_err;
2487         }
2488
2489         reason = GHCB_ERR_MISSING_INPUT;
2490
2491         if (!kvm_ghcb_sw_exit_code_is_valid(svm) ||
2492             !kvm_ghcb_sw_exit_info_1_is_valid(svm) ||
2493             !kvm_ghcb_sw_exit_info_2_is_valid(svm))
2494                 goto vmgexit_err;
2495
2496         switch (exit_code) {
2497         case SVM_EXIT_READ_DR7:
2498                 break;
2499         case SVM_EXIT_WRITE_DR7:
2500                 if (!kvm_ghcb_rax_is_valid(svm))
2501                         goto vmgexit_err;
2502                 break;
2503         case SVM_EXIT_RDTSC:
2504                 break;
2505         case SVM_EXIT_RDPMC:
2506                 if (!kvm_ghcb_rcx_is_valid(svm))
2507                         goto vmgexit_err;
2508                 break;
2509         case SVM_EXIT_CPUID:
2510                 if (!kvm_ghcb_rax_is_valid(svm) ||
2511                     !kvm_ghcb_rcx_is_valid(svm))
2512                         goto vmgexit_err;
2513                 if (vcpu->arch.regs[VCPU_REGS_RAX] == 0xd)
2514                         if (!kvm_ghcb_xcr0_is_valid(svm))
2515                                 goto vmgexit_err;
2516                 break;
2517         case SVM_EXIT_INVD:
2518                 break;
2519         case SVM_EXIT_IOIO:
2520                 if (control->exit_info_1 & SVM_IOIO_STR_MASK) {
2521                         if (!kvm_ghcb_sw_scratch_is_valid(svm))
2522                                 goto vmgexit_err;
2523                 } else {
2524                         if (!(control->exit_info_1 & SVM_IOIO_TYPE_MASK))
2525                                 if (!kvm_ghcb_rax_is_valid(svm))
2526                                         goto vmgexit_err;
2527                 }
2528                 break;
2529         case SVM_EXIT_MSR:
2530                 if (!kvm_ghcb_rcx_is_valid(svm))
2531                         goto vmgexit_err;
2532                 if (control->exit_info_1) {
2533                         if (!kvm_ghcb_rax_is_valid(svm) ||
2534                             !kvm_ghcb_rdx_is_valid(svm))
2535                                 goto vmgexit_err;
2536                 }
2537                 break;
2538         case SVM_EXIT_VMMCALL:
2539                 if (!kvm_ghcb_rax_is_valid(svm) ||
2540                     !kvm_ghcb_cpl_is_valid(svm))
2541                         goto vmgexit_err;
2542                 break;
2543         case SVM_EXIT_RDTSCP:
2544                 break;
2545         case SVM_EXIT_WBINVD:
2546                 break;
2547         case SVM_EXIT_MONITOR:
2548                 if (!kvm_ghcb_rax_is_valid(svm) ||
2549                     !kvm_ghcb_rcx_is_valid(svm) ||
2550                     !kvm_ghcb_rdx_is_valid(svm))
2551                         goto vmgexit_err;
2552                 break;
2553         case SVM_EXIT_MWAIT:
2554                 if (!kvm_ghcb_rax_is_valid(svm) ||
2555                     !kvm_ghcb_rcx_is_valid(svm))
2556                         goto vmgexit_err;
2557                 break;
2558         case SVM_VMGEXIT_MMIO_READ:
2559         case SVM_VMGEXIT_MMIO_WRITE:
2560                 if (!kvm_ghcb_sw_scratch_is_valid(svm))
2561                         goto vmgexit_err;
2562                 break;
2563         case SVM_VMGEXIT_NMI_COMPLETE:
2564         case SVM_VMGEXIT_AP_HLT_LOOP:
2565         case SVM_VMGEXIT_AP_JUMP_TABLE:
2566         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2567                 break;
2568         default:
2569                 reason = GHCB_ERR_INVALID_EVENT;
2570                 goto vmgexit_err;
2571         }
2572
2573         return 0;
2574
2575 vmgexit_err:
2576         if (reason == GHCB_ERR_INVALID_USAGE) {
2577                 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
2578                             svm->sev_es.ghcb->ghcb_usage);
2579         } else if (reason == GHCB_ERR_INVALID_EVENT) {
2580                 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx is not valid\n",
2581                             exit_code);
2582         } else {
2583                 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx input is not valid\n",
2584                             exit_code);
2585                 dump_ghcb(svm);
2586         }
2587
2588         ghcb_set_sw_exit_info_1(svm->sev_es.ghcb, 2);
2589         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, reason);
2590
2591         /* Resume the guest to "return" the error code. */
2592         return 1;
2593 }
2594
2595 void sev_es_unmap_ghcb(struct vcpu_svm *svm)
2596 {
2597         if (!svm->sev_es.ghcb)
2598                 return;
2599
2600         if (svm->sev_es.ghcb_sa_free) {
2601                 /*
2602                  * The scratch area lives outside the GHCB, so there is a
2603                  * buffer that, depending on the operation performed, may
2604                  * need to be synced, then freed.
