arch/x86/kvm/svm/sev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel-based Virtual Machine driver for Linux
   4  *
   5  * AMD SVM-SEV support
   6  *
   7  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   8  */
   9
  10 #include <linux/kvm_types.h>
  11 #include <linux/kvm_host.h>
  12 #include <linux/kernel.h>
  13 #include <linux/highmem.h>
  14 #include <linux/psp-sev.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/swap.h>
  17 #include <linux/misc_cgroup.h>
  18 #include <linux/processor.h>
  19 #include <linux/trace_events.h>
  20
  21 #include <asm/pkru.h>
  22 #include <asm/trapnr.h>
  23 #include <asm/fpu/xcr.h>
  24
  25 #include "x86.h"
  26 #include "svm.h"
  27 #include "svm_ops.h"
  28 #include "cpuid.h"
  29 #include "trace.h"
  30
  31 #ifndef CONFIG_KVM_AMD_SEV
  32 /*
  33  * When this config is not defined, SEV feature is not supported and APIs in
  34  * this file are not used but this file still gets compiled into the KVM AMD
  35  * module.
  36  *
  37  * We will not have MISC_CG_RES_SEV and MISC_CG_RES_SEV_ES entries in the enum
  38  * misc_res_type {} defined in linux/misc_cgroup.h.
  39  *
  40  * Below macros allow compilation to succeed.
  41  */
  42 #define MISC_CG_RES_SEV MISC_CG_RES_TYPES
  43 #define MISC_CG_RES_SEV_ES MISC_CG_RES_TYPES
  44 #endif
  45
  46 #ifdef CONFIG_KVM_AMD_SEV
  47 /* enable/disable SEV support */
  48 static bool sev_enabled = true;
  49 module_param_named(sev, sev_enabled, bool, 0444);
  50
  51 /* enable/disable SEV-ES support */
  52 static bool sev_es_enabled = true;
  53 module_param_named(sev_es, sev_es_enabled, bool, 0444);
  54 #else
  55 #define sev_enabled false
  56 #define sev_es_enabled false
  57 #endif /* CONFIG_KVM_AMD_SEV */
  58
  59 static u8 sev_enc_bit;
  60 static DECLARE_RWSEM(sev_deactivate_lock);
  61 static DEFINE_MUTEX(sev_bitmap_lock);
  62 unsigned int max_sev_asid;
  63 static unsigned int min_sev_asid;
  64 static unsigned long sev_me_mask;
  65 static unsigned int nr_asids;
  66 static unsigned long *sev_asid_bitmap;
  67 static unsigned long *sev_reclaim_asid_bitmap;
  68
  69 struct enc_region {
  70         struct list_head list;
  71         unsigned long npages;
  72         struct page **pages;
  73         unsigned long uaddr;
  74         unsigned long size;
  75 };
  76
  77 /* Called with the sev_bitmap_lock held, or on shutdown  */
  78 static int sev_flush_asids(int min_asid, int max_asid)
  79 {
  80         int ret, asid, error = 0;
  81
  82         /* Check if there are any ASIDs to reclaim before performing a flush */
  83         asid = find_next_bit(sev_reclaim_asid_bitmap, nr_asids, min_asid);
  84         if (asid > max_asid)
  85                 return -EBUSY;
  86
  87         /*
  88          * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
  89          * so it must be guarded.
  90          */
  91         down_write(&sev_deactivate_lock);
  92
  93         wbinvd_on_all_cpus();
  94         ret = sev_guest_df_flush(&error);
  95
  96         up_write(&sev_deactivate_lock);
  97
  98         if (ret)
  99                 pr_err("SEV: DF_FLUSH failed, ret=%d, error=%#x\n", ret, error);
 100
 101         return ret;
 102 }
 103
 104 static inline bool is_mirroring_enc_context(struct kvm *kvm)
 105 {
 106         return !!to_kvm_svm(kvm)->sev_info.enc_context_owner;
 107 }
 108
 109 /* Must be called with the sev_bitmap_lock held */
 110 static bool __sev_recycle_asids(int min_asid, int max_asid)
 111 {
 112         if (sev_flush_asids(min_asid, max_asid))
 113                 return false;
 114
 115         /* The flush process will flush all reclaimable SEV and SEV-ES ASIDs */
 116         bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
 117                    nr_asids);
 118         bitmap_zero(sev_reclaim_asid_bitmap, nr_asids);
 119
 120         return true;
 121 }
 122
 123 static int sev_misc_cg_try_charge(struct kvm_sev_info *sev)
 124 {
 125         enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 126         return misc_cg_try_charge(type, sev->misc_cg, 1);
 127 }
 128
 129 static void sev_misc_cg_uncharge(struct kvm_sev_info *sev)
 130 {
 131         enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 132         misc_cg_uncharge(type, sev->misc_cg, 1);
 133 }
 134
 135 static int sev_asid_new(struct kvm_sev_info *sev)
 136 {
 137         int asid, min_asid, max_asid, ret;
 138         bool retry = true;
 139
 140         WARN_ON(sev->misc_cg);
 141         sev->misc_cg = get_current_misc_cg();
 142         ret = sev_misc_cg_try_charge(sev);
 143         if (ret) {
 144                 put_misc_cg(sev->misc_cg);
 145                 sev->misc_cg = NULL;
 146                 return ret;
 147         }
 148
 149         mutex_lock(&sev_bitmap_lock);
 150
 151         /*
 152          * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
 153          * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.
 154          */
 155         min_asid = sev->es_active ? 1 : min_sev_asid;
 156         max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
 157 again:
 158         asid = find_next_zero_bit(sev_asid_bitmap, max_asid + 1, min_asid);
 159         if (asid > max_asid) {
 160                 if (retry && __sev_recycle_asids(min_asid, max_asid)) {
 161                         retry = false;
 162                         goto again;
 163                 }
 164                 mutex_unlock(&sev_bitmap_lock);
 165                 ret = -EBUSY;
 166                 goto e_uncharge;
 167         }
 168
 169         __set_bit(asid, sev_asid_bitmap);
 170
 171         mutex_unlock(&sev_bitmap_lock);
 172
 173         return asid;
 174 e_uncharge:
 175         sev_misc_cg_uncharge(sev);
 176         put_misc_cg(sev->misc_cg);
 177         sev->misc_cg = NULL;
 178         return ret;
 179 }
 180
 181 static int sev_get_asid(struct kvm *kvm)
 182 {
 183         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 184
 185         return sev->asid;
 186 }
 187
 188 static void sev_asid_free(struct kvm_sev_info *sev)
 189 {
 190         struct svm_cpu_data *sd;
 191         int cpu;
 192
 193         mutex_lock(&sev_bitmap_lock);
 194
 195         __set_bit(sev->asid, sev_reclaim_asid_bitmap);
 196
 197         for_each_possible_cpu(cpu) {
 198                 sd = per_cpu(svm_data, cpu);
 199                 sd->sev_vmcbs[sev->asid] = NULL;
 200         }
 201
 202         mutex_unlock(&sev_bitmap_lock);
 203
 204         sev_misc_cg_uncharge(sev);
 205         put_misc_cg(sev->misc_cg);
 206         sev->misc_cg = NULL;
 207 }
 208
 209 static void sev_decommission(unsigned int handle)
 210 {
 211         struct sev_data_decommission decommission;
 212
 213         if (!handle)
 214                 return;
 215
 216         decommission.handle = handle;
 217         sev_guest_decommission(&decommission, NULL);
 218 }
 219
 220 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
 221 {
 222         struct sev_data_deactivate deactivate;
 223
 224         if (!handle)
 225                 return;
 226
 227         deactivate.handle = handle;
 228
 229         /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
 230         down_read(&sev_deactivate_lock);
 231         sev_guest_deactivate(&deactivate, NULL);
 232         up_read(&sev_deactivate_lock);
 233
 234         sev_decommission(handle);
 235 }
 236
 237 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 238 {
 239         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 240         int asid, ret;
 241
 242         if (kvm->created_vcpus)
 243                 return -EINVAL;
 244
 245         ret = -EBUSY;
 246         if (unlikely(sev->active))
 247                 return ret;
 248
 249         sev->active = true;
 250         sev->es_active = argp->id == KVM_SEV_ES_INIT;
 251         asid = sev_asid_new(sev);
 252         if (asid < 0)
 253                 goto e_no_asid;
 254         sev->asid = asid;
 255
 256         ret = sev_platform_init(&argp->error);
 257         if (ret)
 258                 goto e_free;
 259
 260         INIT_LIST_HEAD(&sev->regions_list);
 261         INIT_LIST_HEAD(&sev->mirror_vms);
 262
 263         return 0;
 264
 265 e_free:
 266         sev_asid_free(sev);
 267         sev->asid = 0;
 268 e_no_asid:
 269         sev->es_active = false;
 270         sev->active = false;
 271         return ret;
 272 }
 273
 274 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
 275 {
 276         struct sev_data_activate activate;
 277         int asid = sev_get_asid(kvm);
 278         int ret;
 279
 280         /* activate ASID on the given handle */
 281         activate.handle = handle;
 282         activate.asid   = asid;
 283         ret = sev_guest_activate(&activate, error);
 284
 285         return ret;
 286 }
 287
 288 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
 289 {
 290         struct fd f;
 291         int ret;
 292
 293         f = fdget(fd);
 294         if (!f.file)
 295                 return -EBADF;
 296
 297         ret = sev_issue_cmd_external_user(f.file, id, data, error);
 298
 299         fdput(f);
 300         return ret;
 301 }
 302
 303 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
 304 {
 305         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 306
 307         return __sev_issue_cmd(sev->fd, id, data, error);
 308 }
 309
 310 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
 311 {
 312         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 313         struct sev_data_launch_start start;
 314         struct kvm_sev_launch_start params;
 315         void *dh_blob, *session_blob;
 316         int *error = &argp->error;
 317         int ret;
 318
 319         if (!sev_guest(kvm))
 320                 return -ENOTTY;
 321
 322         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 323                 return -EFAULT;
 324
 325         memset(&start, 0, sizeof(start));
 326
 327         dh_blob = NULL;
 328         if (params.dh_uaddr) {
 329                 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
 330                 if (IS_ERR(dh_blob))
 331                         return PTR_ERR(dh_blob);
 332
 333                 start.dh_cert_address = __sme_set(__pa(dh_blob));
 334                 start.dh_cert_len = params.dh_len;
 335         }
 336
 337         session_blob = NULL;
 338         if (params.session_uaddr) {
 339                 session_blob = psp_copy_user_blob(params.session_uaddr, params.session_len);
 340                 if (IS_ERR(session_blob)) {
 341                         ret = PTR_ERR(session_blob);
 342                         goto e_free_dh;
 343                 }
 344
 345                 start.session_address = __sme_set(__pa(session_blob));
 346                 start.session_len = params.session_len;
 347         }
 348
 349         start.handle = params.handle;
 350         start.policy = params.policy;
 351
 352         /* create memory encryption context */
 353         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_LAUNCH_START, &start, error);
 354         if (ret)
 355                 goto e_free_session;
 356
 357         /* Bind ASID to this guest */
 358         ret = sev_bind_asid(kvm, start.handle, error);
 359         if (ret) {
 360                 sev_decommission(start.handle);
 361                 goto e_free_session;
 362         }
 363
 364         /* return handle to userspace */
 365         params.handle = start.handle;
 366         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params))) {
 367                 sev_unbind_asid(kvm, start.handle);
 368                 ret = -EFAULT;
 369                 goto e_free_session;
 370         }
 371
 372         sev->handle = start.handle;
 373         sev->fd = argp->sev_fd;
 374
 375 e_free_session:
 376         kfree(session_blob);
 377 e_free_dh:
 378         kfree(dh_blob);
 379         return ret;
 380 }
 381
 382 static struct page **sev_pin_memory(struct kvm *kvm, unsigned long uaddr,
 383                                     unsigned long ulen, unsigned long *n,
 384                                     int write)
 385 {
 386         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 387         unsigned long npages, size;
 388         int npinned;
 389         unsigned long locked, lock_limit;
 390         struct page **pages;
 391         unsigned long first, last;
 392         int ret;
 393
 394         lockdep_assert_held(&kvm->lock);
 395
 396         if (ulen == 0 || uaddr + ulen < uaddr)
 397                 return ERR_PTR(-EINVAL);
 398
 399         /* Calculate number of pages. */
 400         first = (uaddr & PAGE_MASK) >> PAGE_SHIFT;
 401         last = ((uaddr + ulen - 1) & PAGE_MASK) >> PAGE_SHIFT;
 402         npages = (last - first + 1);
 403
 404         locked = sev->pages_locked + npages;
 405         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 406         if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
 407                 pr_err("SEV: %lu locked pages exceed the lock limit of %lu.\n", locked, lock_limit);
 408                 return ERR_PTR(-ENOMEM);
 409         }
 410
 411         if (WARN_ON_ONCE(npages > INT_MAX))
 412                 return ERR_PTR(-EINVAL);
 413
 414         /* Avoid using vmalloc for smaller buffers. */
 415         size = npages * sizeof(struct page *);
 416         if (size > PAGE_SIZE)
 417                 pages = __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 418         else
 419                 pages = kmalloc(size, GFP_KERNEL_ACCOUNT);
 420
 421         if (!pages)
 422                 return ERR_PTR(-ENOMEM);
 423
 424         /* Pin the user virtual address. */
 425         npinned = pin_user_pages_fast(uaddr, npages, write ? FOLL_WRITE : 0, pages);
 426         if (npinned != npages) {
 427                 pr_err("SEV: Failure locking %lu pages.\n", npages);
 428                 ret = -ENOMEM;
 429                 goto err;
 430         }
 431
 432         *n = npages;
 433         sev->pages_locked = locked;
 434
 435         return pages;
 436
 437 err:
 438         if (npinned > 0)
 439                 unpin_user_pages(pages, npinned);
 440
 441         kvfree(pages);
 442         return ERR_PTR(ret);
 443 }
 444
 445 static void sev_unpin_memory(struct kvm *kvm, struct page **pages,
 446                              unsigned long npages)
 447 {
 448         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 449
 450         unpin_user_pages(pages, npages);
 451         kvfree(pages);
 452         sev->pages_locked -= npages;
 453 }
 454
 455 static void sev_clflush_pages(struct page *pages[], unsigned long npages)
 456 {
 457         uint8_t *page_virtual;
 458         unsigned long i;
 459
 460         if (this_cpu_has(X86_FEATURE_SME_COHERENT) || npages == 0 ||
 461             pages == NULL)
 462                 return;
 463
 464         for (i = 0; i < npages; i++) {
 465                 page_virtual = kmap_atomic(pages[i]);
 466                 clflush_cache_range(page_virtual, PAGE_SIZE);
 467                 kunmap_atomic(page_virtual);
 468         }
 469 }
 470
 471 static unsigned long get_num_contig_pages(unsigned long idx,
 472                                 struct page **inpages, unsigned long npages)
 473 {
 474         unsigned long paddr, next_paddr;
 475         unsigned long i = idx + 1, pages = 1;
 476
 477         /* find the number of contiguous pages starting from idx */
 478         paddr = __sme_page_pa(inpages[idx]);
 479         while (i < npages) {
 480                 next_paddr = __sme_page_pa(inpages[i++]);
 481                 if ((paddr + PAGE_SIZE) == next_paddr) {
 482                         pages++;
 483                         paddr = next_paddr;
 484                         continue;
 485                 }
 486                 break;
 487         }
 488
 489         return pages;
 490 }
 491
 492 static int sev_launch_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
 493 {
 494         unsigned long vaddr, vaddr_end, next_vaddr, npages, pages, size, i;
 495         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 496         struct kvm_sev_launch_update_data params;
 497         struct sev_data_launch_update_data data;
 498         struct page **inpages;
 499         int ret;
 500
 501         if (!sev_guest(kvm))
 502                 return -ENOTTY;
 503
 504         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 505                 return -EFAULT;
 506
 507         vaddr = params.uaddr;
 508         size = params.len;
 509         vaddr_end = vaddr + size;
 510
 511         /* Lock the user memory. */
 512         inpages = sev_pin_memory(kvm, vaddr, size, &npages, 1);
 513         if (IS_ERR(inpages))
 514                 return PTR_ERR(inpages);
 515
 516         /*
 517          * Flush (on non-coherent CPUs) before LAUNCH_UPDATE encrypts pages in
 518          * place; the cache may contain the data that was written unencrypted.
