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/processor.h>
  18 #include <linux/trace_events.h>
  19 #include <asm/fpu/internal.h>
  20
  21 #include <asm/trapnr.h>
  22
  23 #include "x86.h"
  24 #include "svm.h"
  25 #include "cpuid.h"
  26 #include "trace.h"
  27
  28 #define __ex(x) __kvm_handle_fault_on_reboot(x)
  29
  30 static u8 sev_enc_bit;
  31 static int sev_flush_asids(void);
  32 static DECLARE_RWSEM(sev_deactivate_lock);
  33 static DEFINE_MUTEX(sev_bitmap_lock);
  34 unsigned int max_sev_asid;
  35 static unsigned int min_sev_asid;
  36 static unsigned long *sev_asid_bitmap;
  37 static unsigned long *sev_reclaim_asid_bitmap;
  38
  39 struct enc_region {
  40         struct list_head list;
  41         unsigned long npages;
  42         struct page **pages;
  43         unsigned long uaddr;
  44         unsigned long size;
  45 };
  46
  47 static int sev_flush_asids(void)
  48 {
  49         int ret, error = 0;
  50
  51         /*
  52          * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
  53          * so it must be guarded.
  54          */
  55         down_write(&sev_deactivate_lock);
  56
  57         wbinvd_on_all_cpus();
  58         ret = sev_guest_df_flush(&error);
  59
  60         up_write(&sev_deactivate_lock);
  61
  62         if (ret)
  63                 pr_err("SEV: DF_FLUSH failed, ret=%d, error=%#x\n", ret, error);
  64
  65         return ret;
  66 }
  67
  68 /* Must be called with the sev_bitmap_lock held */
  69 static bool __sev_recycle_asids(int min_asid, int max_asid)
  70 {
  71         int pos;
  72
  73         /* Check if there are any ASIDs to reclaim before performing a flush */
  74         pos = find_next_bit(sev_reclaim_asid_bitmap, max_sev_asid, min_asid);
  75         if (pos >= max_asid)
  76                 return false;
  77
  78         if (sev_flush_asids())
  79                 return false;
  80
  81         /* The flush process will flush all reclaimable SEV and SEV-ES ASIDs */
  82         bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
  83                    max_sev_asid);
  84         bitmap_zero(sev_reclaim_asid_bitmap, max_sev_asid);
  85
  86         return true;
  87 }
  88
  89 static int sev_asid_new(struct kvm_sev_info *sev)
  90 {
  91         int pos, min_asid, max_asid;
  92         bool retry = true;
  93
  94         mutex_lock(&sev_bitmap_lock);
  95
  96         /*
  97          * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
  98          * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.
  99          */
 100         min_asid = sev->es_active ? 0 : min_sev_asid - 1;
 101         max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
 102 again:
 103         pos = find_next_zero_bit(sev_asid_bitmap, max_sev_asid, min_asid);
 104         if (pos >= max_asid) {
 105                 if (retry && __sev_recycle_asids(min_asid, max_asid)) {
 106                         retry = false;
 107                         goto again;
 108                 }
 109                 mutex_unlock(&sev_bitmap_lock);
 110                 return -EBUSY;
 111         }
 112
 113         __set_bit(pos, sev_asid_bitmap);
 114
 115         mutex_unlock(&sev_bitmap_lock);
 116
 117         return pos + 1;
 118 }
 119
 120 static int sev_get_asid(struct kvm *kvm)
 121 {
 122         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 123
 124         return sev->asid;
 125 }
 126
 127 static void sev_asid_free(int asid)
 128 {
 129         struct svm_cpu_data *sd;
 130         int cpu, pos;
 131
 132         mutex_lock(&sev_bitmap_lock);
 133
 134         pos = asid - 1;
 135         __set_bit(pos, sev_reclaim_asid_bitmap);
 136
 137         for_each_possible_cpu(cpu) {
 138                 sd = per_cpu(svm_data, cpu);
 139                 sd->sev_vmcbs[pos] = NULL;
 140         }
 141
 142         mutex_unlock(&sev_bitmap_lock);
 143 }
 144
 145 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
 146 {
 147         struct sev_data_decommission *decommission;
 148         struct sev_data_deactivate *data;
 149
 150         if (!handle)
 151                 return;
 152
 153         data = kzalloc(sizeof(*data), GFP_KERNEL);
 154         if (!data)
 155                 return;
 156
 157         /* deactivate handle */
 158         data->handle = handle;
 159
 160         /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
 161         down_read(&sev_deactivate_lock);
 162         sev_guest_deactivate(data, NULL);
 163         up_read(&sev_deactivate_lock);
 164
 165         kfree(data);
 166
 167         decommission = kzalloc(sizeof(*decommission), GFP_KERNEL);
 168         if (!decommission)
 169                 return;
 170
 171         /* decommission handle */
 172         decommission->handle = handle;
 173         sev_guest_decommission(decommission, NULL);
 174
 175         kfree(decommission);
 176 }
 177
 178 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 179 {
 180         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 181         int asid, ret;
 182
 183         ret = -EBUSY;
 184         if (unlikely(sev->active))
 185                 return ret;
 186
 187         asid = sev_asid_new(sev);
 188         if (asid < 0)
 189                 return ret;
 190
 191         ret = sev_platform_init(&argp->error);
 192         if (ret)
 193                 goto e_free;
 194
 195         sev->active = true;
 196         sev->asid = asid;
 197         INIT_LIST_HEAD(&sev->regions_list);
 198
 199         return 0;
 200
 201 e_free:
 202         sev_asid_free(asid);
 203         return ret;
 204 }
 205
 206 static int sev_es_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 207 {
 208         if (!sev_es)
 209                 return -ENOTTY;
 210
 211         to_kvm_svm(kvm)->sev_info.es_active = true;
 212
 213         return sev_guest_init(kvm, argp);
 214 }
 215
 216 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
 217 {
 218         struct sev_data_activate *data;
 219         int asid = sev_get_asid(kvm);
 220         int ret;
 221
 222         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 223         if (!data)
 224                 return -ENOMEM;
 225
 226         /* activate ASID on the given handle */
 227         data->handle = handle;
 228         data->asid   = asid;
 229         ret = sev_guest_activate(data, error);
 230         kfree(data);
 231
 232         return ret;
 233 }
 234
 235 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
 236 {
 237         struct fd f;
 238         int ret;
 239
 240         f = fdget(fd);
 241         if (!f.file)
 242                 return -EBADF;
 243
 244         ret = sev_issue_cmd_external_user(f.file, id, data, error);
 245
 246         fdput(f);
 247         return ret;
 248 }
 249
 250 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
 251 {
 252         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 253
 254         return __sev_issue_cmd(sev->fd, id, data, error);
 255 }
 256
 257 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
 258 {
 259         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 260         struct sev_data_launch_start *start;
 261         struct kvm_sev_launch_start params;
 262         void *dh_blob, *session_blob;
 263         int *error = &argp->error;
 264         int ret;
 265
 266         if (!sev_guest(kvm))
 267                 return -ENOTTY;
 268
 269         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 270                 return -EFAULT;
 271
 272         start = kzalloc(sizeof(*start), GFP_KERNEL_ACCOUNT);
 273         if (!start)
 274                 return -ENOMEM;
 275
 276         dh_blob = NULL;
 277         if (params.dh_uaddr) {
 278                 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
 279                 if (IS_ERR(dh_blob)) {
 280                         ret = PTR_ERR(dh_blob);
 281                         goto e_free;
 282                 }
 283
 284                 start->dh_cert_address = __sme_set(__pa(dh_blob));
 285                 start->dh_cert_len = params.dh_len;
 286         }
 287
 288         session_blob = NULL;
 289         if (params.session_uaddr) {
 290                 session_blob = psp_copy_user_blob(params.session_uaddr, params.session_len);
 291                 if (IS_ERR(session_blob)) {
 292                         ret = PTR_ERR(session_blob);
 293                         goto e_free_dh;
 294                 }
 295
 296                 start->session_address = __sme_set(__pa(session_blob));
 297                 start->session_len = params.session_len;
 298         }
 299
 300         start->handle = params.handle;
 301         start->policy = params.policy;
 302
 303         /* create memory encryption context */
 304         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_LAUNCH_START, start, error);
 305         if (ret)
 306                 goto e_free_session;
 307
 308         /* Bind ASID to this guest */
 309         ret = sev_bind_asid(kvm, start->handle, error);
 310         if (ret)
 311                 goto e_free_session;
 312
 313         /* return handle to userspace */
 314         params.handle = start->handle;