2605                  */
2606                 if (svm->sev_es.ghcb_sa_sync) {
2607                         kvm_write_guest(svm->vcpu.kvm,
2608                                         svm->sev_es.sw_scratch,
2609                                         svm->sev_es.ghcb_sa,
2610                                         svm->sev_es.ghcb_sa_len);
2611                         svm->sev_es.ghcb_sa_sync = false;
2612                 }
2613
2614                 kvfree(svm->sev_es.ghcb_sa);
2615                 svm->sev_es.ghcb_sa = NULL;
2616                 svm->sev_es.ghcb_sa_free = false;
2617         }
2618
2619         trace_kvm_vmgexit_exit(svm->vcpu.vcpu_id, svm->sev_es.ghcb);
2620
2621         sev_es_sync_to_ghcb(svm);
2622
2623         kvm_vcpu_unmap(&svm->vcpu, &svm->sev_es.ghcb_map, true);
2624         svm->sev_es.ghcb = NULL;
2625 }
2626
2627 void pre_sev_run(struct vcpu_svm *svm, int cpu)
2628 {
2629         struct svm_cpu_data *sd = per_cpu_ptr(&svm_data, cpu);
2630         int asid = sev_get_asid(svm->vcpu.kvm);
2631
2632         /* Assign the asid allocated with this SEV guest */
2633         svm->asid = asid;
2634
2635         /*
2636          * Flush guest TLB:
2637          *
2638          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
2639          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
2640          */
2641         if (sd->sev_vmcbs[asid] == svm->vmcb &&
2642             svm->vcpu.arch.last_vmentry_cpu == cpu)
2643                 return;
2644
2645         sd->sev_vmcbs[asid] = svm->vmcb;
2646         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
2647         vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
2648 }
2649
2650 #define GHCB_SCRATCH_AREA_LIMIT         (16ULL * PAGE_SIZE)
2651 static int setup_vmgexit_scratch(struct vcpu_svm *svm, bool sync, u64 len)
2652 {
2653         struct vmcb_control_area *control = &svm->vmcb->control;
2654         u64 ghcb_scratch_beg, ghcb_scratch_end;
2655         u64 scratch_gpa_beg, scratch_gpa_end;
2656         void *scratch_va;
2657
2658         scratch_gpa_beg = svm->sev_es.sw_scratch;
2659         if (!scratch_gpa_beg) {
2660                 pr_err("vmgexit: scratch gpa not provided\n");
2661                 goto e_scratch;
2662         }
2663
2664         scratch_gpa_end = scratch_gpa_beg + len;
2665         if (scratch_gpa_end < scratch_gpa_beg) {
2666                 pr_err("vmgexit: scratch length (%#llx) not valid for scratch address (%#llx)\n",
2667                        len, scratch_gpa_beg);
2668                 goto e_scratch;
2669         }
2670
2671         if ((scratch_gpa_beg & PAGE_MASK) == control->ghcb_gpa) {
2672                 /* Scratch area begins within GHCB */
2673                 ghcb_scratch_beg = control->ghcb_gpa +
2674                                    offsetof(struct ghcb, shared_buffer);
2675                 ghcb_scratch_end = control->ghcb_gpa +
2676                                    offsetof(struct ghcb, reserved_0xff0);
2677
2678                 /*
2679                  * If the scratch area begins within the GHCB, it must be
2680                  * completely contained in the GHCB shared buffer area.