 519          */
 520         sev_clflush_pages(inpages, npages);
 521
 522         data.reserved = 0;
 523         data.handle = sev->handle;
 524
 525         for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
 526                 int offset, len;
 527
 528                 /*
 529                  * If the user buffer is not page-aligned, calculate the offset
 530                  * within the page.
 531                  */
 532                 offset = vaddr & (PAGE_SIZE - 1);
 533
 534                 /* Calculate the number of pages that can be encrypted in one go. */
 535                 pages = get_num_contig_pages(i, inpages, npages);
 536
 537                 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
 538
 539                 data.len = len;
 540                 data.address = __sme_page_pa(inpages[i]) + offset;
 541                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, &data, &argp->error);
 542                 if (ret)
 543                         goto e_unpin;
 544
 545                 size -= len;
 546                 next_vaddr = vaddr + len;
 547         }
 548
 549 e_unpin:
 550         /* content of memory is updated, mark pages dirty */
 551         for (i = 0; i < npages; i++) {
 552                 set_page_dirty_lock(inpages[i]);
 553                 mark_page_accessed(inpages[i]);
 554         }
 555         /* unlock the user pages */
 556         sev_unpin_memory(kvm, inpages, npages);
 557         return ret;
 558 }
 559
 560 static int sev_es_sync_vmsa(struct vcpu_svm *svm)
 561 {
 562         struct vmcb_save_area *save = &svm->vmcb->save;
 563
 564         /* Check some debug related fields before encrypting the VMSA */
 565         if (svm->vcpu.guest_debug || (save->dr7 & ~DR7_FIXED_1))
 566                 return -EINVAL;
 567
 568         /* Sync registgers */
 569         save->rax = svm->vcpu.arch.regs[VCPU_REGS_RAX];
 570         save->rbx = svm->vcpu.arch.regs[VCPU_REGS_RBX];
 571         save->rcx = svm->vcpu.arch.regs[VCPU_REGS_RCX];
 572         save->rdx = svm->vcpu.arch.regs[VCPU_REGS_RDX];
 573         save->rsp = svm->vcpu.arch.regs[VCPU_REGS_RSP];
 574         save->rbp = svm->vcpu.arch.regs[VCPU_REGS_RBP];
 575         save->rsi = svm->vcpu.arch.regs[VCPU_REGS_RSI];
 576         save->rdi = svm->vcpu.arch.regs[VCPU_REGS_RDI];
 577 #ifdef CONFIG_X86_64
 578         save->r8  = svm->vcpu.arch.regs[VCPU_REGS_R8];
 579         save->r9  = svm->vcpu.arch.regs[VCPU_REGS_R9];
 580         save->r10 = svm->vcpu.arch.regs[VCPU_REGS_R10];
 581         save->r11 = svm->vcpu.arch.regs[VCPU_REGS_R11];
 582         save->r12 = svm->vcpu.arch.regs[VCPU_REGS_R12];
 583         save->r13 = svm->vcpu.arch.regs[VCPU_REGS_R13];
 584         save->r14 = svm->vcpu.arch.regs[VCPU_REGS_R14];
 585         save->r15 = svm->vcpu.arch.regs[VCPU_REGS_R15];
 586 #endif
 587         save->rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
 588
 589         /* Sync some non-GPR registers before encrypting */
 590         save->xcr0 = svm->vcpu.arch.xcr0;
 591         save->pkru = svm->vcpu.arch.pkru;
 592         save->xss  = svm->vcpu.arch.ia32_xss;
 593         save->dr6  = svm->vcpu.arch.dr6;
 594
 595         /*
 596          * SEV-ES will use a VMSA that is pointed to by the VMCB, not
 597          * the traditional VMSA that is part of the VMCB. Copy the
 598          * traditional VMSA as it has been built so far (in prep
 599          * for LAUNCH_UPDATE_VMSA) to be the initial SEV-ES state.
 600          */
 601         memcpy(svm->sev_es.vmsa, save, sizeof(*save));
 602
 603         return 0;
 604 }
 605
 606 static int __sev_launch_update_vmsa(struct kvm *kvm, struct kvm_vcpu *vcpu,
 607                                     int *error)
 608 {
 609         struct sev_data_launch_update_vmsa vmsa;
 610         struct vcpu_svm *svm = to_svm(vcpu);
 611         int ret;
 612
 613         /* Perform some pre-encryption checks against the VMSA */
 614         ret = sev_es_sync_vmsa(svm);
 615         if (ret)
 616                 return ret;
 617
 618         /*
 619          * The LAUNCH_UPDATE_VMSA command will perform in-place encryption of
 620          * the VMSA memory content (i.e it will write the same memory region
 621          * with the guest's key), so invalidate it first.
 622          */
 623         clflush_cache_range(svm->sev_es.vmsa, PAGE_SIZE);
 624
 625         vmsa.reserved = 0;
 626         vmsa.handle = to_kvm_svm(kvm)->sev_info.handle;
 627         vmsa.address = __sme_pa(svm->sev_es.vmsa);
 628         vmsa.len = PAGE_SIZE;
 629         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_VMSA, &vmsa, error);
 630         if (ret)
 631           return ret;
 632
 633         vcpu->arch.guest_state_protected = true;
 634         return 0;
 635 }
 636
 637 static int sev_launch_update_vmsa(struct kvm *kvm, struct kvm_sev_cmd *argp)
 638 {
 639         struct kvm_vcpu *vcpu;
 640         unsigned long i;
 641         int ret;
 642
 643         if (!sev_es_guest(kvm))
 644                 return -ENOTTY;
 645
 646         kvm_for_each_vcpu(i, vcpu, kvm) {
 647                 ret = mutex_lock_killable(&vcpu->mutex);
 648                 if (ret)
 649                         return ret;
 650
 651                 ret = __sev_launch_update_vmsa(kvm, vcpu, &argp->error);
 652
 653                 mutex_unlock(&vcpu->mutex);
 654                 if (ret)
 655                         return ret;
 656         }
 657
 658         return 0;
 659 }
 660
 661 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 662 {
 663         void __user *measure = (void __user *)(uintptr_t)argp->data;
 664         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 665         struct sev_data_launch_measure data;
 666         struct kvm_sev_launch_measure params;
 667         void __user *p = NULL;
 668         void *blob = NULL;
 669         int ret;
 670
 671         if (!sev_guest(kvm))
 672                 return -ENOTTY;
 673
 674         if (copy_from_user(&params, measure, sizeof(params)))
 675                 return -EFAULT;
 676
 677         memset(&data, 0, sizeof(data));
 678
 679         /* User wants to query the blob length */
 680         if (!params.len)
 681                 goto cmd;
 682
 683         p = (void __user *)(uintptr_t)params.uaddr;
 684         if (p) {
 685                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
 686                         return -EINVAL;
 687
 688                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
 689                 if (!blob)
 690                         return -ENOMEM;
 691
 692                 data.address = __psp_pa(blob);
 693                 data.len = params.len;
 694         }
 695
 696 cmd:
 697         data.handle = sev->handle;
 698         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, &data, &argp->error);
 699
 700         /*
 701          * If we query the session length, FW responded with expected data.
 702          */
 703         if (!params.len)
 704                 goto done;
 705
 706         if (ret)
 707                 goto e_free_blob;
 708
 709         if (blob) {
 710                 if (copy_to_user(p, blob, params.len))
 711                         ret = -EFAULT;
 712         }
 713
 714 done:
 715         params.len = data.len;
 716         if (copy_to_user(measure, &params, sizeof(params)))
 717                 ret = -EFAULT;
 718 e_free_blob:
 719         kfree(blob);
 720         return ret;
 721 }
 722
 723 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 724 {
 725         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 726         struct sev_data_launch_finish data;
 727
 728         if (!sev_guest(kvm))
 729                 return -ENOTTY;
 730
 731         data.handle = sev->handle;
 732         return sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, &data, &argp->error);
 733 }
 734
 735 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 736 {
 737         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 738         struct kvm_sev_guest_status params;
 739         struct sev_data_guest_status data;
 740         int ret;
 741
 742         if (!sev_guest(kvm))
 743                 return -ENOTTY;
 744
 745         memset(&data, 0, sizeof(data));
 746
 747         data.handle = sev->handle;
 748         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, &data, &argp->error);
 749         if (ret)
 750                 return ret;
 751
 752         params.policy = data.policy;
 753         params.state = data.state;
 754         params.handle = data.handle;
 755
 756         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 757                 ret = -EFAULT;
 758
 759         return ret;
 760 }
 761
 762 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 763                                unsigned long dst, int size,
 764                                int *error, bool enc)
 765 {
 766         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 767         struct sev_data_dbg data;
 768
 769         data.reserved = 0;
 770         data.handle = sev->handle;
 771         data.dst_addr = dst;
 772         data.src_addr = src;
 773         data.len = size;
 774
 775         return sev_issue_cmd(kvm,
 776                              enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 777                              &data, error);
 778 }
 779
 780 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 781                              unsigned long dst_paddr, int sz, int *err)
 782 {
 783         int offset;
 784
 785         /*
 786          * Its safe to read more than we are asked, caller should ensure that
 787          * destination has enough space.