 315         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params))) {
 316                 sev_unbind_asid(kvm, start->handle);
 317                 ret = -EFAULT;
 318                 goto e_free_session;
 319         }
 320
 321         sev->handle = start->handle;
 322         sev->fd = argp->sev_fd;
 323
 324 e_free_session:
 325         kfree(session_blob);
 326 e_free_dh:
 327         kfree(dh_blob);
 328 e_free:
 329         kfree(start);
 330         return ret;
 331 }
 332
 333 static struct page **sev_pin_memory(struct kvm *kvm, unsigned long uaddr,
 334                                     unsigned long ulen, unsigned long *n,
 335                                     int write)
 336 {
 337         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 338         unsigned long npages, size;
 339         int npinned;
 340         unsigned long locked, lock_limit;
 341         struct page **pages;
 342         unsigned long first, last;
 343         int ret;
 344
 345         if (ulen == 0 || uaddr + ulen < uaddr)
 346                 return ERR_PTR(-EINVAL);
 347
 348         /* Calculate number of pages. */
 349         first = (uaddr & PAGE_MASK) >> PAGE_SHIFT;
 350         last = ((uaddr + ulen - 1) & PAGE_MASK) >> PAGE_SHIFT;
 351         npages = (last - first + 1);
 352
 353         locked = sev->pages_locked + npages;
 354         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 355         if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
 356                 pr_err("SEV: %lu locked pages exceed the lock limit of %lu.\n", locked, lock_limit);
 357                 return ERR_PTR(-ENOMEM);
 358         }
 359
 360         if (WARN_ON_ONCE(npages > INT_MAX))
 361                 return ERR_PTR(-EINVAL);
 362
 363         /* Avoid using vmalloc for smaller buffers. */
 364         size = npages * sizeof(struct page *);
 365         if (size > PAGE_SIZE)
 366                 pages = __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 367         else
 368                 pages = kmalloc(size, GFP_KERNEL_ACCOUNT);
 369
 370         if (!pages)
 371                 return ERR_PTR(-ENOMEM);
 372
 373         /* Pin the user virtual address. */
 374         npinned = pin_user_pages_fast(uaddr, npages, write ? FOLL_WRITE : 0, pages);
 375         if (npinned != npages) {
 376                 pr_err("SEV: Failure locking %lu pages.\n", npages);
 377                 ret = -ENOMEM;
 378                 goto err;
 379         }
 380
 381         *n = npages;
 382         sev->pages_locked = locked;
 383
 384         return pages;
 385
 386 err:
 387         if (npinned > 0)
 388                 unpin_user_pages(pages, npinned);
 389
 390         kvfree(pages);
 391         return ERR_PTR(ret);
 392 }
 393
 394 static void sev_unpin_memory(struct kvm *kvm, struct page **pages,
 395                              unsigned long npages)
 396 {
 397         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 398
 399         unpin_user_pages(pages, npages);
 400         kvfree(pages);
 401         sev->pages_locked -= npages;
 402 }
 403
 404 static void sev_clflush_pages(struct page *pages[], unsigned long npages)
 405 {
 406         uint8_t *page_virtual;
 407         unsigned long i;
 408
 409         if (this_cpu_has(X86_FEATURE_SME_COHERENT) || npages == 0 ||
 410             pages == NULL)
 411                 return;
 412
 413         for (i = 0; i < npages; i++) {
 414                 page_virtual = kmap_atomic(pages[i]);
 415                 clflush_cache_range(page_virtual, PAGE_SIZE);
 416                 kunmap_atomic(page_virtual);
 417         }
 418 }
 419
 420 static unsigned long get_num_contig_pages(unsigned long idx,
 421                                 struct page **inpages, unsigned long npages)
 422 {
 423         unsigned long paddr, next_paddr;
 424         unsigned long i = idx + 1, pages = 1;
 425
 426         /* find the number of contiguous pages starting from idx */
 427         paddr = __sme_page_pa(inpages[idx]);
 428         while (i < npages) {
 429                 next_paddr = __sme_page_pa(inpages[i++]);
 430                 if ((paddr + PAGE_SIZE) == next_paddr) {
 431                         pages++;
 432                         paddr = next_paddr;
 433                         continue;
 434                 }
 435                 break;
 436         }
 437
 438         return pages;
 439 }
 440
 441 static int sev_launch_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
 442 {
 443         unsigned long vaddr, vaddr_end, next_vaddr, npages, pages, size, i;
 444         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 445         struct kvm_sev_launch_update_data params;
 446         struct sev_data_launch_update_data *data;
 447         struct page **inpages;
 448         int ret;
 449
 450         if (!sev_guest(kvm))
 451                 return -ENOTTY;
 452
 453         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 454                 return -EFAULT;
 455
 456         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 457         if (!data)
 458                 return -ENOMEM;
 459
 460         vaddr = params.uaddr;
 461         size = params.len;
 462         vaddr_end = vaddr + size;
 463
 464         /* Lock the user memory. */
 465         inpages = sev_pin_memory(kvm, vaddr, size, &npages, 1);
 466         if (IS_ERR(inpages)) {
 467                 ret = PTR_ERR(inpages);
 468                 goto e_free;
 469         }
 470
 471         /*
 472          * Flush (on non-coherent CPUs) before LAUNCH_UPDATE encrypts pages in
 473          * place; the cache may contain the data that was written unencrypted.
 474          */
 475         sev_clflush_pages(inpages, npages);
 476
 477         for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
 478                 int offset, len;
 479
 480                 /*
 481                  * If the user buffer is not page-aligned, calculate the offset
 482                  * within the page.
 483                  */
 484                 offset = vaddr & (PAGE_SIZE - 1);
 485
 486                 /* Calculate the number of pages that can be encrypted in one go. */
 487                 pages = get_num_contig_pages(i, inpages, npages);
 488
 489                 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
 490
 491                 data->handle = sev->handle;
 492                 data->len = len;
 493                 data->address = __sme_page_pa(inpages[i]) + offset;
 494                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, data, &argp->error);
 495                 if (ret)
 496                         goto e_unpin;
 497
 498                 size -= len;
 499                 next_vaddr = vaddr + len;
 500         }
 501
 502 e_unpin:
 503         /* content of memory is updated, mark pages dirty */
 504         for (i = 0; i < npages; i++) {
 505                 set_page_dirty_lock(inpages[i]);
 506                 mark_page_accessed(inpages[i]);
 507         }
 508         /* unlock the user pages */
 509         sev_unpin_memory(kvm, inpages, npages);
 510 e_free:
 511         kfree(data);
 512         return ret;
 513 }
 514
 515 static int sev_es_sync_vmsa(struct vcpu_svm *svm)
 516 {
 517         struct vmcb_save_area *save = &svm->vmcb->save;
 518
 519         /* Check some debug related fields before encrypting the VMSA */
 520         if (svm->vcpu.guest_debug || (save->dr7 & ~DR7_FIXED_1))
 521                 return -EINVAL;
 522
 523         /* Sync registgers */
 524         save->rax = svm->vcpu.arch.regs[VCPU_REGS_RAX];
 525         save->rbx = svm->vcpu.arch.regs[VCPU_REGS_RBX];
 526         save->rcx = svm->vcpu.arch.regs[VCPU_REGS_RCX];
 527         save->rdx = svm->vcpu.arch.regs[VCPU_REGS_RDX];
 528         save->rsp = svm->vcpu.arch.regs[VCPU_REGS_RSP];
 529         save->rbp = svm->vcpu.arch.regs[VCPU_REGS_RBP];
 530         save->rsi = svm->vcpu.arch.regs[VCPU_REGS_RSI];
 531         save->rdi = svm->vcpu.arch.regs[VCPU_REGS_RDI];
 532 #ifdef CONFIG_X86_64
 533         save->r8  = svm->vcpu.arch.regs[VCPU_REGS_R8];
 534         save->r9  = svm->vcpu.arch.regs[VCPU_REGS_R9];
 535         save->r10 = svm->vcpu.arch.regs[VCPU_REGS_R10];
 536         save->r11 = svm->vcpu.arch.regs[VCPU_REGS_R11];
 537         save->r12 = svm->vcpu.arch.regs[VCPU_REGS_R12];
 538         save->r13 = svm->vcpu.arch.regs[VCPU_REGS_R13];
 539         save->r14 = svm->vcpu.arch.regs[VCPU_REGS_R14];
 540         save->r15 = svm->vcpu.arch.regs[VCPU_REGS_R15];
 541 #endif
 542         save->rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
 543
 544         /* Sync some non-GPR registers before encrypting */
 545         save->xcr0 = svm->vcpu.arch.xcr0;
 546         save->pkru = svm->vcpu.arch.pkru;
 547         save->xss  = svm->vcpu.arch.ia32_xss;
 548
 549         /*
 550          * SEV-ES will use a VMSA that is pointed to by the VMCB, not
 551          * the traditional VMSA that is part of the VMCB. Copy the
 552          * traditional VMSA as it has been built so far (in prep
 553          * for LAUNCH_UPDATE_VMSA) to be the initial SEV-ES state.