2681                  */
2682                 if (scratch_gpa_beg < ghcb_scratch_beg ||
2683                     scratch_gpa_end > ghcb_scratch_end) {
2684                         pr_err("vmgexit: scratch area is outside of GHCB shared buffer area (%#llx - %#llx)\n",
2685                                scratch_gpa_beg, scratch_gpa_end);
2686                         goto e_scratch;
2687                 }
2688
2689                 scratch_va = (void *)svm->sev_es.ghcb;
2690                 scratch_va += (scratch_gpa_beg - control->ghcb_gpa);
2691         } else {
2692                 /*
2693                  * The guest memory must be read into a kernel buffer, so
2694                  * limit the size
2695                  */
2696                 if (len > GHCB_SCRATCH_AREA_LIMIT) {
2697                         pr_err("vmgexit: scratch area exceeds KVM limits (%#llx requested, %#llx limit)\n",
2698                                len, GHCB_SCRATCH_AREA_LIMIT);
2699                         goto e_scratch;
2700                 }
2701                 scratch_va = kvzalloc(len, GFP_KERNEL_ACCOUNT);
2702                 if (!scratch_va)
2703                         return -ENOMEM;
2704
2705                 if (kvm_read_guest(svm->vcpu.kvm, scratch_gpa_beg, scratch_va, len)) {
2706                         /* Unable to copy scratch area from guest */
2707                         pr_err("vmgexit: kvm_read_guest for scratch area failed\n");
2708
2709                         kvfree(scratch_va);
2710                         return -EFAULT;
2711                 }
2712
2713                 /*
2714                  * The scratch area is outside the GHCB. The operation will
2715                  * dictate whether the buffer needs to be synced before running
2716                  * the vCPU next time (i.e. a read was requested so the data
2717                  * must be written back to the guest memory).
2718                  */
2719                 svm->sev_es.ghcb_sa_sync = sync;
2720                 svm->sev_es.ghcb_sa_free = true;
2721         }
2722
2723         svm->sev_es.ghcb_sa = scratch_va;
2724         svm->sev_es.ghcb_sa_len = len;
2725
2726         return 0;
2727
2728 e_scratch:
2729         ghcb_set_sw_exit_info_1(svm->sev_es.ghcb, 2);
2730         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, GHCB_ERR_INVALID_SCRATCH_AREA);
2731
2732         return 1;
2733 }
2734
2735 static void set_ghcb_msr_bits(struct vcpu_svm *svm, u64 value, u64 mask,
2736                               unsigned int pos)
2737 {
2738         svm->vmcb->control.ghcb_gpa &= ~(mask << pos);
2739         svm->vmcb->control.ghcb_gpa |= (value & mask) << pos;
2740 }
2741
2742 static u64 get_ghcb_msr_bits(struct vcpu_svm *svm, u64 mask, unsigned int pos)
2743 {
2744         return (svm->vmcb->control.ghcb_gpa >> pos) & mask;
2745 }
2746
2747 static void set_ghcb_msr(struct vcpu_svm *svm, u64 value)
2748 {
2749         svm->vmcb->control.ghcb_gpa = value;
2750 }
2751
2752 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
2753 {
2754         struct vmcb_control_area *control = &svm->vmcb->control;
2755         struct kvm_vcpu *vcpu = &svm->vcpu;
2756         u64 ghcb_info;
2757         int ret = 1;
2758
2759         ghcb_info = control->ghcb_gpa & GHCB_MSR_INFO_MASK;
2760
2761         trace_kvm_vmgexit_msr_protocol_enter(svm->vcpu.vcpu_id,
2762                                              control->ghcb_gpa);
2763
2764         switch (ghcb_info) {
2765         case GHCB_MSR_SEV_INFO_REQ:
2766                 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2767                                                     GHCB_VERSION_MIN,
2768                                                     sev_enc_bit));
2769                 break;
2770         case GHCB_MSR_CPUID_REQ: {
2771                 u64 cpuid_fn, cpuid_reg, cpuid_value;
2772
2773                 cpuid_fn = get_ghcb_msr_bits(svm,
2774                                              GHCB_MSR_CPUID_FUNC_MASK,
2775                                              GHCB_MSR_CPUID_FUNC_POS);
2776
2777                 /* Initialize the registers needed by the CPUID intercept */
2778                 vcpu->arch.regs[VCPU_REGS_RAX] = cpuid_fn;
2779                 vcpu->arch.regs[VCPU_REGS_RCX] = 0;
2780