 788          */
 789         offset = src_paddr & 15;
 790         src_paddr = round_down(src_paddr, 16);
 791         sz = round_up(sz + offset, 16);
 792
 793         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 794 }
 795
 796 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 797                                   void __user *dst_uaddr,
 798                                   unsigned long dst_paddr,
 799                                   int size, int *err)
 800 {
 801         struct page *tpage = NULL;
 802         int ret, offset;
 803
 804         /* if inputs are not 16-byte then use intermediate buffer */
 805         if (!IS_ALIGNED(dst_paddr, 16) ||
 806             !IS_ALIGNED(paddr,     16) ||
 807             !IS_ALIGNED(size,      16)) {
 808                 tpage = (void *)alloc_page(GFP_KERNEL);
 809                 if (!tpage)
 810                         return -ENOMEM;
 811
 812                 dst_paddr = __sme_page_pa(tpage);
 813         }
 814
 815         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 816         if (ret)
 817                 goto e_free;
 818
 819         if (tpage) {
 820                 offset = paddr & 15;
 821                 if (copy_to_user(dst_uaddr, page_address(tpage) + offset, size))
 822                         ret = -EFAULT;
 823         }
 824
 825 e_free:
 826         if (tpage)
 827                 __free_page(tpage);
 828
 829         return ret;
 830 }
 831
 832 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 833                                   void __user *vaddr,
 834                                   unsigned long dst_paddr,
 835                                   void __user *dst_vaddr,
 836                                   int size, int *error)
 837 {
 838         struct page *src_tpage = NULL;
 839         struct page *dst_tpage = NULL;
 840         int ret, len = size;
 841
 842         /* If source buffer is not aligned then use an intermediate buffer */
 843         if (!IS_ALIGNED((unsigned long)vaddr, 16)) {
 844                 src_tpage = alloc_page(GFP_KERNEL);
 845                 if (!src_tpage)
 846                         return -ENOMEM;
 847
 848                 if (copy_from_user(page_address(src_tpage), vaddr, size)) {
 849                         __free_page(src_tpage);
 850                         return -EFAULT;
 851                 }
 852
 853                 paddr = __sme_page_pa(src_tpage);
 854         }
 855
 856         /*
 857          *  If destination buffer or length is not aligned then do read-modify-write:
 858          *   - decrypt destination in an intermediate buffer
 859          *   - copy the source buffer in an intermediate buffer
 860          *   - use the intermediate buffer as source buffer
 861          */
 862         if (!IS_ALIGNED((unsigned long)dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 863                 int dst_offset;
 864
 865                 dst_tpage = alloc_page(GFP_KERNEL);
 866                 if (!dst_tpage) {
 867                         ret = -ENOMEM;
 868                         goto e_free;
 869                 }
 870
 871                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 872                                         __sme_page_pa(dst_tpage), size, error);
 873                 if (ret)
 874                         goto e_free;
 875
 876                 /*
 877                  *  If source is kernel buffer then use memcpy() otherwise
 878                  *  copy_from_user().
 879                  */
 880                 dst_offset = dst_paddr & 15;
 881
 882                 if (src_tpage)
 883                         memcpy(page_address(dst_tpage) + dst_offset,
 884                                page_address(src_tpage), size);
 885                 else {
 886                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 887                                            vaddr, size)) {
 888                                 ret = -EFAULT;
 889                                 goto e_free;
 890                         }
 891                 }
 892
 893                 paddr = __sme_page_pa(dst_tpage);
 894                 dst_paddr = round_down(dst_paddr, 16);
 895                 len = round_up(size, 16);
 896         }
 897
 898         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 899
 900 e_free:
 901         if (src_tpage)
 902                 __free_page(src_tpage);
 903         if (dst_tpage)
 904                 __free_page(dst_tpage);
 905         return ret;
 906 }
 907
 908 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 909 {
 910         unsigned long vaddr, vaddr_end, next_vaddr;
 911         unsigned long dst_vaddr;
 912         struct page **src_p, **dst_p;
 913         struct kvm_sev_dbg debug;
 914         unsigned long n;
 915         unsigned int size;
 916         int ret;
 917
 918         if (!sev_guest(kvm))
 919                 return -ENOTTY;
 920
 921         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 922                 return -EFAULT;
 923
 924         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 925                 return -EINVAL;
 926         if (!debug.dst_uaddr)
 927                 return -EINVAL;
 928
 929         vaddr = debug.src_uaddr;
 930         size = debug.len;
 931         vaddr_end = vaddr + size;
 932         dst_vaddr = debug.dst_uaddr;
 933
 934         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 935                 int len, s_off, d_off;
 936
 937                 /* lock userspace source and destination page */
 938                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 939                 if (IS_ERR(src_p))
 940                         return PTR_ERR(src_p);
 941
 942                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 943                 if (IS_ERR(dst_p)) {
 944                         sev_unpin_memory(kvm, src_p, n);
 945                         return PTR_ERR(dst_p);
 946                 }
 947
 948                 /*
 949                  * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
 950                  * the pages; flush the destination too so that future accesses do not
 951                  * see stale data.
 952                  */
 953                 sev_clflush_pages(src_p, 1);
 954                 sev_clflush_pages(dst_p, 1);
 955
 956                 /*
 957                  * Since user buffer may not be page aligned, calculate the
 958                  * offset within the page.
 959                  */
 960                 s_off = vaddr & ~PAGE_MASK;
 961                 d_off = dst_vaddr & ~PAGE_MASK;
 962                 len = min_t(size_t, (PAGE_SIZE - s_off), size);
 963
 964                 if (dec)
 965                         ret = __sev_dbg_decrypt_user(kvm,
 966                                                      __sme_page_pa(src_p[0]) + s_off,
 967                                                      (void __user *)dst_vaddr,
 968                                                      __sme_page_pa(dst_p[0]) + d_off,
 969                                                      len, &argp->error);
 970                 else
 971                         ret = __sev_dbg_encrypt_user(kvm,
 972                                                      __sme_page_pa(src_p[0]) + s_off,
 973                                                      (void __user *)vaddr,
 974                                                      __sme_page_pa(dst_p[0]) + d_off,
 975                                                      (void __user *)dst_vaddr,
 976                                                      len, &argp->error);
 977
 978                 sev_unpin_memory(kvm, src_p, n);
 979                 sev_unpin_memory(kvm, dst_p, n);
 980
 981                 if (ret)
 982                         goto err;
 983
 984                 next_vaddr = vaddr + len;
 985                 dst_vaddr = dst_vaddr + len;
 986                 size -= len;
 987         }
 988 err:
 989         return ret;
 990 }
 991
 992 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
 993 {
 994         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 995         struct sev_data_launch_secret data;
 996         struct kvm_sev_launch_secret params;
 997         struct page **pages;
 998         void *blob, *hdr;
 999         unsigned long n, i;
1000         int ret, offset;
1001
1002         if (!sev_guest(kvm))
1003                 return -ENOTTY;
1004
1005         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1006                 return -EFAULT;
1007
1008         pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, 1);
1009         if (IS_ERR(pages))
1010                 return PTR_ERR(pages);
1011
1012         /*
1013          * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
1014          * place; the cache may contain the data that was written unencrypted.
1015          */
1016         sev_clflush_pages(pages, n);
1017
1018         /*
1019          * The secret must be copied into contiguous memory region, lets verify
1020          * that userspace memory pages are contiguous before we issue command.
1021          */
1022         if (get_num_contig_pages(0, pages, n) != n) {
1023                 ret = -EINVAL;
1024                 goto e_unpin_memory;
1025         }
1026
1027         memset(&data, 0, sizeof(data));
1028
1029         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1030         data.guest_address = __sme_page_pa(pages[0]) + offset;
1031         data.guest_len = params.guest_len;
1032
1033         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1034         if (IS_ERR(blob)) {
1035                 ret = PTR_ERR(blob);
1036                 goto e_unpin_memory;
1037         }
1038
1039         data.trans_address = __psp_pa(blob);
1040         data.trans_len = params.trans_len;
1041
1042         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1043         if (IS_ERR(hdr)) {
1044                 ret = PTR_ERR(hdr);
1045                 goto e_free_blob;
1046         }
1047         data.hdr_address = __psp_pa(hdr);
1048         data.hdr_len = params.hdr_len;
1049
1050         data.handle = sev->handle;
1051         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, &data, &argp->error);
1052
1053         kfree(hdr);
1054
1055 e_free_blob:
1056         kfree(blob);
1057 e_unpin_memory:
1058         /* content of memory is updated, mark pages dirty */
1059         for (i = 0; i < n; i++) {
1060                 set_page_dirty_lock(pages[i]);
1061                 mark_page_accessed(pages[i]);
1062         }
1063         sev_unpin_memory(kvm, pages, n);
1064         return ret;
1065 }
1066
1067 static int sev_get_attestation_report(struct kvm *kvm, struct kvm_sev_cmd *argp)
1068 {
1069         void __user *report = (void __user *)(uintptr_t)argp->data;
1070         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1071         struct sev_data_attestation_report data;
1072         struct kvm_sev_attestation_report params;
1073         void __user *p;
1074         void *blob = NULL;
1075         int ret;
1076
1077         if (!sev_guest(kvm))
1078                 return -ENOTTY;
1079
1080         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1081                 return -EFAULT;
1082
1083         memset(&data, 0, sizeof(data));
1084
1085         /* User wants to query the blob length */
1086         if (!params.len)
1087                 goto cmd;
1088
1089         p = (void __user *)(uintptr_t)params.uaddr;
1090         if (p) {
1091                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
1092                         return -EINVAL;
1093
1094                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
1095                 if (!blob)
1096                         return -ENOMEM;
1097
1098                 data.address = __psp_pa(blob);
1099                 data.len = params.len;
1100                 memcpy(data.mnonce, params.mnonce, sizeof(params.mnonce));
1101         }
1102 cmd:
1103         data.handle = sev->handle;
1104         ret = sev_issue_cmd(kvm, SEV_CMD_ATTESTATION_REPORT, &data, &argp->error);
1105         /*
1106          * If we query the session length, FW responded with expected data.
1107          */
1108         if (!params.len)
1109                 goto done;
1110
1111         if (ret)
1112                 goto e_free_blob;
1113
1114         if (blob) {
1115                 if (copy_to_user(p, blob, params.len))
1116                         ret = -EFAULT;
1117         }
1118
1119 done:
1120         params.len = data.len;
1121         if (copy_to_user(report, &params, sizeof(params)))
1122                 ret = -EFAULT;
1123 e_free_blob:
1124         kfree(blob);
1125         return ret;
1126 }
1127
1128 /* Userspace wants to query session length. */
1129 static int
1130 __sev_send_start_query_session_length(struct kvm *kvm, struct kvm_sev_cmd *argp,
1131                                       struct kvm_sev_send_start *params)
1132 {
1133         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1134         struct sev_data_send_start data;
1135         int ret;
1136
1137         memset(&data, 0, sizeof(data));
1138         data.handle = sev->handle;
1139         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1140
1141         params->session_len = data.session_len;
1142         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1143                                 sizeof(struct kvm_sev_send_start)))
1144                 ret = -EFAULT;
1145
1146         return ret;
1147 }
1148
1149 static int sev_send_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1150 {
1151         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1152         struct sev_data_send_start data;
1153         struct kvm_sev_send_start params;
1154         void *amd_certs, *session_data;
1155         void *pdh_cert, *plat_certs;
1156         int ret;
1157
1158         if (!sev_guest(kvm))
1159                 return -ENOTTY;
1160
1161         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1162                                 sizeof(struct kvm_sev_send_start)))
1163                 return -EFAULT;
1164
1165         /* if session_len is zero, userspace wants to query the session length */
1166         if (!params.session_len)
1167                 return __sev_send_start_query_session_length(kvm, argp,
1168                                 &params);
1169
1170         /* some sanity checks */
1171         if (!params.pdh_cert_uaddr || !params.pdh_cert_len ||
1172             !params.session_uaddr || params.session_len > SEV_FW_BLOB_MAX_SIZE)
1173                 return -EINVAL;
1174
1175         /* allocate the memory to hold the session data blob */
1176         session_data = kmalloc(params.session_len, GFP_KERNEL_ACCOUNT);
1177         if (!session_data)
1178                 return -ENOMEM;
1179
1180         /* copy the certificate blobs from userspace */
1181         pdh_cert = psp_copy_user_blob(params.pdh_cert_uaddr,
1182                                 params.pdh_cert_len);
1183         if (IS_ERR(pdh_cert)) {
1184                 ret = PTR_ERR(pdh_cert);
1185                 goto e_free_session;
1186         }
1187
1188         plat_certs = psp_copy_user_blob(params.plat_certs_uaddr,
1189                                 params.plat_certs_len);
1190         if (IS_ERR(plat_certs)) {
1191                 ret = PTR_ERR(plat_certs);
1192                 goto e_free_pdh;
1193         }
1194
1195         amd_certs = psp_copy_user_blob(params.amd_certs_uaddr,
1196                                 params.amd_certs_len);
1197         if (IS_ERR(amd_certs)) {
1198                 ret = PTR_ERR(amd_certs);
1199                 goto e_free_plat_cert;
1200         }
1201
1202         /* populate the FW SEND_START field with system physical address */
1203         memset(&data, 0, sizeof(data));
1204         data.pdh_cert_address = __psp_pa(pdh_cert);
1205         data.pdh_cert_len = params.pdh_cert_len;
1206         data.plat_certs_address = __psp_pa(plat_certs);
1207         data.plat_certs_len = params.plat_certs_len;
1208         data.amd_certs_address = __psp_pa(amd_certs);
1209         data.amd_certs_len = params.amd_certs_len;