 554          */
 555         memcpy(svm->vmsa, save, sizeof(*save));
 556
 557         return 0;
 558 }
 559
 560 static int sev_launch_update_vmsa(struct kvm *kvm, struct kvm_sev_cmd *argp)
 561 {
 562         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 563         struct sev_data_launch_update_vmsa *vmsa;
 564         int i, ret;
 565
 566         if (!sev_es_guest(kvm))
 567                 return -ENOTTY;
 568
 569         vmsa = kzalloc(sizeof(*vmsa), GFP_KERNEL);
 570         if (!vmsa)
 571                 return -ENOMEM;
 572
 573         for (i = 0; i < kvm->created_vcpus; i++) {
 574                 struct vcpu_svm *svm = to_svm(kvm->vcpus[i]);
 575
 576                 /* Perform some pre-encryption checks against the VMSA */
 577                 ret = sev_es_sync_vmsa(svm);
 578                 if (ret)
 579                         goto e_free;
 580
 581                 /*
 582                  * The LAUNCH_UPDATE_VMSA command will perform in-place
 583                  * encryption of the VMSA memory content (i.e it will write
 584                  * the same memory region with the guest's key), so invalidate
 585                  * it first.
 586                  */
 587                 clflush_cache_range(svm->vmsa, PAGE_SIZE);
 588
 589                 vmsa->handle = sev->handle;
 590                 vmsa->address = __sme_pa(svm->vmsa);
 591                 vmsa->len = PAGE_SIZE;
 592                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_VMSA, vmsa,
 593                                     &argp->error);
 594                 if (ret)
 595                         goto e_free;
 596
 597                 svm->vcpu.arch.guest_state_protected = true;
 598         }
 599
 600 e_free:
 601         kfree(vmsa);
 602         return ret;
 603 }
 604
 605 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 606 {
 607         void __user *measure = (void __user *)(uintptr_t)argp->data;
 608         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 609         struct sev_data_launch_measure *data;
 610         struct kvm_sev_launch_measure params;
 611         void __user *p = NULL;
 612         void *blob = NULL;
 613         int ret;
 614
 615         if (!sev_guest(kvm))
 616                 return -ENOTTY;
 617
 618         if (copy_from_user(&params, measure, sizeof(params)))
 619                 return -EFAULT;
 620
 621         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 622         if (!data)
 623                 return -ENOMEM;
 624
 625         /* User wants to query the blob length */
 626         if (!params.len)
 627                 goto cmd;
 628
 629         p = (void __user *)(uintptr_t)params.uaddr;
 630         if (p) {
 631                 if (params.len > SEV_FW_BLOB_MAX_SIZE) {
 632                         ret = -EINVAL;
 633                         goto e_free;
 634                 }
 635
 636                 ret = -ENOMEM;
 637                 blob = kmalloc(params.len, GFP_KERNEL);
 638                 if (!blob)
 639                         goto e_free;
 640
 641                 data->address = __psp_pa(blob);
 642                 data->len = params.len;
 643         }
 644
 645 cmd:
 646         data->handle = sev->handle;
 647         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, data, &argp->error);
 648
 649         /*
 650          * If we query the session length, FW responded with expected data.
 651          */
 652         if (!params.len)
 653                 goto done;
 654
 655         if (ret)
 656                 goto e_free_blob;
 657
 658         if (blob) {
 659                 if (copy_to_user(p, blob, params.len))
 660                         ret = -EFAULT;
 661         }
 662
 663 done:
 664         params.len = data->len;
 665         if (copy_to_user(measure, &params, sizeof(params)))
 666                 ret = -EFAULT;
 667 e_free_blob:
 668         kfree(blob);
 669 e_free:
 670         kfree(data);
 671         return ret;
 672 }
 673
 674 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 675 {
 676         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 677         struct sev_data_launch_finish *data;
 678         int ret;
 679
 680         if (!sev_guest(kvm))
 681                 return -ENOTTY;
 682
 683         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 684         if (!data)
 685                 return -ENOMEM;
 686
 687         data->handle = sev->handle;
 688         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, data, &argp->error);
 689
 690         kfree(data);
 691         return ret;
 692 }
 693
 694 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 695 {
 696         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 697         struct kvm_sev_guest_status params;
 698         struct sev_data_guest_status *data;
 699         int ret;
 700
 701         if (!sev_guest(kvm))
 702                 return -ENOTTY;
 703
 704         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 705         if (!data)
 706                 return -ENOMEM;
 707
 708         data->handle = sev->handle;
 709         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, data, &argp->error);
 710         if (ret)
 711                 goto e_free;
 712
 713         params.policy = data->policy;
 714         params.state = data->state;
 715         params.handle = data->handle;
 716
 717         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 718                 ret = -EFAULT;
 719 e_free:
 720         kfree(data);
 721         return ret;
 722 }
 723
 724 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 725                                unsigned long dst, int size,
 726                                int *error, bool enc)
 727 {
 728         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 729         struct sev_data_dbg *data;
 730         int ret;
 731
 732         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 733         if (!data)
 734                 return -ENOMEM;
 735
 736         data->handle = sev->handle;
 737         data->dst_addr = dst;
 738         data->src_addr = src;
 739         data->len = size;
 740
 741         ret = sev_issue_cmd(kvm,
 742                             enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 743                             data, error);
 744         kfree(data);
 745         return ret;
 746 }
 747
 748 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 749                              unsigned long dst_paddr, int sz, int *err)
 750 {
 751         int offset;
 752
 753         /*
 754          * Its safe to read more than we are asked, caller should ensure that
 755          * destination has enough space.
 756          */
 757         offset = src_paddr & 15;
 758         src_paddr = round_down(src_paddr, 16);
 759         sz = round_up(sz + offset, 16);
 760
 761         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 762 }
 763
 764 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 765                                   unsigned long __user dst_uaddr,
 766                                   unsigned long dst_paddr,
 767                                   int size, int *err)
 768 {
 769         struct page *tpage = NULL;
 770         int ret, offset;
 771
 772         /* if inputs are not 16-byte then use intermediate buffer */
 773         if (!IS_ALIGNED(dst_paddr, 16) ||
 774             !IS_ALIGNED(paddr,     16) ||
 775             !IS_ALIGNED(size,      16)) {
 776                 tpage = (void *)alloc_page(GFP_KERNEL);
 777                 if (!tpage)
 778                         return -ENOMEM;
 779
 780                 dst_paddr = __sme_page_pa(tpage);
 781         }
 782
 783         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 784         if (ret)
 785                 goto e_free;
 786
 787         if (tpage) {
 788                 offset = paddr & 15;
 789                 if (copy_to_user((void __user *)(uintptr_t)dst_uaddr,
 790                                  page_address(tpage) + offset, size))
 791                         ret = -EFAULT;
 792         }
 793
 794 e_free:
 795         if (tpage)
 796                 __free_page(tpage);
 797
 798         return ret;
 799 }
 800
 801 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 802                                   unsigned long __user vaddr,
 803                                   unsigned long dst_paddr,
 804                                   unsigned long __user dst_vaddr,
 805                                   int size, int *error)
 806 {
 807         struct page *src_tpage = NULL;
 808         struct page *dst_tpage = NULL;
 809         int ret, len = size;
 810
 811         /* If source buffer is not aligned then use an intermediate buffer */
 812         if (!IS_ALIGNED(vaddr, 16)) {
 813                 src_tpage = alloc_page(GFP_KERNEL);
 814                 if (!src_tpage)
 815                         return -ENOMEM;
 816
 817                 if (copy_from_user(page_address(src_tpage),
 818                                 (void __user *)(uintptr_t)vaddr, size)) {
 819                         __free_page(src_tpage);
 820                         return -EFAULT;
 821                 }
 822
 823                 paddr = __sme_page_pa(src_tpage);
 824         }
 825
 826         /*
 827          *  If destination buffer or length is not aligned then do read-modify-write:
 828          *   - decrypt destination in an intermediate buffer
 829          *   - copy the source buffer in an intermediate buffer
 830          *   - use the intermediate buffer as source buffer
 831          */
 832         if (!IS_ALIGNED(dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 833                 int dst_offset;
 834
 835                 dst_tpage = alloc_page(GFP_KERNEL);
 836                 if (!dst_tpage) {
 837                         ret = -ENOMEM;
 838                         goto e_free;
 839                 }
 840
 841                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 842                                         __sme_page_pa(dst_tpage), size, error);
 843                 if (ret)
 844                         goto e_free;
 845
 846                 /*
 847                  *  If source is kernel buffer then use memcpy() otherwise
 848                  *  copy_from_user().