2781                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_CPUID);
2782                 if (!ret) {
2783                         /* Error, keep GHCB MSR value as-is */
2784                         break;
2785                 }
2786
2787                 cpuid_reg = get_ghcb_msr_bits(svm,
2788                                               GHCB_MSR_CPUID_REG_MASK,
2789                                               GHCB_MSR_CPUID_REG_POS);
2790                 if (cpuid_reg == 0)
2791                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RAX];
2792                 else if (cpuid_reg == 1)
2793                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RBX];
2794                 else if (cpuid_reg == 2)
2795                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RCX];
2796                 else
2797                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RDX];
2798
2799                 set_ghcb_msr_bits(svm, cpuid_value,
2800                                   GHCB_MSR_CPUID_VALUE_MASK,
2801                                   GHCB_MSR_CPUID_VALUE_POS);
2802
2803                 set_ghcb_msr_bits(svm, GHCB_MSR_CPUID_RESP,
2804                                   GHCB_MSR_INFO_MASK,
2805                                   GHCB_MSR_INFO_POS);
2806                 break;
2807         }
2808         case GHCB_MSR_TERM_REQ: {
2809                 u64 reason_set, reason_code;
2810
2811                 reason_set = get_ghcb_msr_bits(svm,
2812                                                GHCB_MSR_TERM_REASON_SET_MASK,
2813                                                GHCB_MSR_TERM_REASON_SET_POS);
2814                 reason_code = get_ghcb_msr_bits(svm,
2815                                                 GHCB_MSR_TERM_REASON_MASK,
2816                                                 GHCB_MSR_TERM_REASON_POS);
2817                 pr_info("SEV-ES guest requested termination: %#llx:%#llx\n",
2818                         reason_set, reason_code);
2819
2820                 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
2821                 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_SEV_TERM;
2822                 vcpu->run->system_event.ndata = 1;
2823                 vcpu->run->system_event.data[0] = control->ghcb_gpa;
2824
2825                 return 0;
2826         }
2827         default:
2828                 /* Error, keep GHCB MSR value as-is */
2829                 break;
2830         }
2831
2832         trace_kvm_vmgexit_msr_protocol_exit(svm->vcpu.vcpu_id,
2833                                             control->ghcb_gpa, ret);
2834
2835         return ret;
2836 }
2837
2838 int sev_handle_vmgexit(struct kvm_vcpu *vcpu)
2839 {
2840         struct vcpu_svm *svm = to_svm(vcpu);
2841         struct vmcb_control_area *control = &svm->vmcb->control;
2842         u64 ghcb_gpa, exit_code;
2843         int ret;
2844
2845         /* Validate the GHCB */
2846         ghcb_gpa = control->ghcb_gpa;
2847         if (ghcb_gpa & GHCB_MSR_INFO_MASK)
2848                 return sev_handle_vmgexit_msr_protocol(svm);
2849
2850         if (!ghcb_gpa) {
2851                 vcpu_unimpl(vcpu, "vmgexit: GHCB gpa is not set\n");
2852
2853                 /* Without a GHCB, just return right back to the guest */
2854                 return 1;
2855         }
2856
2857         if (kvm_vcpu_map(vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->sev_es.ghcb_map)) {
2858                 /* Unable to map GHCB from guest */
2859                 vcpu_unimpl(vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
2860                             ghcb_gpa);
2861
2862                 /* Without a GHCB, just return right back to the guest */
2863                 return 1;
2864         }
2865
2866         svm->sev_es.ghcb = svm->sev_es.ghcb_map.hva;
2867
2868         trace_kvm_vmgexit_enter(vcpu->vcpu_id, svm->sev_es.ghcb);
2869
2870         sev_es_sync_from_ghcb(svm);
2871         ret = sev_es_validate_vmgexit(svm);
2872         if (ret)
2873                 return ret;
2874
2875         ghcb_set_sw_exit_info_1(svm->sev_es.ghcb, 0);
2876         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, 0);
2877
2878         exit_code = kvm_ghcb_get_sw_exit_code(control);