1210         data.session_address = __psp_pa(session_data);
1211         data.session_len = params.session_len;
1212         data.handle = sev->handle;
1213
1214         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1215
1216         if (!ret && copy_to_user((void __user *)(uintptr_t)params.session_uaddr,
1217                         session_data, params.session_len)) {
1218                 ret = -EFAULT;
1219                 goto e_free_amd_cert;
1220         }
1221
1222         params.policy = data.policy;
1223         params.session_len = data.session_len;
1224         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params,
1225                                 sizeof(struct kvm_sev_send_start)))
1226                 ret = -EFAULT;
1227
1228 e_free_amd_cert:
1229         kfree(amd_certs);
1230 e_free_plat_cert:
1231         kfree(plat_certs);
1232 e_free_pdh:
1233         kfree(pdh_cert);
1234 e_free_session:
1235         kfree(session_data);
1236         return ret;
1237 }
1238
1239 /* Userspace wants to query either header or trans length. */
1240 static int
1241 __sev_send_update_data_query_lengths(struct kvm *kvm, struct kvm_sev_cmd *argp,
1242                                      struct kvm_sev_send_update_data *params)
1243 {
1244         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1245         struct sev_data_send_update_data data;
1246         int ret;
1247
1248         memset(&data, 0, sizeof(data));
1249         data.handle = sev->handle;
1250         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1251
1252         params->hdr_len = data.hdr_len;
1253         params->trans_len = data.trans_len;
1254
1255         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1256                          sizeof(struct kvm_sev_send_update_data)))
1257                 ret = -EFAULT;
1258
1259         return ret;
1260 }
1261
1262 static int sev_send_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1263 {
1264         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1265         struct sev_data_send_update_data data;
1266         struct kvm_sev_send_update_data params;
1267         void *hdr, *trans_data;
1268         struct page **guest_page;
1269         unsigned long n;
1270         int ret, offset;
1271
1272         if (!sev_guest(kvm))
1273                 return -ENOTTY;
1274
1275         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1276                         sizeof(struct kvm_sev_send_update_data)))
1277                 return -EFAULT;
1278
1279         /* userspace wants to query either header or trans length */
1280         if (!params.trans_len || !params.hdr_len)
1281                 return __sev_send_update_data_query_lengths(kvm, argp, &params);
1282
1283         if (!params.trans_uaddr || !params.guest_uaddr ||
1284             !params.guest_len || !params.hdr_uaddr)
1285                 return -EINVAL;
1286
1287         /* Check if we are crossing the page boundary */
1288         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1289         if ((params.guest_len + offset > PAGE_SIZE))
1290                 return -EINVAL;
1291
1292         /* Pin guest memory */
1293         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1294                                     PAGE_SIZE, &n, 0);
1295         if (IS_ERR(guest_page))
1296                 return PTR_ERR(guest_page);
1297
1298         /* allocate memory for header and transport buffer */
1299         ret = -ENOMEM;
1300         hdr = kmalloc(params.hdr_len, GFP_KERNEL_ACCOUNT);
1301         if (!hdr)
1302                 goto e_unpin;
1303
1304         trans_data = kmalloc(params.trans_len, GFP_KERNEL_ACCOUNT);
1305         if (!trans_data)
1306                 goto e_free_hdr;
1307
1308         memset(&data, 0, sizeof(data));
1309         data.hdr_address = __psp_pa(hdr);
1310         data.hdr_len = params.hdr_len;
1311         data.trans_address = __psp_pa(trans_data);
1312         data.trans_len = params.trans_len;
1313
1314         /* The SEND_UPDATE_DATA command requires C-bit to be always set. */
1315         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1316         data.guest_address |= sev_me_mask;
1317         data.guest_len = params.guest_len;
1318         data.handle = sev->handle;
1319
1320         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1321
1322         if (ret)
1323                 goto e_free_trans_data;
1324
1325         /* copy transport buffer to user space */
1326         if (copy_to_user((void __user *)(uintptr_t)params.trans_uaddr,
1327                          trans_data, params.trans_len)) {
1328                 ret = -EFAULT;
1329                 goto e_free_trans_data;
1330         }
1331
1332         /* Copy packet header to userspace. */
1333         if (copy_to_user((void __user *)(uintptr_t)params.hdr_uaddr, hdr,
1334                          params.hdr_len))
1335                 ret = -EFAULT;
1336
1337 e_free_trans_data:
1338         kfree(trans_data);
1339 e_free_hdr:
1340         kfree(hdr);
1341 e_unpin:
1342         sev_unpin_memory(kvm, guest_page, n);
1343
1344         return ret;
1345 }
1346
1347 static int sev_send_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1348 {
1349         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1350         struct sev_data_send_finish data;
1351
1352         if (!sev_guest(kvm))
1353                 return -ENOTTY;
1354
1355         data.handle = sev->handle;
1356         return sev_issue_cmd(kvm, SEV_CMD_SEND_FINISH, &data, &argp->error);
1357 }
1358
1359 static int sev_send_cancel(struct kvm *kvm, struct kvm_sev_cmd *argp)
1360 {
1361         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1362         struct sev_data_send_cancel data;
1363
1364         if (!sev_guest(kvm))
1365                 return -ENOTTY;
1366
1367         data.handle = sev->handle;
1368         return sev_issue_cmd(kvm, SEV_CMD_SEND_CANCEL, &data, &argp->error);
1369 }
1370
1371 static int sev_receive_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1372 {
1373         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1374         struct sev_data_receive_start start;
1375         struct kvm_sev_receive_start params;
1376         int *error = &argp->error;
1377         void *session_data;
1378         void *pdh_data;
1379         int ret;
1380
1381         if (!sev_guest(kvm))
1382                 return -ENOTTY;
1383
1384         /* Get parameter from the userspace */
1385         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1386                         sizeof(struct kvm_sev_receive_start)))
1387                 return -EFAULT;
1388
1389         /* some sanity checks */
1390         if (!params.pdh_uaddr || !params.pdh_len ||
1391             !params.session_uaddr || !params.session_len)
1392                 return -EINVAL;
1393
1394         pdh_data = psp_copy_user_blob(params.pdh_uaddr, params.pdh_len);
1395         if (IS_ERR(pdh_data))
1396                 return PTR_ERR(pdh_data);
1397
1398         session_data = psp_copy_user_blob(params.session_uaddr,
1399                         params.session_len);
1400         if (IS_ERR(session_data)) {
1401                 ret = PTR_ERR(session_data);
1402                 goto e_free_pdh;
1403         }
1404
1405         memset(&start, 0, sizeof(start));
1406         start.handle = params.handle;
1407         start.policy = params.policy;
1408         start.pdh_cert_address = __psp_pa(pdh_data);
1409         start.pdh_cert_len = params.pdh_len;
1410         start.session_address = __psp_pa(session_data);
1411         start.session_len = params.session_len;
1412
1413         /* create memory encryption context */
1414         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_RECEIVE_START, &start,
1415                                 error);
1416         if (ret)
1417                 goto e_free_session;
1418
1419         /* Bind ASID to this guest */
1420         ret = sev_bind_asid(kvm, start.handle, error);
1421         if (ret) {
1422                 sev_decommission(start.handle);
1423                 goto e_free_session;
1424         }
1425
1426         params.handle = start.handle;
1427         if (copy_to_user((void __user *)(uintptr_t)argp->data,
1428                          &params, sizeof(struct kvm_sev_receive_start))) {
1429                 ret = -EFAULT;
1430                 sev_unbind_asid(kvm, start.handle);
1431                 goto e_free_session;
1432         }
1433
1434         sev->handle = start.handle;
1435         sev->fd = argp->sev_fd;
1436
1437 e_free_session:
1438         kfree(session_data);
1439 e_free_pdh:
1440         kfree(pdh_data);
1441
1442         return ret;
1443 }
1444
1445 static int sev_receive_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1446 {
1447         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1448         struct kvm_sev_receive_update_data params;
1449         struct sev_data_receive_update_data data;
1450         void *hdr = NULL, *trans = NULL;
1451         struct page **guest_page;
1452         unsigned long n;
1453         int ret, offset;
1454
1455         if (!sev_guest(kvm))
1456                 return -EINVAL;
1457
1458         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1459                         sizeof(struct kvm_sev_receive_update_data)))
1460                 return -EFAULT;
1461
1462         if (!params.hdr_uaddr || !params.hdr_len ||
1463             !params.guest_uaddr || !params.guest_len ||
1464             !params.trans_uaddr || !params.trans_len)
1465                 return -EINVAL;
1466
1467         /* Check if we are crossing the page boundary */
1468         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1469         if ((params.guest_len + offset > PAGE_SIZE))
1470                 return -EINVAL;
1471
1472         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1473         if (IS_ERR(hdr))
1474                 return PTR_ERR(hdr);
1475
1476         trans = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1477         if (IS_ERR(trans)) {
1478                 ret = PTR_ERR(trans);
1479                 goto e_free_hdr;
1480         }
1481
1482         memset(&data, 0, sizeof(data));
1483         data.hdr_address = __psp_pa(hdr);
1484         data.hdr_len = params.hdr_len;
1485         data.trans_address = __psp_pa(trans);
1486         data.trans_len = params.trans_len;
1487
1488         /* Pin guest memory */
1489         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1490                                     PAGE_SIZE, &n, 1);
1491         if (IS_ERR(guest_page)) {
1492                 ret = PTR_ERR(guest_page);
1493                 goto e_free_trans;
1494         }
1495
1496         /*
1497          * Flush (on non-coherent CPUs) before RECEIVE_UPDATE_DATA, the PSP
1498          * encrypts the written data with the guest's key, and the cache may
1499          * contain dirty, unencrypted data.
1500          */
1501         sev_clflush_pages(guest_page, n);
1502
1503         /* The RECEIVE_UPDATE_DATA command requires C-bit to be always set. */
1504         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1505         data.guest_address |= sev_me_mask;
1506         data.guest_len = params.guest_len;
1507         data.handle = sev->handle;
1508
1509         ret = sev_issue_cmd(kvm, SEV_CMD_RECEIVE_UPDATE_DATA, &data,
1510                                 &argp->error);
1511
1512         sev_unpin_memory(kvm, guest_page, n);
1513
1514 e_free_trans:
1515         kfree(trans);
1516 e_free_hdr:
1517         kfree(hdr);
1518
1519         return ret;
1520 }
1521
1522 static int sev_receive_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1523 {
1524         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1525         struct sev_data_receive_finish data;
1526
1527         if (!sev_guest(kvm))
1528                 return -ENOTTY;
1529
1530         data.handle = sev->handle;
1531         return sev_issue_cmd(kvm, SEV_CMD_RECEIVE_FINISH, &data, &argp->error);
1532 }
1533
1534 static bool is_cmd_allowed_from_mirror(u32 cmd_id)
1535 {
1536         /*
1537          * Allow mirrors VM to call KVM_SEV_LAUNCH_UPDATE_VMSA to enable SEV-ES
1538          * active mirror VMs. Also allow the debugging and status commands.
1539          */
1540         if (cmd_id == KVM_SEV_LAUNCH_UPDATE_VMSA ||
1541             cmd_id == KVM_SEV_GUEST_STATUS || cmd_id == KVM_SEV_DBG_DECRYPT ||
1542             cmd_id == KVM_SEV_DBG_ENCRYPT)
1543                 return true;
1544
1545         return false;
1546 }
1547
1548 static int sev_lock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1549 {
1550         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1551         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1552         int r = -EBUSY;
1553
1554         if (dst_kvm == src_kvm)
1555                 return -EINVAL;
1556
1557         /*
1558          * Bail if these VMs are already involved in a migration to avoid
1559          * deadlock between two VMs trying to migrate to/from each other.
1560          */
1561         if (atomic_cmpxchg_acquire(&dst_sev->migration_in_progress, 0, 1))
1562                 return -EBUSY;
1563
1564         if (atomic_cmpxchg_acquire(&src_sev->migration_in_progress, 0, 1))
1565                 goto release_dst;
1566
1567         r = -EINTR;
1568         if (mutex_lock_killable(&dst_kvm->lock))
1569                 goto release_src;
1570         if (mutex_lock_killable_nested(&src_kvm->lock, SINGLE_DEPTH_NESTING))
1571                 goto unlock_dst;
1572         return 0;
1573
1574 unlock_dst:
1575         mutex_unlock(&dst_kvm->lock);
1576 release_src:
1577         atomic_set_release(&src_sev->migration_in_progress, 0);
1578 release_dst:
1579         atomic_set_release(&dst_sev->migration_in_progress, 0);
1580         return r;
1581 }
1582
1583 static void sev_unlock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1584 {
1585         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1586         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1587
1588         mutex_unlock(&dst_kvm->lock);
1589         mutex_unlock(&src_kvm->lock);
1590         atomic_set_release(&dst_sev->migration_in_progress, 0);
1591         atomic_set_release(&src_sev->migration_in_progress, 0);
1592 }
1593
1594
1595 static int sev_lock_vcpus_for_migration(struct kvm *kvm)
1596 {
1597         struct kvm_vcpu *vcpu;
1598         unsigned long i, j;
1599
1600         kvm_for_each_vcpu(i, vcpu, kvm) {
1601                 if (mutex_lock_killable(&vcpu->mutex))
1602                         goto out_unlock;
1603         }
1604
1605         return 0;
1606
1607 out_unlock:
1608         kvm_for_each_vcpu(j, vcpu, kvm) {
1609                 if (i == j)
1610                         break;
1611
1612                 mutex_unlock(&vcpu->mutex);
1613         }
1614         return -EINTR;
1615 }
1616
1617 static void sev_unlock_vcpus_for_migration(struct kvm *kvm)
1618 {
1619         struct kvm_vcpu *vcpu;
1620         unsigned long i;
1621
1622         kvm_for_each_vcpu(i, vcpu, kvm) {
1623                 mutex_unlock(&vcpu->mutex);
1624         }
1625 }
1626
1627 static void sev_migrate_from(struct kvm *dst_kvm, struct kvm *src_kvm)
1628 {
1629         struct kvm_sev_info *dst = &to_kvm_svm(dst_kvm)->sev_info;
1630         struct kvm_sev_info *src = &to_kvm_svm(src_kvm)->sev_info;
1631         struct kvm_sev_info *mirror;
1632
1633         dst->active = true;
1634         dst->asid = src->asid;
1635         dst->handle = src->handle;
1636         dst->pages_locked = src->pages_locked;
1637         dst->enc_context_owner = src->enc_context_owner;
1638
1639         src->asid = 0;
1640         src->active = false;
1641         src->handle = 0;
1642         src->pages_locked = 0;
1643         src->enc_context_owner = NULL;
1644
1645         list_cut_before(&dst->regions_list, &src->regions_list, &src->regions_list);
1646
1647         /*
1648          * If this VM has mirrors, "transfer" each mirror's refcount of the
1649          * source to the destination (this KVM).  The caller holds a reference
1650          * to the source, so there's no danger of use-after-free.