 849                  */
 850                 dst_offset = dst_paddr & 15;
 851
 852                 if (src_tpage)
 853                         memcpy(page_address(dst_tpage) + dst_offset,
 854                                page_address(src_tpage), size);
 855                 else {
 856                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 857                                            (void __user *)(uintptr_t)vaddr, size)) {
 858                                 ret = -EFAULT;
 859                                 goto e_free;
 860                         }
 861                 }
 862
 863                 paddr = __sme_page_pa(dst_tpage);
 864                 dst_paddr = round_down(dst_paddr, 16);
 865                 len = round_up(size, 16);
 866         }
 867
 868         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 869
 870 e_free:
 871         if (src_tpage)
 872                 __free_page(src_tpage);
 873         if (dst_tpage)
 874                 __free_page(dst_tpage);
 875         return ret;
 876 }
 877
 878 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 879 {
 880         unsigned long vaddr, vaddr_end, next_vaddr;
 881         unsigned long dst_vaddr;
 882         struct page **src_p, **dst_p;
 883         struct kvm_sev_dbg debug;
 884         unsigned long n;
 885         unsigned int size;
 886         int ret;
 887
 888         if (!sev_guest(kvm))
 889                 return -ENOTTY;
 890
 891         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 892                 return -EFAULT;
 893
 894         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 895                 return -EINVAL;
 896         if (!debug.dst_uaddr)
 897                 return -EINVAL;
 898
 899         vaddr = debug.src_uaddr;
 900         size = debug.len;
 901         vaddr_end = vaddr + size;
 902         dst_vaddr = debug.dst_uaddr;
 903
 904         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 905                 int len, s_off, d_off;
 906
 907                 /* lock userspace source and destination page */
 908                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 909                 if (IS_ERR(src_p))
 910                         return PTR_ERR(src_p);
 911
 912                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 913                 if (IS_ERR(dst_p)) {
 914                         sev_unpin_memory(kvm, src_p, n);
 915                         return PTR_ERR(dst_p);
 916                 }
 917
 918                 /*
 919                  * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
 920                  * the pages; flush the destination too so that future accesses do not
 921                  * see stale data.
 922                  */
 923                 sev_clflush_pages(src_p, 1);
 924                 sev_clflush_pages(dst_p, 1);
 925
 926                 /*
 927                  * Since user buffer may not be page aligned, calculate the
 928                  * offset within the page.
 929                  */
 930                 s_off = vaddr & ~PAGE_MASK;
 931                 d_off = dst_vaddr & ~PAGE_MASK;
 932                 len = min_t(size_t, (PAGE_SIZE - s_off), size);
 933
 934                 if (dec)
 935                         ret = __sev_dbg_decrypt_user(kvm,
 936                                                      __sme_page_pa(src_p[0]) + s_off,
 937                                                      dst_vaddr,
 938                                                      __sme_page_pa(dst_p[0]) + d_off,
 939                                                      len, &argp->error);
 940                 else
 941                         ret = __sev_dbg_encrypt_user(kvm,
 942                                                      __sme_page_pa(src_p[0]) + s_off,
 943                                                      vaddr,
 944                                                      __sme_page_pa(dst_p[0]) + d_off,
 945                                                      dst_vaddr,
 946                                                      len, &argp->error);
 947
 948                 sev_unpin_memory(kvm, src_p, n);
 949                 sev_unpin_memory(kvm, dst_p, n);
 950
 951                 if (ret)
 952                         goto err;
 953
 954                 next_vaddr = vaddr + len;
 955                 dst_vaddr = dst_vaddr + len;
 956                 size -= len;
 957         }
 958 err:
 959         return ret;
 960 }
 961
 962 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
 963 {
 964         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 965         struct sev_data_launch_secret *data;
 966         struct kvm_sev_launch_secret params;
 967         struct page **pages;
 968         void *blob, *hdr;
 969         unsigned long n, i;
 970         int ret, offset;
 971
 972         if (!sev_guest(kvm))
 973                 return -ENOTTY;
 974
 975         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 976                 return -EFAULT;
 977
 978         pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, 1);
 979         if (IS_ERR(pages))
 980                 return PTR_ERR(pages);
 981
 982         /*
 983          * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
 984          * place; the cache may contain the data that was written unencrypted.
 985          */
 986         sev_clflush_pages(pages, n);
 987
 988         /*
 989          * The secret must be copied into contiguous memory region, lets verify
 990          * that userspace memory pages are contiguous before we issue command.
 991          */
 992         if (get_num_contig_pages(0, pages, n) != n) {
 993                 ret = -EINVAL;
 994                 goto e_unpin_memory;
 995         }
 996
 997         ret = -ENOMEM;
 998         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 999         if (!data)
1000                 goto e_unpin_memory;
1001
1002         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1003         data->guest_address = __sme_page_pa(pages[0]) + offset;
1004         data->guest_len = params.guest_len;
1005
1006         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1007         if (IS_ERR(blob)) {
1008                 ret = PTR_ERR(blob);
1009                 goto e_free;
1010         }
1011
1012         data->trans_address = __psp_pa(blob);
1013         data->trans_len = params.trans_len;
1014
1015         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1016         if (IS_ERR(hdr)) {
1017                 ret = PTR_ERR(hdr);
1018                 goto e_free_blob;
1019         }
1020         data->hdr_address = __psp_pa(hdr);
1021         data->hdr_len = params.hdr_len;
1022
1023         data->handle = sev->handle;
1024         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, data, &argp->error);
1025
1026         kfree(hdr);
1027
1028 e_free_blob:
1029         kfree(blob);
1030 e_free:
1031         kfree(data);
1032 e_unpin_memory:
1033         /* content of memory is updated, mark pages dirty */
1034         for (i = 0; i < n; i++) {
1035                 set_page_dirty_lock(pages[i]);
1036                 mark_page_accessed(pages[i]);
1037         }
1038         sev_unpin_memory(kvm, pages, n);
1039         return ret;
1040 }
1041
1042 int svm_mem_enc_op(struct kvm *kvm, void __user *argp)
1043 {
1044         struct kvm_sev_cmd sev_cmd;
1045         int r;
1046
1047         if (!svm_sev_enabled() || !sev)
1048                 return -ENOTTY;
1049
1050         if (!argp)
1051                 return 0;
1052
1053         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
1054                 return -EFAULT;
1055
1056         mutex_lock(&kvm->lock);
1057
1058         switch (sev_cmd.id) {
1059         case KVM_SEV_INIT:
1060                 r = sev_guest_init(kvm, &sev_cmd);
1061                 break;
1062         case KVM_SEV_ES_INIT:
1063                 r = sev_es_guest_init(kvm, &sev_cmd);
1064                 break;
1065         case KVM_SEV_LAUNCH_START:
1066                 r = sev_launch_start(kvm, &sev_cmd);
1067                 break;
1068         case KVM_SEV_LAUNCH_UPDATE_DATA:
1069                 r = sev_launch_update_data(kvm, &sev_cmd);
1070                 break;
1071         case KVM_SEV_LAUNCH_UPDATE_VMSA:
1072                 r = sev_launch_update_vmsa(kvm, &sev_cmd);
1073                 break;
1074         case KVM_SEV_LAUNCH_MEASURE:
1075                 r = sev_launch_measure(kvm, &sev_cmd);
1076                 break;
1077         case KVM_SEV_LAUNCH_FINISH:
1078                 r = sev_launch_finish(kvm, &sev_cmd);
1079                 break;
1080         case KVM_SEV_GUEST_STATUS:
1081                 r = sev_guest_status(kvm, &sev_cmd);
1082                 break;
1083         case KVM_SEV_DBG_DECRYPT:
1084                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
1085                 break;
1086         case KVM_SEV_DBG_ENCRYPT:
1087                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
1088                 break;
1089         case KVM_SEV_LAUNCH_SECRET:
1090                 r = sev_launch_secret(kvm, &sev_cmd);
1091                 break;
1092         default:
1093                 r = -EINVAL;
1094                 goto out;
1095         }
1096
1097         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
1098                 r = -EFAULT;
1099
1100 out:
1101         mutex_unlock(&kvm->lock);
1102         return r;
1103 }
1104
1105 int svm_register_enc_region(struct kvm *kvm,
1106                             struct kvm_enc_region *range)
1107 {
1108         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1109         struct enc_region *region;
1110         int ret = 0;
1111
1112         if (!sev_guest(kvm))
1113                 return -ENOTTY;
1114
1115         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
1116                 return -EINVAL;
1117
1118         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
1119         if (!region)
1120                 return -ENOMEM;
1121
1122         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
1123         if (IS_ERR(region->pages)) {
1124                 ret = PTR_ERR(region->pages);
1125                 goto e_free;
1126         }
1127
1128         /*
1129          * The guest may change the memory encryption attribute from C=0 -> C=1
1130          * or vice versa for this memory range. Lets make sure caches are
1131          * flushed to ensure that guest data gets written into memory with
1132          * correct C-bit.