2879         switch (exit_code) {
2880         case SVM_VMGEXIT_MMIO_READ:
2881                 ret = setup_vmgexit_scratch(svm, true, control->exit_info_2);
2882                 if (ret)
2883                         break;
2884
2885                 ret = kvm_sev_es_mmio_read(vcpu,
2886                                            control->exit_info_1,
2887                                            control->exit_info_2,
2888                                            svm->sev_es.ghcb_sa);
2889                 break;
2890         case SVM_VMGEXIT_MMIO_WRITE:
2891                 ret = setup_vmgexit_scratch(svm, false, control->exit_info_2);
2892                 if (ret)
2893                         break;
2894
2895                 ret = kvm_sev_es_mmio_write(vcpu,
2896                                             control->exit_info_1,
2897                                             control->exit_info_2,
2898                                             svm->sev_es.ghcb_sa);
2899                 break;
2900         case SVM_VMGEXIT_NMI_COMPLETE:
2901                 ++vcpu->stat.nmi_window_exits;
2902                 svm->nmi_masked = false;
2903                 kvm_make_request(KVM_REQ_EVENT, vcpu);
2904                 ret = 1;
2905                 break;
2906         case SVM_VMGEXIT_AP_HLT_LOOP:
2907                 ret = kvm_emulate_ap_reset_hold(vcpu);
2908                 break;
2909         case SVM_VMGEXIT_AP_JUMP_TABLE: {
2910                 struct kvm_sev_info *sev = &to_kvm_svm(vcpu->kvm)->sev_info;
2911
2912                 switch (control->exit_info_1) {
2913                 case 0:
2914                         /* Set AP jump table address */
2915                         sev->ap_jump_table = control->exit_info_2;
2916                         break;
2917                 case 1:
2918                         /* Get AP jump table address */
2919                         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, sev->ap_jump_table);
2920                         break;
2921                 default:
2922                         pr_err("svm: vmgexit: unsupported AP jump table request - exit_info_1=%#llx\n",
2923                                control->exit_info_1);
2924                         ghcb_set_sw_exit_info_1(svm->sev_es.ghcb, 2);
2925                         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, GHCB_ERR_INVALID_INPUT);
2926                 }
2927
2928                 ret = 1;
2929                 break;
2930         }
2931         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2932                 vcpu_unimpl(vcpu,
2933                             "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
2934                             control->exit_info_1, control->exit_info_2);
2935                 ret = -EINVAL;
2936                 break;
2937         default:
2938                 ret = svm_invoke_exit_handler(vcpu, exit_code);
2939         }
2940
2941         return ret;
2942 }
2943
2944 int sev_es_string_io(struct vcpu_svm *svm, int size, unsigned int port, int in)
2945 {
2946         int count;
2947         int bytes;
2948         int r;
2949
2950         if (svm->vmcb->control.exit_info_2 > INT_MAX)
2951                 return -EINVAL;
2952
2953         count = svm->vmcb->control.exit_info_2;
2954         if (unlikely(check_mul_overflow(count, size, &bytes)))
2955                 return -EINVAL;
2956
2957         r = setup_vmgexit_scratch(svm, in, bytes);
2958         if (r)
2959                 return r;
2960
2961         return kvm_sev_es_string_io(&svm->vcpu, size, port, svm->sev_es.ghcb_sa,
2962                                     count, in);
2963 }
2964
2965 static void sev_es_vcpu_after_set_cpuid(struct vcpu_svm *svm)
2966 {
2967         struct kvm_vcpu *vcpu = &svm->vcpu;
2968
2969         if (boot_cpu_has(X86_FEATURE_V_TSC_AUX)) {
2970                 bool v_tsc_aux = guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP) ||
2971                                  guest_cpuid_has(vcpu, X86_FEATURE_RDPID);
2972
2973                 set_msr_interception(vcpu, svm->msrpm, MSR_TSC_AUX, v_tsc_aux, v_tsc_aux);
2974         }
2975
2976         /*
2977          * For SEV-ES, accesses to MSR_IA32_XSS should not be intercepted if
2978          * the host/guest supports its use.