1651          */
1652         list_cut_before(&dst->mirror_vms, &src->mirror_vms, &src->mirror_vms);
1653         list_for_each_entry(mirror, &dst->mirror_vms, mirror_entry) {
1654                 kvm_get_kvm(dst_kvm);
1655                 kvm_put_kvm(src_kvm);
1656                 mirror->enc_context_owner = dst_kvm;
1657         }
1658
1659         /*
1660          * If this VM is a mirror, remove the old mirror from the owners list
1661          * and add the new mirror to the list.
1662          */
1663         if (is_mirroring_enc_context(dst_kvm)) {
1664                 struct kvm_sev_info *owner_sev_info =
1665                         &to_kvm_svm(dst->enc_context_owner)->sev_info;
1666
1667                 list_del(&src->mirror_entry);
1668                 list_add_tail(&dst->mirror_entry, &owner_sev_info->mirror_vms);
1669         }
1670 }
1671
1672 static int sev_es_migrate_from(struct kvm *dst, struct kvm *src)
1673 {
1674         unsigned long i;
1675         struct kvm_vcpu *dst_vcpu, *src_vcpu;
1676         struct vcpu_svm *dst_svm, *src_svm;
1677
1678         if (atomic_read(&src->online_vcpus) != atomic_read(&dst->online_vcpus))
1679                 return -EINVAL;
1680
1681         kvm_for_each_vcpu(i, src_vcpu, src) {
1682                 if (!src_vcpu->arch.guest_state_protected)
1683                         return -EINVAL;
1684         }
1685
1686         kvm_for_each_vcpu(i, src_vcpu, src) {
1687                 src_svm = to_svm(src_vcpu);
1688                 dst_vcpu = kvm_get_vcpu(dst, i);
1689                 dst_svm = to_svm(dst_vcpu);
1690
1691                 /*
1692                  * Transfer VMSA and GHCB state to the destination.  Nullify and
1693                  * clear source fields as appropriate, the state now belongs to
1694                  * the destination.
1695                  */
1696                 memcpy(&dst_svm->sev_es, &src_svm->sev_es, sizeof(src_svm->sev_es));
1697                 dst_svm->vmcb->control.ghcb_gpa = src_svm->vmcb->control.ghcb_gpa;
1698                 dst_svm->vmcb->control.vmsa_pa = src_svm->vmcb->control.vmsa_pa;
1699                 dst_vcpu->arch.guest_state_protected = true;
1700
1701                 memset(&src_svm->sev_es, 0, sizeof(src_svm->sev_es));
1702                 src_svm->vmcb->control.ghcb_gpa = INVALID_PAGE;
1703                 src_svm->vmcb->control.vmsa_pa = INVALID_PAGE;
1704                 src_vcpu->arch.guest_state_protected = false;
1705         }
1706         to_kvm_svm(src)->sev_info.es_active = false;
1707         to_kvm_svm(dst)->sev_info.es_active = true;
1708
1709         return 0;
1710 }
1711
1712 int sev_vm_move_enc_context_from(struct kvm *kvm, unsigned int source_fd)
1713 {
1714         struct kvm_sev_info *dst_sev = &to_kvm_svm(kvm)->sev_info;
1715         struct kvm_sev_info *src_sev, *cg_cleanup_sev;
1716         struct file *source_kvm_file;
1717         struct kvm *source_kvm;
1718         bool charged = false;
1719         int ret;
1720
1721         source_kvm_file = fget(source_fd);
1722         if (!file_is_kvm(source_kvm_file)) {
1723                 ret = -EBADF;
1724                 goto out_fput;
1725         }
1726
1727         source_kvm = source_kvm_file->private_data;
1728         ret = sev_lock_two_vms(kvm, source_kvm);
1729         if (ret)
1730                 goto out_fput;
1731
1732         if (sev_guest(kvm) || !sev_guest(source_kvm)) {
1733                 ret = -EINVAL;
1734                 goto out_unlock;
1735         }
1736
1737         src_sev = &to_kvm_svm(source_kvm)->sev_info;
1738
1739         dst_sev->misc_cg = get_current_misc_cg();
1740         cg_cleanup_sev = dst_sev;
1741         if (dst_sev->misc_cg != src_sev->misc_cg) {
1742                 ret = sev_misc_cg_try_charge(dst_sev);
1743                 if (ret)
1744                         goto out_dst_cgroup;
1745                 charged = true;
1746         }
1747
1748         ret = sev_lock_vcpus_for_migration(kvm);
1749         if (ret)
1750                 goto out_dst_cgroup;
1751         ret = sev_lock_vcpus_for_migration(source_kvm);
1752         if (ret)
1753                 goto out_dst_vcpu;
1754
1755         if (sev_es_guest(source_kvm)) {
1756                 ret = sev_es_migrate_from(kvm, source_kvm);
1757                 if (ret)
1758                         goto out_source_vcpu;
1759         }
1760
1761         sev_migrate_from(kvm, source_kvm);
1762         kvm_vm_dead(source_kvm);
1763         cg_cleanup_sev = src_sev;
1764         ret = 0;
1765
1766 out_source_vcpu:
1767         sev_unlock_vcpus_for_migration(source_kvm);
1768 out_dst_vcpu:
1769         sev_unlock_vcpus_for_migration(kvm);
1770 out_dst_cgroup:
1771         /* Operates on the source on success, on the destination on failure.  */
1772         if (charged)
1773                 sev_misc_cg_uncharge(cg_cleanup_sev);
1774         put_misc_cg(cg_cleanup_sev->misc_cg);
1775         cg_cleanup_sev->misc_cg = NULL;
1776 out_unlock:
1777         sev_unlock_two_vms(kvm, source_kvm);
1778 out_fput:
1779         if (source_kvm_file)
1780                 fput(source_kvm_file);
1781         return ret;
1782 }
1783
1784 int sev_mem_enc_ioctl(struct kvm *kvm, void __user *argp)
1785 {
1786         struct kvm_sev_cmd sev_cmd;
1787         int r;
1788
1789         if (!sev_enabled)
1790                 return -ENOTTY;
1791
1792         if (!argp)
1793                 return 0;
1794
1795         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
1796                 return -EFAULT;
1797
1798         mutex_lock(&kvm->lock);
1799
1800         /* Only the enc_context_owner handles some memory enc operations. */
1801         if (is_mirroring_enc_context(kvm) &&
1802             !is_cmd_allowed_from_mirror(sev_cmd.id)) {
1803                 r = -EINVAL;
1804                 goto out;
1805         }
1806
1807         switch (sev_cmd.id) {
1808         case KVM_SEV_ES_INIT:
1809                 if (!sev_es_enabled) {
1810                         r = -ENOTTY;
1811                         goto out;
1812                 }
1813                 fallthrough;
1814         case KVM_SEV_INIT:
1815                 r = sev_guest_init(kvm, &sev_cmd);
1816                 break;
1817         case KVM_SEV_LAUNCH_START:
1818                 r = sev_launch_start(kvm, &sev_cmd);
1819                 break;
1820         case KVM_SEV_LAUNCH_UPDATE_DATA:
1821                 r = sev_launch_update_data(kvm, &sev_cmd);
1822                 break;
1823         case KVM_SEV_LAUNCH_UPDATE_VMSA:
1824                 r = sev_launch_update_vmsa(kvm, &sev_cmd);
1825                 break;
1826         case KVM_SEV_LAUNCH_MEASURE:
1827                 r = sev_launch_measure(kvm, &sev_cmd);
1828                 break;
1829         case KVM_SEV_LAUNCH_FINISH:
1830                 r = sev_launch_finish(kvm, &sev_cmd);
1831                 break;
1832         case KVM_SEV_GUEST_STATUS:
1833                 r = sev_guest_status(kvm, &sev_cmd);
1834                 break;
1835         case KVM_SEV_DBG_DECRYPT:
1836                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
1837                 break;
1838         case KVM_SEV_DBG_ENCRYPT:
1839                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
1840                 break;
1841         case KVM_SEV_LAUNCH_SECRET:
1842                 r = sev_launch_secret(kvm, &sev_cmd);
1843                 break;
1844         case KVM_SEV_GET_ATTESTATION_REPORT:
1845                 r = sev_get_attestation_report(kvm, &sev_cmd);
1846                 break;
1847         case KVM_SEV_SEND_START:
1848                 r = sev_send_start(kvm, &sev_cmd);
1849                 break;
1850         case KVM_SEV_SEND_UPDATE_DATA:
1851                 r = sev_send_update_data(kvm, &sev_cmd);
1852                 break;
1853         case KVM_SEV_SEND_FINISH:
1854                 r = sev_send_finish(kvm, &sev_cmd);
1855                 break;
1856         case KVM_SEV_SEND_CANCEL:
1857                 r = sev_send_cancel(kvm, &sev_cmd);
1858                 break;
1859         case KVM_SEV_RECEIVE_START:
1860                 r = sev_receive_start(kvm, &sev_cmd);
1861                 break;
1862         case KVM_SEV_RECEIVE_UPDATE_DATA:
1863                 r = sev_receive_update_data(kvm, &sev_cmd);
1864                 break;
1865         case KVM_SEV_RECEIVE_FINISH:
1866                 r = sev_receive_finish(kvm, &sev_cmd);
1867                 break;
1868         default:
1869                 r = -EINVAL;
1870                 goto out;
1871         }
1872
1873         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
1874                 r = -EFAULT;
1875
1876 out:
1877         mutex_unlock(&kvm->lock);
1878         return r;
1879 }
1880
1881 int sev_mem_enc_register_region(struct kvm *kvm,
1882                                 struct kvm_enc_region *range)
1883 {
1884         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1885         struct enc_region *region;
1886         int ret = 0;
1887
1888         if (!sev_guest(kvm))
1889                 return -ENOTTY;
1890
1891         /* If kvm is mirroring encryption context it isn't responsible for it */
1892         if (is_mirroring_enc_context(kvm))
1893                 return -EINVAL;
1894
1895         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
1896                 return -EINVAL;
1897
1898         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
1899         if (!region)
1900                 return -ENOMEM;
1901
1902         mutex_lock(&kvm->lock);
1903         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
1904         if (IS_ERR(region->pages)) {
1905                 ret = PTR_ERR(region->pages);
1906                 mutex_unlock(&kvm->lock);
1907                 goto e_free;
1908         }
1909
1910         region->uaddr = range->addr;
1911         region->size = range->size;
1912
1913         list_add_tail(&region->list, &sev->regions_list);
1914         mutex_unlock(&kvm->lock);
1915
1916         /*
1917          * The guest may change the memory encryption attribute from C=0 -> C=1
1918          * or vice versa for this memory range. Lets make sure caches are
1919          * flushed to ensure that guest data gets written into memory with
1920          * correct C-bit.
1921          */
1922         sev_clflush_pages(region->pages, region->npages);
1923
1924         return ret;
1925
1926 e_free:
1927         kfree(region);
1928         return ret;
1929 }
1930
1931 static struct enc_region *
1932 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
1933 {
1934         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1935         struct list_head *head = &sev->regions_list;
1936         struct enc_region *i;
1937
1938         list_for_each_entry(i, head, list) {
1939                 if (i->uaddr == range->addr &&
1940                     i->size == range->size)
1941                         return i;
1942         }
1943
1944         return NULL;
1945 }
1946
1947 static void __unregister_enc_region_locked(struct kvm *kvm,
1948                                            struct enc_region *region)
1949 {
1950         sev_unpin_memory(kvm, region->pages, region->npages);
1951         list_del(&region->list);
1952         kfree(region);
1953 }
1954
1955 int sev_mem_enc_unregister_region(struct kvm *kvm,
1956                                   struct kvm_enc_region *range)
1957 {
1958         struct enc_region *region;
1959         int ret;
1960
1961         /* If kvm is mirroring encryption context it isn't responsible for it */
1962         if (is_mirroring_enc_context(kvm))
1963                 return -EINVAL;
1964
1965         mutex_lock(&kvm->lock);
1966
1967         if (!sev_guest(kvm)) {
1968                 ret = -ENOTTY;
1969                 goto failed;
1970         }
1971
1972         region = find_enc_region(kvm, range);
1973         if (!region) {
1974                 ret = -EINVAL;
1975                 goto failed;
1976         }
1977
1978         /*
1979          * Ensure that all guest tagged cache entries are flushed before
1980          * releasing the pages back to the system for use. CLFLUSH will
1981          * not do this, so issue a WBINVD.