1133          */
1134         sev_clflush_pages(region->pages, region->npages);
1135
1136         region->uaddr = range->addr;
1137         region->size = range->size;
1138
1139         mutex_lock(&kvm->lock);
1140         list_add_tail(&region->list, &sev->regions_list);
1141         mutex_unlock(&kvm->lock);
1142
1143         return ret;
1144
1145 e_free:
1146         kfree(region);
1147         return ret;
1148 }
1149
1150 static struct enc_region *
1151 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
1152 {
1153         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1154         struct list_head *head = &sev->regions_list;
1155         struct enc_region *i;
1156
1157         list_for_each_entry(i, head, list) {
1158                 if (i->uaddr == range->addr &&
1159                     i->size == range->size)
1160                         return i;
1161         }
1162
1163         return NULL;
1164 }
1165
1166 static void __unregister_enc_region_locked(struct kvm *kvm,
1167                                            struct enc_region *region)
1168 {
1169         sev_unpin_memory(kvm, region->pages, region->npages);
1170         list_del(&region->list);
1171         kfree(region);
1172 }
1173
1174 int svm_unregister_enc_region(struct kvm *kvm,
1175                               struct kvm_enc_region *range)
1176 {
1177         struct enc_region *region;
1178         int ret;
1179
1180         mutex_lock(&kvm->lock);
1181
1182         if (!sev_guest(kvm)) {
1183                 ret = -ENOTTY;
1184                 goto failed;
1185         }
1186
1187         region = find_enc_region(kvm, range);
1188         if (!region) {
1189                 ret = -EINVAL;
1190                 goto failed;
1191         }
1192
1193         /*
1194          * Ensure that all guest tagged cache entries are flushed before
1195          * releasing the pages back to the system for use. CLFLUSH will
1196          * not do this, so issue a WBINVD.
1197          */
1198         wbinvd_on_all_cpus();
1199
1200         __unregister_enc_region_locked(kvm, region);
1201
1202         mutex_unlock(&kvm->lock);
1203         return 0;
1204
1205 failed:
1206         mutex_unlock(&kvm->lock);
1207         return ret;
1208 }
1209
1210 void sev_vm_destroy(struct kvm *kvm)
1211 {
1212         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1213         struct list_head *head = &sev->regions_list;
1214         struct list_head *pos, *q;
1215
1216         if (!sev_guest(kvm))
1217                 return;
1218
1219         mutex_lock(&kvm->lock);
1220
1221         /*
1222          * Ensure that all guest tagged cache entries are flushed before
1223          * releasing the pages back to the system for use. CLFLUSH will
1224          * not do this, so issue a WBINVD.
1225          */
1226         wbinvd_on_all_cpus();
1227
1228         /*
1229          * if userspace was terminated before unregistering the memory regions
1230          * then lets unpin all the registered memory.
1231          */
1232         if (!list_empty(head)) {
1233                 list_for_each_safe(pos, q, head) {
1234                         __unregister_enc_region_locked(kvm,
1235                                 list_entry(pos, struct enc_region, list));
1236                         cond_resched();
1237                 }
1238         }
1239
1240         mutex_unlock(&kvm->lock);
1241
1242         sev_unbind_asid(kvm, sev->handle);
1243         sev_asid_free(sev->asid);
1244 }
1245
1246 void __init sev_hardware_setup(void)
1247 {
1248         unsigned int eax, ebx, ecx, edx;
1249         bool sev_es_supported = false;
1250         bool sev_supported = false;
1251
1252         /* Does the CPU support SEV? */
1253         if (!boot_cpu_has(X86_FEATURE_SEV))
1254                 goto out;
1255
1256         /* Retrieve SEV CPUID information */
1257         cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
1258
1259         /* Set encryption bit location for SEV-ES guests */
1260         sev_enc_bit = ebx & 0x3f;
1261
1262         /* Maximum number of encrypted guests supported simultaneously */
1263         max_sev_asid = ecx;
1264
1265         if (!svm_sev_enabled())
1266                 goto out;
1267
1268         /* Minimum ASID value that should be used for SEV guest */
1269         min_sev_asid = edx;
1270
1271         /* Initialize SEV ASID bitmaps */
1272         sev_asid_bitmap = bitmap_zalloc(max_sev_asid, GFP_KERNEL);
1273         if (!sev_asid_bitmap)
1274                 goto out;
1275
1276         sev_reclaim_asid_bitmap = bitmap_zalloc(max_sev_asid, GFP_KERNEL);
1277         if (!sev_reclaim_asid_bitmap)
1278                 goto out;
1279
1280         pr_info("SEV supported: %u ASIDs\n", max_sev_asid - min_sev_asid + 1);
1281         sev_supported = true;
1282
1283         /* SEV-ES support requested? */
1284         if (!sev_es)
1285                 goto out;
1286
1287         /* Does the CPU support SEV-ES? */
1288         if (!boot_cpu_has(X86_FEATURE_SEV_ES))
1289                 goto out;
1290
1291         /* Has the system been allocated ASIDs for SEV-ES? */
1292         if (min_sev_asid == 1)
1293                 goto out;
1294
1295         pr_info("SEV-ES supported: %u ASIDs\n", min_sev_asid - 1);
1296         sev_es_supported = true;
1297
1298 out:
1299         sev = sev_supported;
1300         sev_es = sev_es_supported;
1301 }
1302
1303 void sev_hardware_teardown(void)
1304 {
1305         if (!svm_sev_enabled())
1306                 return;
1307
1308         bitmap_free(sev_asid_bitmap);
1309         bitmap_free(sev_reclaim_asid_bitmap);
1310
1311         sev_flush_asids();
1312 }
1313
1314 /*
1315  * Pages used by hardware to hold guest encrypted state must be flushed before
1316  * returning them to the system.
1317  */
1318 static void sev_flush_guest_memory(struct vcpu_svm *svm, void *va,
1319                                    unsigned long len)
1320 {
1321         /*
1322          * If hardware enforced cache coherency for encrypted mappings of the
1323          * same physical page is supported, nothing to do.
1324          */
1325         if (boot_cpu_has(X86_FEATURE_SME_COHERENT))
1326                 return;
1327
1328         /*
1329          * If the VM Page Flush MSR is supported, use it to flush the page
1330          * (using the page virtual address and the guest ASID).
1331          */
1332         if (boot_cpu_has(X86_FEATURE_VM_PAGE_FLUSH)) {
1333                 struct kvm_sev_info *sev;
1334                 unsigned long va_start;
1335                 u64 start, stop;
1336
1337                 /* Align start and stop to page boundaries. */
1338                 va_start = (unsigned long)va;
1339                 start = (u64)va_start & PAGE_MASK;
1340                 stop = PAGE_ALIGN((u64)va_start + len);
1341
1342                 if (start < stop) {
1343                         sev = &to_kvm_svm(svm->vcpu.kvm)->sev_info;
1344
1345                         while (start < stop) {
1346                                 wrmsrl(MSR_AMD64_VM_PAGE_FLUSH,
1347                                        start | sev->asid);
1348
1349                                 start += PAGE_SIZE;
1350                         }
1351
1352                         return;
1353                 }
1354
1355                 WARN(1, "Address overflow, using WBINVD\n");
1356         }
1357
1358         /*
1359          * Hardware should always have one of the above features,
1360          * but if not, use WBINVD and issue a warning.