2979          *
2980          * guest_can_use() checks a number of requirements on the host/guest to
2981          * ensure that MSR_IA32_XSS is available, but it might report true even
2982          * if X86_FEATURE_XSAVES isn't configured in the guest to ensure host
2983          * MSR_IA32_XSS is always properly restored. For SEV-ES, it is better
2984          * to further check that the guest CPUID actually supports
2985          * X86_FEATURE_XSAVES so that accesses to MSR_IA32_XSS by misbehaved
2986          * guests will still get intercepted and caught in the normal
2987          * kvm_emulate_rdmsr()/kvm_emulated_wrmsr() paths.
2988          */
2989         if (guest_can_use(vcpu, X86_FEATURE_XSAVES) &&
2990             guest_cpuid_has(vcpu, X86_FEATURE_XSAVES))
2991                 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_XSS, 1, 1);
2992         else
2993                 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_XSS, 0, 0);
2994 }
2995
2996 void sev_vcpu_after_set_cpuid(struct vcpu_svm *svm)
2997 {
2998         struct kvm_vcpu *vcpu = &svm->vcpu;
2999         struct kvm_cpuid_entry2 *best;
3000
3001         /* For sev guests, the memory encryption bit is not reserved in CR3.  */
3002         best = kvm_find_cpuid_entry(vcpu, 0x8000001F);
3003         if (best)
3004                 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
3005
3006         if (sev_es_guest(svm->vcpu.kvm))
3007                 sev_es_vcpu_after_set_cpuid(svm);
3008 }
3009
3010 static void sev_es_init_vmcb(struct vcpu_svm *svm)
3011 {
3012         struct vmcb *vmcb = svm->vmcb01.ptr;
3013         struct kvm_vcpu *vcpu = &svm->vcpu;
3014
3015         svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ES_ENABLE;
3016         svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
3017
3018         /*
3019          * An SEV-ES guest requires a VMSA area that is a separate from the
3020          * VMCB page. Do not include the encryption mask on the VMSA physical
3021          * address since hardware will access it using the guest key.  Note,
3022          * the VMSA will be NULL if this vCPU is the destination for intrahost
3023          * migration, and will be copied later.
3024          */
3025         if (svm->sev_es.vmsa)
3026                 svm->vmcb->control.vmsa_pa = __pa(svm->sev_es.vmsa);
3027
3028         /* Can't intercept CR register access, HV can't modify CR registers */
3029         svm_clr_intercept(svm, INTERCEPT_CR0_READ);
3030         svm_clr_intercept(svm, INTERCEPT_CR4_READ);
3031         svm_clr_intercept(svm, INTERCEPT_CR8_READ);
3032         svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
3033         svm_clr_intercept(svm, INTERCEPT_CR4_WRITE);
3034         svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
3035
3036         svm_clr_intercept(svm, INTERCEPT_SELECTIVE_CR0);
3037
3038         /* Track EFER/CR register changes */
3039         svm_set_intercept(svm, TRAP_EFER_WRITE);
3040         svm_set_intercept(svm, TRAP_CR0_WRITE);
3041         svm_set_intercept(svm, TRAP_CR4_WRITE);
3042         svm_set_intercept(svm, TRAP_CR8_WRITE);
3043
3044         vmcb->control.intercepts[INTERCEPT_DR] = 0;
3045         if (!sev_es_debug_swap_enabled) {
3046                 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_READ);
3047                 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_WRITE);
3048                 recalc_intercepts(svm);
3049         } else {
3050                 /*
3051                  * Disable #DB intercept iff DebugSwap is enabled.  KVM doesn't
3052                  * allow debugging SEV-ES guests, and enables DebugSwap iff
3053                  * NO_NESTED_DATA_BP is supported, so there's no reason to
3054                  * intercept #DB when DebugSwap is enabled.  For simplicity
3055                  * with respect to guest debug, intercept #DB for other VMs
3056                  * even if NO_NESTED_DATA_BP is supported, i.e. even if the
3057                  * guest can't DoS the CPU with infinite #DB vectoring.