1982          */
1983         wbinvd_on_all_cpus();
1984
1985         __unregister_enc_region_locked(kvm, region);
1986
1987         mutex_unlock(&kvm->lock);
1988         return 0;
1989
1990 failed:
1991         mutex_unlock(&kvm->lock);
1992         return ret;
1993 }
1994
1995 int sev_vm_copy_enc_context_from(struct kvm *kvm, unsigned int source_fd)
1996 {
1997         struct file *source_kvm_file;
1998         struct kvm *source_kvm;
1999         struct kvm_sev_info *source_sev, *mirror_sev;
2000         int ret;
2001
2002         source_kvm_file = fget(source_fd);
2003         if (!file_is_kvm(source_kvm_file)) {
2004                 ret = -EBADF;
2005                 goto e_source_fput;
2006         }
2007
2008         source_kvm = source_kvm_file->private_data;
2009         ret = sev_lock_two_vms(kvm, source_kvm);
2010         if (ret)
2011                 goto e_source_fput;
2012
2013         /*
2014          * Mirrors of mirrors should work, but let's not get silly.  Also
2015          * disallow out-of-band SEV/SEV-ES init if the target is already an
2016          * SEV guest, or if vCPUs have been created.  KVM relies on vCPUs being
2017          * created after SEV/SEV-ES initialization, e.g. to init intercepts.
2018          */
2019         if (sev_guest(kvm) || !sev_guest(source_kvm) ||
2020             is_mirroring_enc_context(source_kvm) || kvm->created_vcpus) {
2021                 ret = -EINVAL;
2022                 goto e_unlock;
2023         }
2024
2025         /*
2026          * The mirror kvm holds an enc_context_owner ref so its asid can't
2027          * disappear until we're done with it
2028          */
2029         source_sev = &to_kvm_svm(source_kvm)->sev_info;
2030         kvm_get_kvm(source_kvm);
2031         mirror_sev = &to_kvm_svm(kvm)->sev_info;
2032         list_add_tail(&mirror_sev->mirror_entry, &source_sev->mirror_vms);
2033
2034         /* Set enc_context_owner and copy its encryption context over */
2035         mirror_sev->enc_context_owner = source_kvm;
2036         mirror_sev->active = true;
2037         mirror_sev->asid = source_sev->asid;
2038         mirror_sev->fd = source_sev->fd;
2039         mirror_sev->es_active = source_sev->es_active;
2040         mirror_sev->handle = source_sev->handle;
2041         INIT_LIST_HEAD(&mirror_sev->regions_list);
2042         INIT_LIST_HEAD(&mirror_sev->mirror_vms);
2043         ret = 0;
2044
2045         /*
2046          * Do not copy ap_jump_table. Since the mirror does not share the same
2047          * KVM contexts as the original, and they may have different
2048          * memory-views.
2049          */
2050
2051 e_unlock:
2052         sev_unlock_two_vms(kvm, source_kvm);
2053 e_source_fput:
2054         if (source_kvm_file)
2055                 fput(source_kvm_file);
2056         return ret;
2057 }
2058
2059 void sev_vm_destroy(struct kvm *kvm)
2060 {
2061         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
2062         struct list_head *head = &sev->regions_list;
2063         struct list_head *pos, *q;
2064
2065         if (!sev_guest(kvm))
2066                 return;
2067
2068         WARN_ON(!list_empty(&sev->mirror_vms));
2069
2070         /* If this is a mirror_kvm release the enc_context_owner and skip sev cleanup */
2071         if (is_mirroring_enc_context(kvm)) {
2072                 struct kvm *owner_kvm = sev->enc_context_owner;
2073
2074                 mutex_lock(&owner_kvm->lock);
2075                 list_del(&sev->mirror_entry);
2076                 mutex_unlock(&owner_kvm->lock);
2077                 kvm_put_kvm(owner_kvm);
2078                 return;
2079         }
2080
2081         /*
2082          * Ensure that all guest tagged cache entries are flushed before
2083          * releasing the pages back to the system for use. CLFLUSH will
2084          * not do this, so issue a WBINVD.
2085          */
2086         wbinvd_on_all_cpus();
2087
2088         /*
2089          * if userspace was terminated before unregistering the memory regions
2090          * then lets unpin all the registered memory.
2091          */
2092         if (!list_empty(head)) {
2093                 list_for_each_safe(pos, q, head) {
2094                         __unregister_enc_region_locked(kvm,
2095                                 list_entry(pos, struct enc_region, list));
2096                         cond_resched();
2097                 }
2098         }
2099
2100         sev_unbind_asid(kvm, sev->handle);
2101         sev_asid_free(sev);
2102 }
2103
2104 void __init sev_set_cpu_caps(void)
2105 {
2106         if (!sev_enabled)
2107                 kvm_cpu_cap_clear(X86_FEATURE_SEV);
2108         if (!sev_es_enabled)
2109                 kvm_cpu_cap_clear(X86_FEATURE_SEV_ES);
2110 }
2111
2112 void __init sev_hardware_setup(void)
2113 {
2114 #ifdef CONFIG_KVM_AMD_SEV
2115         unsigned int eax, ebx, ecx, edx, sev_asid_count, sev_es_asid_count;
2116         bool sev_es_supported = false;
2117         bool sev_supported = false;
2118
2119         if (!sev_enabled || !npt_enabled)
2120                 goto out;
2121
2122         /*
2123          * SEV must obviously be supported in hardware.  Sanity check that the
2124          * CPU supports decode assists, which is mandatory for SEV guests to
2125          * support instruction emulation.
2126          */
2127         if (!boot_cpu_has(X86_FEATURE_SEV) ||
2128             WARN_ON_ONCE(!boot_cpu_has(X86_FEATURE_DECODEASSISTS)))
2129                 goto out;
2130
2131         /* Retrieve SEV CPUID information */
2132         cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
2133
2134         /* Set encryption bit location for SEV-ES guests */
2135         sev_enc_bit = ebx & 0x3f;
2136
2137         /* Maximum number of encrypted guests supported simultaneously */
2138         max_sev_asid = ecx;
2139         if (!max_sev_asid)
2140                 goto out;
2141
2142         /* Minimum ASID value that should be used for SEV guest */
2143         min_sev_asid = edx;
2144         sev_me_mask = 1UL << (ebx & 0x3f);
2145
2146         /*
2147          * Initialize SEV ASID bitmaps. Allocate space for ASID 0 in the bitmap,
2148          * even though it's never used, so that the bitmap is indexed by the
2149          * actual ASID.
2150          */
2151         nr_asids = max_sev_asid + 1;
2152         sev_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2153         if (!sev_asid_bitmap)
2154                 goto out;
2155
2156         sev_reclaim_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2157         if (!sev_reclaim_asid_bitmap) {
2158                 bitmap_free(sev_asid_bitmap);
2159                 sev_asid_bitmap = NULL;
2160                 goto out;
2161         }
2162
2163         sev_asid_count = max_sev_asid - min_sev_asid + 1;
2164         if (misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count))
2165                 goto out;
2166
2167         pr_info("SEV supported: %u ASIDs\n", sev_asid_count);
2168         sev_supported = true;
2169
2170         /* SEV-ES support requested? */
2171         if (!sev_es_enabled)
2172                 goto out;
2173
2174         /* Does the CPU support SEV-ES? */
2175         if (!boot_cpu_has(X86_FEATURE_SEV_ES))
2176                 goto out;
2177
2178         /* Has the system been allocated ASIDs for SEV-ES? */
2179         if (min_sev_asid == 1)
2180                 goto out;
2181
2182         sev_es_asid_count = min_sev_asid - 1;
2183         if (misc_cg_set_capacity(MISC_CG_RES_SEV_ES, sev_es_asid_count))
2184                 goto out;
2185
2186         pr_info("SEV-ES supported: %u ASIDs\n", sev_es_asid_count);
2187         sev_es_supported = true;
2188
2189 out:
2190         sev_enabled = sev_supported;
2191         sev_es_enabled = sev_es_supported;
2192 #endif
2193 }
2194
2195 void sev_hardware_unsetup(void)
2196 {
2197         if (!sev_enabled)
2198                 return;
2199
2200         /* No need to take sev_bitmap_lock, all VMs have been destroyed. */
2201         sev_flush_asids(1, max_sev_asid);
2202
2203         bitmap_free(sev_asid_bitmap);
2204         bitmap_free(sev_reclaim_asid_bitmap);
2205
2206         misc_cg_set_capacity(MISC_CG_RES_SEV, 0);
2207         misc_cg_set_capacity(MISC_CG_RES_SEV_ES, 0);
2208 }
2209
2210 int sev_cpu_init(struct svm_cpu_data *sd)
2211 {
2212         if (!sev_enabled)
2213                 return 0;
2214
2215         sd->sev_vmcbs = kcalloc(nr_asids, sizeof(void *), GFP_KERNEL);
2216         if (!sd->sev_vmcbs)
2217                 return -ENOMEM;
2218
2219         return 0;
2220 }
2221
2222 /*
2223  * Pages used by hardware to hold guest encrypted state must be flushed before
2224  * returning them to the system.
2225  */
2226 static void sev_flush_guest_memory(struct vcpu_svm *svm, void *va,
2227                                    unsigned long len)
2228 {
2229         /*
2230          * If hardware enforced cache coherency for encrypted mappings of the
2231          * same physical page is supported, nothing to do.
2232          */
2233         if (boot_cpu_has(X86_FEATURE_SME_COHERENT))
2234                 return;
2235
2236         /*
2237          * If the VM Page Flush MSR is supported, use it to flush the page
2238          * (using the page virtual address and the guest ASID).
2239          */
2240         if (boot_cpu_has(X86_FEATURE_VM_PAGE_FLUSH)) {
2241                 struct kvm_sev_info *sev;
2242                 unsigned long va_start;
2243                 u64 start, stop;
2244
2245                 /* Align start and stop to page boundaries. */
2246                 va_start = (unsigned long)va;
2247                 start = (u64)va_start & PAGE_MASK;
2248                 stop = PAGE_ALIGN((u64)va_start + len);
2249
2250                 if (start < stop) {
2251                         sev = &to_kvm_svm(svm->vcpu.kvm)->sev_info;
2252
2253                         while (start < stop) {
2254                                 wrmsrl(MSR_AMD64_VM_PAGE_FLUSH,
2255                                        start | sev->asid);
2256
2257                                 start += PAGE_SIZE;
2258                         }
2259
2260                         return;
2261                 }
2262
2263                 WARN(1, "Address overflow, using WBINVD\n");
2264         }
2265
2266         /*
2267          * Hardware should always have one of the above features,
2268          * but if not, use WBINVD and issue a warning.
2269          */
2270         WARN_ONCE(1, "Using WBINVD to flush guest memory\n");
2271         wbinvd_on_all_cpus();
2272 }
2273
2274 void sev_free_vcpu(struct kvm_vcpu *vcpu)
2275 {
2276         struct vcpu_svm *svm;
2277
2278         if (!sev_es_guest(vcpu->kvm))
2279                 return;
2280
2281         svm = to_svm(vcpu);
2282
2283         if (vcpu->arch.guest_state_protected)
2284                 sev_flush_guest_memory(svm, svm->sev_es.vmsa, PAGE_SIZE);
2285         __free_page(virt_to_page(svm->sev_es.vmsa));
2286
2287         if (svm->sev_es.ghcb_sa_free)
2288                 kvfree(svm->sev_es.ghcb_sa);
2289 }
2290
2291 static void dump_ghcb(struct vcpu_svm *svm)
2292 {
2293         struct ghcb *ghcb = svm->sev_es.ghcb;
2294         unsigned int nbits;
2295
2296         /* Re-use the dump_invalid_vmcb module parameter */
2297         if (!dump_invalid_vmcb) {
2298                 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
2299                 return;
2300         }
2301
2302         nbits = sizeof(ghcb->save.valid_bitmap) * 8;
2303
2304         pr_err("GHCB (GPA=%016llx):\n", svm->vmcb->control.ghcb_gpa);
2305         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
2306                ghcb->save.sw_exit_code, ghcb_sw_exit_code_is_valid(ghcb));
2307         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
2308                ghcb->save.sw_exit_info_1, ghcb_sw_exit_info_1_is_valid(ghcb));
2309         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
2310                ghcb->save.sw_exit_info_2, ghcb_sw_exit_info_2_is_valid(ghcb));
2311         pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
2312                ghcb->save.sw_scratch, ghcb_sw_scratch_is_valid(ghcb));
2313         pr_err("%-20s%*pb\n", "valid_bitmap", nbits, ghcb->save.valid_bitmap);
2314 }
2315
2316 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
2317 {
2318         struct kvm_vcpu *vcpu = &svm->vcpu;
2319         struct ghcb *ghcb = svm->sev_es.ghcb;
2320
2321         /*
2322          * The GHCB protocol so far allows for the following data
2323          * to be returned:
2324          *   GPRs RAX, RBX, RCX, RDX
2325          *
2326          * Copy their values, even if they may not have been written during the
2327          * VM-Exit.  It's the guest's responsibility to not consume random data.