1361          */
1362         WARN_ONCE(1, "Using WBINVD to flush guest memory\n");
1363         wbinvd_on_all_cpus();
1364 }
1365
1366 void sev_free_vcpu(struct kvm_vcpu *vcpu)
1367 {
1368         struct vcpu_svm *svm;
1369
1370         if (!sev_es_guest(vcpu->kvm))
1371                 return;
1372
1373         svm = to_svm(vcpu);
1374
1375         if (vcpu->arch.guest_state_protected)
1376                 sev_flush_guest_memory(svm, svm->vmsa, PAGE_SIZE);
1377         __free_page(virt_to_page(svm->vmsa));
1378
1379         if (svm->ghcb_sa_free)
1380                 kfree(svm->ghcb_sa);
1381 }
1382
1383 static void dump_ghcb(struct vcpu_svm *svm)
1384 {
1385         struct ghcb *ghcb = svm->ghcb;
1386         unsigned int nbits;
1387
1388         /* Re-use the dump_invalid_vmcb module parameter */
1389         if (!dump_invalid_vmcb) {
1390                 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
1391                 return;
1392         }
1393
1394         nbits = sizeof(ghcb->save.valid_bitmap) * 8;
1395
1396         pr_err("GHCB (GPA=%016llx):\n", svm->vmcb->control.ghcb_gpa);
1397         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
1398                ghcb->save.sw_exit_code, ghcb_sw_exit_code_is_valid(ghcb));
1399         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
1400                ghcb->save.sw_exit_info_1, ghcb_sw_exit_info_1_is_valid(ghcb));
1401         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
1402                ghcb->save.sw_exit_info_2, ghcb_sw_exit_info_2_is_valid(ghcb));
1403         pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
1404                ghcb->save.sw_scratch, ghcb_sw_scratch_is_valid(ghcb));
1405         pr_err("%-20s%*pb\n", "valid_bitmap", nbits, ghcb->save.valid_bitmap);
1406 }
1407
1408 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
1409 {
1410         struct kvm_vcpu *vcpu = &svm->vcpu;
1411         struct ghcb *ghcb = svm->ghcb;
1412
1413         /*
1414          * The GHCB protocol so far allows for the following data
1415          * to be returned:
1416          *   GPRs RAX, RBX, RCX, RDX
1417          *
1418          * Copy their values to the GHCB if they are dirty.
1419          */
1420         if (kvm_register_is_dirty(vcpu, VCPU_REGS_RAX))
1421                 ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
1422         if (kvm_register_is_dirty(vcpu, VCPU_REGS_RBX))
1423                 ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
1424         if (kvm_register_is_dirty(vcpu, VCPU_REGS_RCX))
1425                 ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
1426         if (kvm_register_is_dirty(vcpu, VCPU_REGS_RDX))
1427                 ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
1428 }
1429
1430 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
1431 {
1432         struct vmcb_control_area *control = &svm->vmcb->control;
1433         struct kvm_vcpu *vcpu = &svm->vcpu;
1434         struct ghcb *ghcb = svm->ghcb;
1435         u64 exit_code;
1436
1437         /*
1438          * The GHCB protocol so far allows for the following data
1439          * to be supplied:
1440          *   GPRs RAX, RBX, RCX, RDX
1441          *   XCR0
1442          *   CPL
1443          *
1444          * VMMCALL allows the guest to provide extra registers. KVM also
1445          * expects RSI for hypercalls, so include that, too.
1446          *
1447          * Copy their values to the appropriate location if supplied.
1448          */
1449         memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
1450
1451         vcpu->arch.regs[VCPU_REGS_RAX] = ghcb_get_rax_if_valid(ghcb);
1452         vcpu->arch.regs[VCPU_REGS_RBX] = ghcb_get_rbx_if_valid(ghcb);
1453         vcpu->arch.regs[VCPU_REGS_RCX] = ghcb_get_rcx_if_valid(ghcb);
1454         vcpu->arch.regs[VCPU_REGS_RDX] = ghcb_get_rdx_if_valid(ghcb);
1455         vcpu->arch.regs[VCPU_REGS_RSI] = ghcb_get_rsi_if_valid(ghcb);
1456
1457         svm->vmcb->save.cpl = ghcb_get_cpl_if_valid(ghcb);
1458
1459         if (ghcb_xcr0_is_valid(ghcb)) {
1460                 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
1461                 kvm_update_cpuid_runtime(vcpu);
1462         }
1463
1464         /* Copy the GHCB exit information into the VMCB fields */
1465         exit_code = ghcb_get_sw_exit_code(ghcb);
1466         control->exit_code = lower_32_bits(exit_code);
1467         control->exit_code_hi = upper_32_bits(exit_code);
1468         control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
1469         control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
1470
1471         /* Clear the valid entries fields */
1472         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
1473 }
1474
1475 static int sev_es_validate_vmgexit(struct vcpu_svm *svm)
1476 {
1477         struct kvm_vcpu *vcpu;
1478         struct ghcb *ghcb;
1479         u64 exit_code = 0;
1480
1481         ghcb = svm->ghcb;
1482
1483         /* Only GHCB Usage code 0 is supported */
1484         if (ghcb->ghcb_usage)
1485                 goto vmgexit_err;
1486
1487         /*
1488          * Retrieve the exit code now even though is may not be marked valid
1489          * as it could help with debugging.
1490          */
1491         exit_code = ghcb_get_sw_exit_code(ghcb);
1492
1493         if (!ghcb_sw_exit_code_is_valid(ghcb) ||
1494             !ghcb_sw_exit_info_1_is_valid(ghcb) ||
1495             !ghcb_sw_exit_info_2_is_valid(ghcb))
1496                 goto vmgexit_err;
1497
1498         switch (ghcb_get_sw_exit_code(ghcb)) {
1499         case SVM_EXIT_READ_DR7:
1500                 break;
1501         case SVM_EXIT_WRITE_DR7:
1502                 if (!ghcb_rax_is_valid(ghcb))
1503                         goto vmgexit_err;
1504                 break;
1505         case SVM_EXIT_RDTSC:
1506                 break;
1507         case SVM_EXIT_RDPMC:
1508                 if (!ghcb_rcx_is_valid(ghcb))
1509                         goto vmgexit_err;
1510                 break;
1511         case SVM_EXIT_CPUID:
1512                 if (!ghcb_rax_is_valid(ghcb) ||
1513                     !ghcb_rcx_is_valid(ghcb))
1514                         goto vmgexit_err;
1515                 if (ghcb_get_rax(ghcb) == 0xd)
1516                         if (!ghcb_xcr0_is_valid(ghcb))
1517                                 goto vmgexit_err;
1518                 break;
1519         case SVM_EXIT_INVD:
1520                 break;
1521         case SVM_EXIT_IOIO:
1522                 if (ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_STR_MASK) {
1523                         if (!ghcb_sw_scratch_is_valid(ghcb))
1524                                 goto vmgexit_err;
1525                 } else {
1526                         if (!(ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_TYPE_MASK))
1527                                 if (!ghcb_rax_is_valid(ghcb))
1528                                         goto vmgexit_err;
1529                 }
1530                 break;
1531         case SVM_EXIT_MSR:
1532                 if (!ghcb_rcx_is_valid(ghcb))
1533                         goto vmgexit_err;
1534                 if (ghcb_get_sw_exit_info_1(ghcb)) {
1535                         if (!ghcb_rax_is_valid(ghcb) ||
1536                             !ghcb_rdx_is_valid(ghcb))
1537                                 goto vmgexit_err;
1538                 }
1539                 break;
1540         case SVM_EXIT_VMMCALL:
1541                 if (!ghcb_rax_is_valid(ghcb) ||
1542                     !ghcb_cpl_is_valid(ghcb))
1543                         goto vmgexit_err;
1544                 break;
1545         case SVM_EXIT_RDTSCP:
1546                 break;
1547         case SVM_EXIT_WBINVD:
1548                 break;
1549         case SVM_EXIT_MONITOR:
1550                 if (!ghcb_rax_is_valid(ghcb) ||
1551                     !ghcb_rcx_is_valid(ghcb) ||
1552                     !ghcb_rdx_is_valid(ghcb))
1553                         goto vmgexit_err;
1554                 break;
1555         case SVM_EXIT_MWAIT:
1556                 if (!ghcb_rax_is_valid(ghcb) ||
1557                     !ghcb_rcx_is_valid(ghcb))
1558                         goto vmgexit_err;
1559                 break;
1560         case SVM_VMGEXIT_MMIO_READ:
1561         case SVM_VMGEXIT_MMIO_WRITE:
1562                 if (!ghcb_sw_scratch_is_valid(ghcb))
1563                         goto vmgexit_err;
1564                 break;
1565         case SVM_VMGEXIT_NMI_COMPLETE:
1566         case SVM_VMGEXIT_AP_JUMP_TABLE:
1567         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
1568                 break;
1569         default:
1570                 goto vmgexit_err;
1571         }
1572
1573         return 0;
1574
1575 vmgexit_err:
1576         vcpu = &svm->vcpu;
1577
1578         if (ghcb->ghcb_usage) {
1579                 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
1580                             ghcb->ghcb_usage);
1581         } else {
1582                 vcpu_unimpl(vcpu, "vmgexit: exit reason %#llx is not valid\n",
1583                             exit_code);
1584                 dump_ghcb(svm);
1585         }
1586
1587         vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
1588         vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
1589         vcpu->run->internal.ndata = 2;
1590         vcpu->run->internal.data[0] = exit_code;
1591         vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
1592
1593         return -EINVAL;
1594 }
1595
1596 static void pre_sev_es_run(struct vcpu_svm *svm)
1597 {
1598         if (!svm->ghcb)
1599                 return;
1600
1601         if (svm->ghcb_sa_free) {
1602                 /*
1603                  * The scratch area lives outside the GHCB, so there is a
1604                  * buffer that, depending on the operation performed, may
1605                  * need to be synced, then freed.