3058                  */
3059                 clr_exception_intercept(svm, DB_VECTOR);
3060         }
3061
3062         /* Can't intercept XSETBV, HV can't modify XCR0 directly */
3063         svm_clr_intercept(svm, INTERCEPT_XSETBV);
3064
3065         /* Clear intercepts on selected MSRs */
3066         set_msr_interception(vcpu, svm->msrpm, MSR_EFER, 1, 1);
3067         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_CR_PAT, 1, 1);
3068         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
3069         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
3070         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
3071         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
3072 }
3073
3074 void sev_init_vmcb(struct vcpu_svm *svm)
3075 {
3076         svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
3077         clr_exception_intercept(svm, UD_VECTOR);
3078
3079         /*
3080          * Don't intercept #GP for SEV guests, e.g. for the VMware backdoor, as
3081          * KVM can't decrypt guest memory to decode the faulting instruction.
3082          */
3083         clr_exception_intercept(svm, GP_VECTOR);
3084
3085         if (sev_es_guest(svm->vcpu.kvm))
3086                 sev_es_init_vmcb(svm);
3087 }
3088
3089 void sev_es_vcpu_reset(struct vcpu_svm *svm)
3090 {
3091         /*
3092          * Set the GHCB MSR value as per the GHCB specification when emulating
3093          * vCPU RESET for an SEV-ES guest.
3094          */
3095         set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
3096                                             GHCB_VERSION_MIN,
3097                                             sev_enc_bit));
3098 }
3099
3100 void sev_es_prepare_switch_to_guest(struct sev_es_save_area *hostsa)
3101 {
3102         /*
3103          * All host state for SEV-ES guests is categorized into three swap types
3104          * based on how it is handled by hardware during a world switch:
3105          *
3106          * A: VMRUN:   Host state saved in host save area
3107          *    VMEXIT:  Host state loaded from host save area
3108          *
3109          * B: VMRUN:   Host state _NOT_ saved in host save area
3110          *    VMEXIT:  Host state loaded from host save area
3111          *
3112          * C: VMRUN:   Host state _NOT_ saved in host save area
3113          *    VMEXIT:  Host state initialized to default(reset) values
3114          *
3115          * Manually save type-B state, i.e. state that is loaded by VMEXIT but
3116          * isn't saved by VMRUN, that isn't already saved by VMSAVE (performed
3117          * by common SVM code).
3118          */
3119         hostsa->xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
3120         hostsa->pkru = read_pkru();
3121         hostsa->xss = host_xss;
3122
3123         /*
3124          * If DebugSwap is enabled, debug registers are loaded but NOT saved by
3125          * the CPU (Type-B). If DebugSwap is disabled/unsupported, the CPU both
3126          * saves and loads debug registers (Type-A).
3127          */
3128         if (sev_es_debug_swap_enabled) {
3129                 hostsa->dr0 = native_get_debugreg(0);
3130                 hostsa->dr1 = native_get_debugreg(1);
3131                 hostsa->dr2 = native_get_debugreg(2);
3132                 hostsa->dr3 = native_get_debugreg(3);
3133                 hostsa->dr0_addr_mask = amd_get_dr_addr_mask(0);
3134                 hostsa->dr1_addr_mask = amd_get_dr_addr_mask(1);
3135                 hostsa->dr2_addr_mask = amd_get_dr_addr_mask(2);
3136                 hostsa->dr3_addr_mask = amd_get_dr_addr_mask(3);
3137         }
3138 }
3139
3140 void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
3141 {
3142         struct vcpu_svm *svm = to_svm(vcpu);
3143
3144         /* First SIPI: Use the values as initially set by the VMM */
3145         if (!svm->sev_es.received_first_sipi) {
3146                 svm->sev_es.received_first_sipi = true;
3147                 return;
3148         }
3149
3150         /*
3151          * Subsequent SIPI: Return from an AP Reset Hold VMGEXIT, where
3152          * the guest will set the CS and RIP. Set SW_EXIT_INFO_2 to a
3153          * non-zero value.
3154          */
3155         if (!svm->sev_es.ghcb)
3156                 return;
3157
3158         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, 1);
3159 }