2328          */
2329         ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
2330         ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
2331         ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
2332         ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
2333 }
2334
2335 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
2336 {
2337         struct vmcb_control_area *control = &svm->vmcb->control;
2338         struct kvm_vcpu *vcpu = &svm->vcpu;
2339         struct ghcb *ghcb = svm->sev_es.ghcb;
2340         u64 exit_code;
2341
2342         /*
2343          * The GHCB protocol so far allows for the following data
2344          * to be supplied:
2345          *   GPRs RAX, RBX, RCX, RDX
2346          *   XCR0
2347          *   CPL
2348          *
2349          * VMMCALL allows the guest to provide extra registers. KVM also
2350          * expects RSI for hypercalls, so include that, too.
2351          *
2352          * Copy their values to the appropriate location if supplied.
2353          */
2354         memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
2355
2356         vcpu->arch.regs[VCPU_REGS_RAX] = ghcb_get_rax_if_valid(ghcb);
2357         vcpu->arch.regs[VCPU_REGS_RBX] = ghcb_get_rbx_if_valid(ghcb);
2358         vcpu->arch.regs[VCPU_REGS_RCX] = ghcb_get_rcx_if_valid(ghcb);
2359         vcpu->arch.regs[VCPU_REGS_RDX] = ghcb_get_rdx_if_valid(ghcb);
2360         vcpu->arch.regs[VCPU_REGS_RSI] = ghcb_get_rsi_if_valid(ghcb);
2361
2362         svm->vmcb->save.cpl = ghcb_get_cpl_if_valid(ghcb);
2363
2364         if (ghcb_xcr0_is_valid(ghcb)) {
2365                 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
2366                 kvm_update_cpuid_runtime(vcpu);
2367         }
2368
2369         /* Copy the GHCB exit information into the VMCB fields */
2370         exit_code = ghcb_get_sw_exit_code(ghcb);
2371         control->exit_code = lower_32_bits(exit_code);
2372         control->exit_code_hi = upper_32_bits(exit_code);
2373         control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
2374         control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
2375
2376         /* Clear the valid entries fields */
2377         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
2378 }
2379
2380 static bool sev_es_validate_vmgexit(struct vcpu_svm *svm)
2381 {
2382         struct kvm_vcpu *vcpu;
2383         struct ghcb *ghcb;
2384         u64 exit_code;
2385         u64 reason;
2386
2387         ghcb = svm->sev_es.ghcb;
2388
2389         /*
2390          * Retrieve the exit code now even though it may not be marked valid
2391          * as it could help with debugging.
2392          */
2393         exit_code = ghcb_get_sw_exit_code(ghcb);
2394
2395         /* Only GHCB Usage code 0 is supported */
2396         if (ghcb->ghcb_usage) {
2397                 reason = GHCB_ERR_INVALID_USAGE;
2398                 goto vmgexit_err;
2399         }
2400
2401         reason = GHCB_ERR_MISSING_INPUT;
2402
2403         if (!ghcb_sw_exit_code_is_valid(ghcb) ||
2404             !ghcb_sw_exit_info_1_is_valid(ghcb) ||
2405             !ghcb_sw_exit_info_2_is_valid(ghcb))
2406                 goto vmgexit_err;
2407
2408         switch (ghcb_get_sw_exit_code(ghcb)) {
2409         case SVM_EXIT_READ_DR7:
2410                 break;
2411         case SVM_EXIT_WRITE_DR7:
2412                 if (!ghcb_rax_is_valid(ghcb))
2413                         goto vmgexit_err;
2414                 break;
2415         case SVM_EXIT_RDTSC:
2416                 break;
2417         case SVM_EXIT_RDPMC:
2418                 if (!ghcb_rcx_is_valid(ghcb))
2419                         goto vmgexit_err;
2420                 break;
2421         case SVM_EXIT_CPUID:
2422                 if (!ghcb_rax_is_valid(ghcb) ||
2423                     !ghcb_rcx_is_valid(ghcb))
2424                         goto vmgexit_err;
2425                 if (ghcb_get_rax(ghcb) == 0xd)
2426                         if (!ghcb_xcr0_is_valid(ghcb))
2427                                 goto vmgexit_err;
2428                 break;
2429         case SVM_EXIT_INVD:
2430                 break;
2431         case SVM_EXIT_IOIO:
2432                 if (ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_STR_MASK) {
2433                         if (!ghcb_sw_scratch_is_valid(ghcb))
2434                                 goto vmgexit_err;
2435                 } else {
2436                         if (!(ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_TYPE_MASK))
2437                                 if (!ghcb_rax_is_valid(ghcb))
2438                                         goto vmgexit_err;
2439                 }
2440                 break;
2441         case SVM_EXIT_MSR:
2442                 if (!ghcb_rcx_is_valid(ghcb))
2443                         goto vmgexit_err;
2444                 if (ghcb_get_sw_exit_info_1(ghcb)) {
2445                         if (!ghcb_rax_is_valid(ghcb) ||
2446                             !ghcb_rdx_is_valid(ghcb))
2447                                 goto vmgexit_err;
2448                 }
2449                 break;
2450         case SVM_EXIT_VMMCALL:
2451                 if (!ghcb_rax_is_valid(ghcb) ||
2452                     !ghcb_cpl_is_valid(ghcb))
2453                         goto vmgexit_err;
2454                 break;
2455         case SVM_EXIT_RDTSCP:
2456                 break;
2457         case SVM_EXIT_WBINVD:
2458                 break;
2459         case SVM_EXIT_MONITOR:
2460                 if (!ghcb_rax_is_valid(ghcb) ||
2461                     !ghcb_rcx_is_valid(ghcb) ||
2462                     !ghcb_rdx_is_valid(ghcb))
2463                         goto vmgexit_err;
2464                 break;
2465         case SVM_EXIT_MWAIT:
2466                 if (!ghcb_rax_is_valid(ghcb) ||
2467                     !ghcb_rcx_is_valid(ghcb))
2468                         goto vmgexit_err;
2469                 break;
2470         case SVM_VMGEXIT_MMIO_READ:
2471         case SVM_VMGEXIT_MMIO_WRITE:
2472                 if (!ghcb_sw_scratch_is_valid(ghcb))
2473                         goto vmgexit_err;
2474                 break;
2475         case SVM_VMGEXIT_NMI_COMPLETE:
2476         case SVM_VMGEXIT_AP_HLT_LOOP:
2477         case SVM_VMGEXIT_AP_JUMP_TABLE:
2478         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2479                 break;
2480         default:
2481                 reason = GHCB_ERR_INVALID_EVENT;
2482                 goto vmgexit_err;
2483         }
2484
2485         return true;
2486
2487 vmgexit_err:
2488         vcpu = &svm->vcpu;
2489
2490         if (reason == GHCB_ERR_INVALID_USAGE) {
2491                 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
2492                             ghcb->ghcb_usage);
2493         } else if (reason == GHCB_ERR_INVALID_EVENT) {
2494                 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx is not valid\n",
2495                             exit_code);
2496         } else {
2497                 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx input is not valid\n",
2498                             exit_code);
2499                 dump_ghcb(svm);
2500         }
2501
2502         /* Clear the valid entries fields */
2503         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
2504
2505         ghcb_set_sw_exit_info_1(ghcb, 2);
2506         ghcb_set_sw_exit_info_2(ghcb, reason);
2507
2508         return false;
2509 }
2510
2511 void sev_es_unmap_ghcb(struct vcpu_svm *svm)
2512 {
2513         if (!svm->sev_es.ghcb)
2514                 return;
2515
2516         if (svm->sev_es.ghcb_sa_free) {
2517                 /*
2518                  * The scratch area lives outside the GHCB, so there is a
2519                  * buffer that, depending on the operation performed, may
2520                  * need to be synced, then freed.
2521                  */
2522                 if (svm->sev_es.ghcb_sa_sync) {
2523                         kvm_write_guest(svm->vcpu.kvm,
2524                                         ghcb_get_sw_scratch(svm->sev_es.ghcb),
2525                                         svm->sev_es.ghcb_sa,
2526                                         svm->sev_es.ghcb_sa_len);
2527                         svm->sev_es.ghcb_sa_sync = false;
2528                 }
2529
2530                 kvfree(svm->sev_es.ghcb_sa);
2531                 svm->sev_es.ghcb_sa = NULL;
2532                 svm->sev_es.ghcb_sa_free = false;
2533         }
2534
2535         trace_kvm_vmgexit_exit(svm->vcpu.vcpu_id, svm->sev_es.ghcb);
2536
2537         sev_es_sync_to_ghcb(svm);
2538
2539         kvm_vcpu_unmap(&svm->vcpu, &svm->sev_es.ghcb_map, true);
2540         svm->sev_es.ghcb = NULL;
2541 }
2542
2543 void pre_sev_run(struct vcpu_svm *svm, int cpu)
2544 {
2545         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
2546         int asid = sev_get_asid(svm->vcpu.kvm);
2547
2548         /* Assign the asid allocated with this SEV guest */
2549         svm->asid = asid;
2550
2551         /*
2552          * Flush guest TLB:
2553          *
2554          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
2555          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
2556          */
2557         if (sd->sev_vmcbs[asid] == svm->vmcb &&
2558             svm->vcpu.arch.last_vmentry_cpu == cpu)
2559                 return;
2560
2561         sd->sev_vmcbs[asid] = svm->vmcb;
2562         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
2563         vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
2564 }
2565
2566 #define GHCB_SCRATCH_AREA_LIMIT         (16ULL * PAGE_SIZE)
2567 static bool setup_vmgexit_scratch(struct vcpu_svm *svm, bool sync, u64 len)
2568 {
2569         struct vmcb_control_area *control = &svm->vmcb->control;
2570         struct ghcb *ghcb = svm->sev_es.ghcb;
2571         u64 ghcb_scratch_beg, ghcb_scratch_end;
2572         u64 scratch_gpa_beg, scratch_gpa_end;
2573         void *scratch_va;
2574
2575         scratch_gpa_beg = ghcb_get_sw_scratch(ghcb);
2576         if (!scratch_gpa_beg) {
2577                 pr_err("vmgexit: scratch gpa not provided\n");
2578                 goto e_scratch;
2579         }
2580
2581         scratch_gpa_end = scratch_gpa_beg + len;
2582         if (scratch_gpa_end < scratch_gpa_beg) {
2583                 pr_err("vmgexit: scratch length (%#llx) not valid for scratch address (%#llx)\n",
2584                        len, scratch_gpa_beg);
2585                 goto e_scratch;
2586         }
2587
2588         if ((scratch_gpa_beg & PAGE_MASK) == control->ghcb_gpa) {
2589                 /* Scratch area begins within GHCB */
2590                 ghcb_scratch_beg = control->ghcb_gpa +
2591                                    offsetof(struct ghcb, shared_buffer);
2592                 ghcb_scratch_end = control->ghcb_gpa +
2593                                    offsetof(struct ghcb, reserved_1);
2594
2595                 /*
2596                  * If the scratch area begins within the GHCB, it must be
2597                  * completely contained in the GHCB shared buffer area.
2598                  */
2599                 if (scratch_gpa_beg < ghcb_scratch_beg ||
2600                     scratch_gpa_end > ghcb_scratch_end) {
2601                         pr_err("vmgexit: scratch area is outside of GHCB shared buffer area (%#llx - %#llx)\n",
2602                                scratch_gpa_beg, scratch_gpa_end);
2603                         goto e_scratch;
2604                 }
2605
2606                 scratch_va = (void *)svm->sev_es.ghcb;
2607                 scratch_va += (scratch_gpa_beg - control->ghcb_gpa);
2608         } else {
2609                 /*
2610                  * The guest memory must be read into a kernel buffer, so
2611                  * limit the size
2612                  */
2613                 if (len > GHCB_SCRATCH_AREA_LIMIT) {
2614                         pr_err("vmgexit: scratch area exceeds KVM limits (%#llx requested, %#llx limit)\n",
2615                                len, GHCB_SCRATCH_AREA_LIMIT);
2616                         goto e_scratch;
2617                 }
2618                 scratch_va = kvzalloc(len, GFP_KERNEL_ACCOUNT);
2619                 if (!scratch_va)
2620                         goto e_scratch;
2621
2622                 if (kvm_read_guest(svm->vcpu.kvm, scratch_gpa_beg, scratch_va, len)) {
2623                         /* Unable to copy scratch area from guest */
2624                         pr_err("vmgexit: kvm_read_guest for scratch area failed\n");
2625
2626                         kvfree(scratch_va);
2627                         goto e_scratch;
2628                 }
2629
2630                 /*
2631                  * The scratch area is outside the GHCB. The operation will
2632                  * dictate whether the buffer needs to be synced before running
2633                  * the vCPU next time (i.e. a read was requested so the data
2634                  * must be written back to the guest memory).