1606                  */
1607                 if (svm->ghcb_sa_sync) {
1608                         kvm_write_guest(svm->vcpu.kvm,
1609                                         ghcb_get_sw_scratch(svm->ghcb),
1610                                         svm->ghcb_sa, svm->ghcb_sa_len);
1611                         svm->ghcb_sa_sync = false;
1612                 }
1613
1614                 kfree(svm->ghcb_sa);
1615                 svm->ghcb_sa = NULL;
1616                 svm->ghcb_sa_free = false;
1617         }
1618
1619         trace_kvm_vmgexit_exit(svm->vcpu.vcpu_id, svm->ghcb);
1620
1621         sev_es_sync_to_ghcb(svm);
1622
1623         kvm_vcpu_unmap(&svm->vcpu, &svm->ghcb_map, true);
1624         svm->ghcb = NULL;
1625 }
1626
1627 void pre_sev_run(struct vcpu_svm *svm, int cpu)
1628 {
1629         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1630         int asid = sev_get_asid(svm->vcpu.kvm);
1631
1632         /* Perform any SEV-ES pre-run actions */
1633         pre_sev_es_run(svm);
1634
1635         /* Assign the asid allocated with this SEV guest */
1636         svm->asid = asid;
1637
1638         /*
1639          * Flush guest TLB:
1640          *
1641          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
1642          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
1643          */
1644         if (sd->sev_vmcbs[asid] == svm->vmcb &&
1645             svm->vcpu.arch.last_vmentry_cpu == cpu)
1646                 return;
1647
1648         sd->sev_vmcbs[asid] = svm->vmcb;
1649         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
1650         vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1651 }
1652
1653 #define GHCB_SCRATCH_AREA_LIMIT         (16ULL * PAGE_SIZE)
1654 static bool setup_vmgexit_scratch(struct vcpu_svm *svm, bool sync, u64 len)
1655 {
1656         struct vmcb_control_area *control = &svm->vmcb->control;
1657         struct ghcb *ghcb = svm->ghcb;
1658         u64 ghcb_scratch_beg, ghcb_scratch_end;
1659         u64 scratch_gpa_beg, scratch_gpa_end;
1660         void *scratch_va;
1661
1662         scratch_gpa_beg = ghcb_get_sw_scratch(ghcb);
1663         if (!scratch_gpa_beg) {
1664                 pr_err("vmgexit: scratch gpa not provided\n");
1665                 return false;
1666         }
1667
1668         scratch_gpa_end = scratch_gpa_beg + len;
1669         if (scratch_gpa_end < scratch_gpa_beg) {
1670                 pr_err("vmgexit: scratch length (%#llx) not valid for scratch address (%#llx)\n",
1671                        len, scratch_gpa_beg);
1672                 return false;
1673         }
1674
1675         if ((scratch_gpa_beg & PAGE_MASK) == control->ghcb_gpa) {
1676                 /* Scratch area begins within GHCB */
1677                 ghcb_scratch_beg = control->ghcb_gpa +
1678                                    offsetof(struct ghcb, shared_buffer);
1679                 ghcb_scratch_end = control->ghcb_gpa +
1680                                    offsetof(struct ghcb, reserved_1);
1681
1682                 /*
1683                  * If the scratch area begins within the GHCB, it must be
1684                  * completely contained in the GHCB shared buffer area.
1685                  */
1686                 if (scratch_gpa_beg < ghcb_scratch_beg ||
1687                     scratch_gpa_end > ghcb_scratch_end) {
1688                         pr_err("vmgexit: scratch area is outside of GHCB shared buffer area (%#llx - %#llx)\n",
1689                                scratch_gpa_beg, scratch_gpa_end);
1690                         return false;
1691                 }
1692
1693                 scratch_va = (void *)svm->ghcb;
1694                 scratch_va += (scratch_gpa_beg - control->ghcb_gpa);
1695         } else {
1696                 /*
1697                  * The guest memory must be read into a kernel buffer, so
1698                  * limit the size
1699                  */
1700                 if (len > GHCB_SCRATCH_AREA_LIMIT) {
1701                         pr_err("vmgexit: scratch area exceeds KVM limits (%#llx requested, %#llx limit)\n",
1702                                len, GHCB_SCRATCH_AREA_LIMIT);
1703                         return false;
1704                 }
1705                 scratch_va = kzalloc(len, GFP_KERNEL);
1706                 if (!scratch_va)
1707                         return false;
1708
1709                 if (kvm_read_guest(svm->vcpu.kvm, scratch_gpa_beg, scratch_va, len)) {
1710                         /* Unable to copy scratch area from guest */
1711                         pr_err("vmgexit: kvm_read_guest for scratch area failed\n");
1712
1713                         kfree(scratch_va);
1714                         return false;
1715                 }
1716
1717                 /*
1718                  * The scratch area is outside the GHCB. The operation will
1719                  * dictate whether the buffer needs to be synced before running
1720                  * the vCPU next time (i.e. a read was requested so the data
1721                  * must be written back to the guest memory).
1722                  */
1723                 svm->ghcb_sa_sync = sync;
1724                 svm->ghcb_sa_free = true;
1725         }
1726
1727         svm->ghcb_sa = scratch_va;
1728         svm->ghcb_sa_len = len;
1729
1730         return true;
1731 }
1732
1733 static void set_ghcb_msr_bits(struct vcpu_svm *svm, u64 value, u64 mask,
1734                               unsigned int pos)
1735 {
1736         svm->vmcb->control.ghcb_gpa &= ~(mask << pos);
1737         svm->vmcb->control.ghcb_gpa |= (value & mask) << pos;
1738 }
1739
1740 static u64 get_ghcb_msr_bits(struct vcpu_svm *svm, u64 mask, unsigned int pos)
1741 {
1742         return (svm->vmcb->control.ghcb_gpa >> pos) & mask;
1743 }
1744
1745 static void set_ghcb_msr(struct vcpu_svm *svm, u64 value)
1746 {
1747         svm->vmcb->control.ghcb_gpa = value;
1748 }
1749
1750 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
1751 {
1752         struct vmcb_control_area *control = &svm->vmcb->control;
1753         struct kvm_vcpu *vcpu = &svm->vcpu;
1754         u64 ghcb_info;
1755         int ret = 1;
1756
1757         ghcb_info = control->ghcb_gpa & GHCB_MSR_INFO_MASK;
1758
1759         trace_kvm_vmgexit_msr_protocol_enter(svm->vcpu.vcpu_id,
1760                                              control->ghcb_gpa);
1761
1762         switch (ghcb_info) {
1763         case GHCB_MSR_SEV_INFO_REQ:
1764                 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
1765                                                     GHCB_VERSION_MIN,
1766                                                     sev_enc_bit));
1767                 break;
1768         case GHCB_MSR_CPUID_REQ: {
1769                 u64 cpuid_fn, cpuid_reg, cpuid_value;
1770
1771                 cpuid_fn = get_ghcb_msr_bits(svm,
1772                                              GHCB_MSR_CPUID_FUNC_MASK,
1773                                              GHCB_MSR_CPUID_FUNC_POS);
1774
1775                 /* Initialize the registers needed by the CPUID intercept */
1776                 vcpu->arch.regs[VCPU_REGS_RAX] = cpuid_fn;
1777                 vcpu->arch.regs[VCPU_REGS_RCX] = 0;
1778
1779                 ret = svm_invoke_exit_handler(svm, SVM_EXIT_CPUID);
1780                 if (!ret) {
1781                         ret = -EINVAL;
1782                         break;
1783                 }
1784
1785                 cpuid_reg = get_ghcb_msr_bits(svm,
1786                                               GHCB_MSR_CPUID_REG_MASK,
1787                                               GHCB_MSR_CPUID_REG_POS);
1788                 if (cpuid_reg == 0)
1789                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RAX];
1790                 else if (cpuid_reg == 1)
1791                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RBX];
1792                 else if (cpuid_reg == 2)
1793                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RCX];
1794                 else
1795                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RDX];
1796
1797                 set_ghcb_msr_bits(svm, cpuid_value,
1798                                   GHCB_MSR_CPUID_VALUE_MASK,
1799                                   GHCB_MSR_CPUID_VALUE_POS);
1800
1801                 set_ghcb_msr_bits(svm, GHCB_MSR_CPUID_RESP,
1802                                   GHCB_MSR_INFO_MASK,
1803                                   GHCB_MSR_INFO_POS);
1804                 break;
1805         }
1806         case GHCB_MSR_TERM_REQ: {
1807                 u64 reason_set, reason_code;
1808