2635                  */
2636                 svm->sev_es.ghcb_sa_sync = sync;
2637                 svm->sev_es.ghcb_sa_free = true;
2638         }
2639
2640         svm->sev_es.ghcb_sa = scratch_va;
2641         svm->sev_es.ghcb_sa_len = len;
2642
2643         return true;
2644
2645 e_scratch:
2646         ghcb_set_sw_exit_info_1(ghcb, 2);
2647         ghcb_set_sw_exit_info_2(ghcb, GHCB_ERR_INVALID_SCRATCH_AREA);
2648
2649         return false;
2650 }
2651
2652 static void set_ghcb_msr_bits(struct vcpu_svm *svm, u64 value, u64 mask,
2653                               unsigned int pos)
2654 {
2655         svm->vmcb->control.ghcb_gpa &= ~(mask << pos);
2656         svm->vmcb->control.ghcb_gpa |= (value & mask) << pos;
2657 }
2658
2659 static u64 get_ghcb_msr_bits(struct vcpu_svm *svm, u64 mask, unsigned int pos)
2660 {
2661         return (svm->vmcb->control.ghcb_gpa >> pos) & mask;
2662 }
2663
2664 static void set_ghcb_msr(struct vcpu_svm *svm, u64 value)
2665 {
2666         svm->vmcb->control.ghcb_gpa = value;
2667 }
2668
2669 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
2670 {
2671         struct vmcb_control_area *control = &svm->vmcb->control;
2672         struct kvm_vcpu *vcpu = &svm->vcpu;
2673         u64 ghcb_info;
2674         int ret = 1;
2675
2676         ghcb_info = control->ghcb_gpa & GHCB_MSR_INFO_MASK;
2677
2678         trace_kvm_vmgexit_msr_protocol_enter(svm->vcpu.vcpu_id,
2679                                              control->ghcb_gpa);
2680
2681         switch (ghcb_info) {
2682         case GHCB_MSR_SEV_INFO_REQ:
2683                 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2684                                                     GHCB_VERSION_MIN,
2685                                                     sev_enc_bit));
2686                 break;
2687         case GHCB_MSR_CPUID_REQ: {
2688                 u64 cpuid_fn, cpuid_reg, cpuid_value;
2689
2690                 cpuid_fn = get_ghcb_msr_bits(svm,
2691                                              GHCB_MSR_CPUID_FUNC_MASK,
2692                                              GHCB_MSR_CPUID_FUNC_POS);
2693
2694                 /* Initialize the registers needed by the CPUID intercept */
2695                 vcpu->arch.regs[VCPU_REGS_RAX] = cpuid_fn;
2696                 vcpu->arch.regs[VCPU_REGS_RCX] = 0;
2697
2698                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_CPUID);
2699                 if (!ret) {
2700                         /* Error, keep GHCB MSR value as-is */
2701                         break;
2702                 }
2703
2704                 cpuid_reg = get_ghcb_msr_bits(svm,
2705                                               GHCB_MSR_CPUID_REG_MASK,
2706                                               GHCB_MSR_CPUID_REG_POS);
2707                 if (cpuid_reg == 0)
2708                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RAX];
2709                 else if (cpuid_reg == 1)
2710                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RBX];
2711                 else if (cpuid_reg == 2)
2712                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RCX];
2713                 else
2714                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RDX];
2715
2716                 set_ghcb_msr_bits(svm, cpuid_value,
2717                                   GHCB_MSR_CPUID_VALUE_MASK,
2718                                   GHCB_MSR_CPUID_VALUE_POS);
2719
2720                 set_ghcb_msr_bits(svm, GHCB_MSR_CPUID_RESP,
2721                                   GHCB_MSR_INFO_MASK,
2722                                   GHCB_MSR_INFO_POS);
2723                 break;
2724         }
2725         case GHCB_MSR_TERM_REQ: {
2726                 u64 reason_set, reason_code;
2727
2728                 reason_set = get_ghcb_msr_bits(svm,
2729                                                GHCB_MSR_TERM_REASON_SET_MASK,
2730                                                GHCB_MSR_TERM_REASON_SET_POS);
2731                 reason_code = get_ghcb_msr_bits(svm,
2732                                                 GHCB_MSR_TERM_REASON_MASK,
2733                                                 GHCB_MSR_TERM_REASON_POS);
2734                 pr_info("SEV-ES guest requested termination: %#llx:%#llx\n",
2735                         reason_set, reason_code);
2736
2737                 ret = -EINVAL;
2738                 break;
2739         }
2740         default:
2741                 /* Error, keep GHCB MSR value as-is */
2742                 break;
2743         }
2744
2745         trace_kvm_vmgexit_msr_protocol_exit(svm->vcpu.vcpu_id,
2746                                             control->ghcb_gpa, ret);
2747
2748         return ret;
2749 }
2750
2751 int sev_handle_vmgexit(struct kvm_vcpu *vcpu)
2752 {
2753         struct vcpu_svm *svm = to_svm(vcpu);
2754         struct vmcb_control_area *control = &svm->vmcb->control;
2755         u64 ghcb_gpa, exit_code;
2756         struct ghcb *ghcb;
2757         int ret;
2758
2759         /* Validate the GHCB */
2760         ghcb_gpa = control->ghcb_gpa;
2761         if (ghcb_gpa & GHCB_MSR_INFO_MASK)
2762                 return sev_handle_vmgexit_msr_protocol(svm);
2763
2764         if (!ghcb_gpa) {
2765                 vcpu_unimpl(vcpu, "vmgexit: GHCB gpa is not set\n");
2766
2767                 /* Without a GHCB, just return right back to the guest */
2768                 return 1;
2769         }
2770
2771         if (kvm_vcpu_map(vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->sev_es.ghcb_map)) {
2772                 /* Unable to map GHCB from guest */
2773                 vcpu_unimpl(vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
2774                             ghcb_gpa);
2775
2776                 /* Without a GHCB, just return right back to the guest */
2777                 return 1;
2778         }
2779
2780         svm->sev_es.ghcb = svm->sev_es.ghcb_map.hva;
2781         ghcb = svm->sev_es.ghcb_map.hva;
2782
2783         trace_kvm_vmgexit_enter(vcpu->vcpu_id, ghcb);
2784
2785         exit_code = ghcb_get_sw_exit_code(ghcb);
2786
2787         if (!sev_es_validate_vmgexit(svm))
2788                 return 1;
2789
2790         sev_es_sync_from_ghcb(svm);
2791         ghcb_set_sw_exit_info_1(ghcb, 0);
2792         ghcb_set_sw_exit_info_2(ghcb, 0);
2793
2794         ret = 1;
2795         switch (exit_code) {
2796         case SVM_VMGEXIT_MMIO_READ:
2797                 if (!setup_vmgexit_scratch(svm, true, control->exit_info_2))
2798                         break;
2799
2800                 ret = kvm_sev_es_mmio_read(vcpu,
2801                                            control->exit_info_1,
2802                                            control->exit_info_2,
2803                                            svm->sev_es.ghcb_sa);
2804                 break;
2805         case SVM_VMGEXIT_MMIO_WRITE:
2806                 if (!setup_vmgexit_scratch(svm, false, control->exit_info_2))
2807                         break;
2808
2809                 ret = kvm_sev_es_mmio_write(vcpu,
2810                                             control->exit_info_1,
2811                                             control->exit_info_2,
2812                                             svm->sev_es.ghcb_sa);
2813                 break;
2814         case SVM_VMGEXIT_NMI_COMPLETE:
2815                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_IRET);
2816                 break;
2817         case SVM_VMGEXIT_AP_HLT_LOOP:
2818                 ret = kvm_emulate_ap_reset_hold(vcpu);
2819                 break;
2820         case SVM_VMGEXIT_AP_JUMP_TABLE: {
2821                 struct kvm_sev_info *sev = &to_kvm_svm(vcpu->kvm)->sev_info;
2822
2823                 switch (control->exit_info_1) {
2824                 case 0:
2825                         /* Set AP jump table address */
2826                         sev->ap_jump_table = control->exit_info_2;
2827                         break;
2828                 case 1:
2829                         /* Get AP jump table address */
2830                         ghcb_set_sw_exit_info_2(ghcb, sev->ap_jump_table);
2831                         break;
2832                 default:
2833                         pr_err("svm: vmgexit: unsupported AP jump table request - exit_info_1=%#llx\n",
2834                                control->exit_info_1);
2835                         ghcb_set_sw_exit_info_1(ghcb, 2);
2836                         ghcb_set_sw_exit_info_2(ghcb, GHCB_ERR_INVALID_INPUT);
2837                 }
2838
2839                 break;
2840         }
2841         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2842                 vcpu_unimpl(vcpu,
2843                             "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
2844                             control->exit_info_1, control->exit_info_2);
2845                 ret = -EINVAL;
2846                 break;
2847         default:
2848                 ret = svm_invoke_exit_handler(vcpu, exit_code);
2849         }
2850
2851         return ret;
2852 }
2853
2854 int sev_es_string_io(struct vcpu_svm *svm, int size, unsigned int port, int in)
2855 {
2856         int count;
2857         int bytes;
2858
2859         if (svm->vmcb->control.exit_info_2 > INT_MAX)
2860                 return -EINVAL;
2861
2862         count = svm->vmcb->control.exit_info_2;
2863         if (unlikely(check_mul_overflow(count, size, &bytes)))
2864                 return -EINVAL;
2865
2866         if (!setup_vmgexit_scratch(svm, in, bytes))
2867                 return 1;
2868
2869         return kvm_sev_es_string_io(&svm->vcpu, size, port, svm->sev_es.ghcb_sa,
2870                                     count, in);
2871 }
2872
2873 void sev_es_init_vmcb(struct vcpu_svm *svm)
2874 {
2875         struct kvm_vcpu *vcpu = &svm->vcpu;
2876
2877         svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ES_ENABLE;
2878         svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
2879
2880         /*
2881          * An SEV-ES guest requires a VMSA area that is a separate from the
2882          * VMCB page. Do not include the encryption mask on the VMSA physical
2883          * address since hardware will access it using the guest key.
2884          */
2885         svm->vmcb->control.vmsa_pa = __pa(svm->sev_es.vmsa);
2886
2887         /* Can't intercept CR register access, HV can't modify CR registers */
2888         svm_clr_intercept(svm, INTERCEPT_CR0_READ);
2889         svm_clr_intercept(svm, INTERCEPT_CR4_READ);
2890         svm_clr_intercept(svm, INTERCEPT_CR8_READ);
2891         svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
2892         svm_clr_intercept(svm, INTERCEPT_CR4_WRITE);
2893         svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
2894
2895         svm_clr_intercept(svm, INTERCEPT_SELECTIVE_CR0);
2896
2897         /* Track EFER/CR register changes */
2898         svm_set_intercept(svm, TRAP_EFER_WRITE);
2899         svm_set_intercept(svm, TRAP_CR0_WRITE);
2900         svm_set_intercept(svm, TRAP_CR4_WRITE);
2901         svm_set_intercept(svm, TRAP_CR8_WRITE);
2902
2903         /* No support for enable_vmware_backdoor */
2904         clr_exception_intercept(svm, GP_VECTOR);
2905
2906         /* Can't intercept XSETBV, HV can't modify XCR0 directly */
2907         svm_clr_intercept(svm, INTERCEPT_XSETBV);
2908
2909         /* Clear intercepts on selected MSRs */
2910         set_msr_interception(vcpu, svm->msrpm, MSR_EFER, 1, 1);
2911         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_CR_PAT, 1, 1);
2912         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
2913         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
2914         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
2915         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
2916 }
2917
2918 void sev_es_vcpu_reset(struct vcpu_svm *svm)
2919 {
2920         /*
2921          * Set the GHCB MSR value as per the GHCB specification when emulating
2922          * vCPU RESET for an SEV-ES guest.
2923          */
2924         set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2925                                             GHCB_VERSION_MIN,
2926                                             sev_enc_bit));
2927 }
2928
2929 void sev_es_prepare_switch_to_guest(struct vmcb_save_area *hostsa)
2930 {
2931         /*
2932          * As an SEV-ES guest, hardware will restore the host state on VMEXIT,
2933          * of which one step is to perform a VMLOAD.  KVM performs the
2934          * corresponding VMSAVE in svm_prepare_guest_switch for both
2935          * traditional and SEV-ES guests.
2936          */
2937
2938         /* XCR0 is restored on VMEXIT, save the current host value */
2939         hostsa->xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
2940
2941         /* PKRU is restored on VMEXIT, save the current host value */
2942         hostsa->pkru = read_pkru();
2943
2944         /* MSR_IA32_XSS is restored on VMEXIT, save the currnet host value */
2945         hostsa->xss = host_xss;
2946 }
2947
2948 void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
2949 {
2950         struct vcpu_svm *svm = to_svm(vcpu);
2951
2952         /* First SIPI: Use the values as initially set by the VMM */
2953         if (!svm->sev_es.received_first_sipi) {
2954                 svm->sev_es.received_first_sipi = true;
2955                 return;
2956         }
2957
2958         /*
2959          * Subsequent SIPI: Return from an AP Reset Hold VMGEXIT, where
2960          * the guest will set the CS and RIP. Set SW_EXIT_INFO_2 to a
2961          * non-zero value.
2962          */
2963         if (!svm->sev_es.ghcb)
2964                 return;
2965
2966         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, 1);
2967 }