1809                 reason_set = get_ghcb_msr_bits(svm,
1810                                                GHCB_MSR_TERM_REASON_SET_MASK,
1811                                                GHCB_MSR_TERM_REASON_SET_POS);
1812                 reason_code = get_ghcb_msr_bits(svm,
1813                                                 GHCB_MSR_TERM_REASON_MASK,
1814                                                 GHCB_MSR_TERM_REASON_POS);
1815                 pr_info("SEV-ES guest requested termination: %#llx:%#llx\n",
1816                         reason_set, reason_code);
1817                 fallthrough;
1818         }
1819         default:
1820                 ret = -EINVAL;
1821         }
1822
1823         trace_kvm_vmgexit_msr_protocol_exit(svm->vcpu.vcpu_id,
1824                                             control->ghcb_gpa, ret);
1825
1826         return ret;
1827 }
1828
1829 int sev_handle_vmgexit(struct vcpu_svm *svm)
1830 {
1831         struct vmcb_control_area *control = &svm->vmcb->control;
1832         u64 ghcb_gpa, exit_code;
1833         struct ghcb *ghcb;
1834         int ret;
1835
1836         /* Validate the GHCB */
1837         ghcb_gpa = control->ghcb_gpa;
1838         if (ghcb_gpa & GHCB_MSR_INFO_MASK)
1839                 return sev_handle_vmgexit_msr_protocol(svm);
1840
1841         if (!ghcb_gpa) {
1842                 vcpu_unimpl(&svm->vcpu, "vmgexit: GHCB gpa is not set\n");
1843                 return -EINVAL;
1844         }
1845
1846         if (kvm_vcpu_map(&svm->vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->ghcb_map)) {
1847                 /* Unable to map GHCB from guest */
1848                 vcpu_unimpl(&svm->vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
1849                             ghcb_gpa);
1850                 return -EINVAL;
1851         }
1852
1853         svm->ghcb = svm->ghcb_map.hva;
1854         ghcb = svm->ghcb_map.hva;
1855
1856         trace_kvm_vmgexit_enter(svm->vcpu.vcpu_id, ghcb);
1857
1858         exit_code = ghcb_get_sw_exit_code(ghcb);
1859
1860         ret = sev_es_validate_vmgexit(svm);
1861         if (ret)
1862                 return ret;
1863
1864         sev_es_sync_from_ghcb(svm);
1865         ghcb_set_sw_exit_info_1(ghcb, 0);
1866         ghcb_set_sw_exit_info_2(ghcb, 0);
1867
1868         ret = -EINVAL;
1869         switch (exit_code) {
1870         case SVM_VMGEXIT_MMIO_READ:
1871                 if (!setup_vmgexit_scratch(svm, true, control->exit_info_2))
1872                         break;
1873
1874                 ret = kvm_sev_es_mmio_read(&svm->vcpu,
1875                                            control->exit_info_1,
1876                                            control->exit_info_2,
1877                                            svm->ghcb_sa);
1878                 break;
1879         case SVM_VMGEXIT_MMIO_WRITE:
1880                 if (!setup_vmgexit_scratch(svm, false, control->exit_info_2))
1881                         break;
1882
1883                 ret = kvm_sev_es_mmio_write(&svm->vcpu,
1884                                             control->exit_info_1,
1885                                             control->exit_info_2,
1886                                             svm->ghcb_sa);
1887                 break;
1888         case SVM_VMGEXIT_NMI_COMPLETE:
1889                 ret = svm_invoke_exit_handler(svm, SVM_EXIT_IRET);
1890                 break;
1891         case SVM_VMGEXIT_AP_JUMP_TABLE: {
1892                 struct kvm_sev_info *sev = &to_kvm_svm(svm->vcpu.kvm)->sev_info;
1893
1894                 switch (control->exit_info_1) {
1895                 case 0:
1896                         /* Set AP jump table address */
1897                         sev->ap_jump_table = control->exit_info_2;
1898                         break;
1899                 case 1:
1900                         /* Get AP jump table address */
1901                         ghcb_set_sw_exit_info_2(ghcb, sev->ap_jump_table);
1902                         break;
1903                 default:
1904                         pr_err("svm: vmgexit: unsupported AP jump table request - exit_info_1=%#llx\n",
1905                                control->exit_info_1);
1906                         ghcb_set_sw_exit_info_1(ghcb, 1);
1907                         ghcb_set_sw_exit_info_2(ghcb,
1908                                                 X86_TRAP_UD |
1909                                                 SVM_EVTINJ_TYPE_EXEPT |
1910                                                 SVM_EVTINJ_VALID);
1911                 }
1912
1913                 ret = 1;
1914                 break;
1915         }
1916         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
1917                 vcpu_unimpl(&svm->vcpu,
1918                             "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
1919                             control->exit_info_1, control->exit_info_2);
1920                 break;
1921         default:
1922                 ret = svm_invoke_exit_handler(svm, exit_code);
1923         }
1924
1925         return ret;
1926 }
1927
1928 int sev_es_string_io(struct vcpu_svm *svm, int size, unsigned int port, int in)
1929 {
1930         if (!setup_vmgexit_scratch(svm, in, svm->vmcb->control.exit_info_2))
1931                 return -EINVAL;
1932
1933         return kvm_sev_es_string_io(&svm->vcpu, size, port,
1934                                     svm->ghcb_sa, svm->ghcb_sa_len, in);
1935 }
1936
1937 void sev_es_init_vmcb(struct vcpu_svm *svm)
1938 {
1939         struct kvm_vcpu *vcpu = &svm->vcpu;
1940
1941         svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ES_ENABLE;
1942         svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
1943
1944         /*
1945          * An SEV-ES guest requires a VMSA area that is a separate from the
1946          * VMCB page. Do not include the encryption mask on the VMSA physical
1947          * address since hardware will access it using the guest key.
1948          */
1949         svm->vmcb->control.vmsa_pa = __pa(svm->vmsa);
1950
1951         /* Can't intercept CR register access, HV can't modify CR registers */
1952         svm_clr_intercept(svm, INTERCEPT_CR0_READ);
1953         svm_clr_intercept(svm, INTERCEPT_CR4_READ);
1954         svm_clr_intercept(svm, INTERCEPT_CR8_READ);
1955         svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
1956         svm_clr_intercept(svm, INTERCEPT_CR4_WRITE);
1957         svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
1958
1959         svm_clr_intercept(svm, INTERCEPT_SELECTIVE_CR0);
1960
1961         /* Track EFER/CR register changes */
1962         svm_set_intercept(svm, TRAP_EFER_WRITE);
1963         svm_set_intercept(svm, TRAP_CR0_WRITE);
1964         svm_set_intercept(svm, TRAP_CR4_WRITE);
1965         svm_set_intercept(svm, TRAP_CR8_WRITE);
1966
1967         /* No support for enable_vmware_backdoor */
1968         clr_exception_intercept(svm, GP_VECTOR);
1969
1970         /* Can't intercept XSETBV, HV can't modify XCR0 directly */
1971         svm_clr_intercept(svm, INTERCEPT_XSETBV);
1972
1973         /* Clear intercepts on selected MSRs */
1974         set_msr_interception(vcpu, svm->msrpm, MSR_EFER, 1, 1);
1975         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_CR_PAT, 1, 1);
1976         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
1977         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
1978         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
1979         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
1980 }
1981
1982 void sev_es_create_vcpu(struct vcpu_svm *svm)
1983 {
1984         /*
1985          * Set the GHCB MSR value as per the GHCB specification when creating
1986          * a vCPU for an SEV-ES guest.
1987          */
1988         set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
1989                                             GHCB_VERSION_MIN,
1990                                             sev_enc_bit));
1991 }
1992
1993 void sev_es_vcpu_load(struct vcpu_svm *svm, int cpu)
1994 {
1995         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1996         struct vmcb_save_area *hostsa;
1997         unsigned int i;
1998
1999         /*
2000          * As an SEV-ES guest, hardware will restore the host state on VMEXIT,
2001          * of which one step is to perform a VMLOAD. Since hardware does not
2002          * perform a VMSAVE on VMRUN, the host savearea must be updated.
2003          */
2004         asm volatile(__ex("vmsave") : : "a" (__sme_page_pa(sd->save_area)) : "memory");
2005
2006         /*
2007          * Certain MSRs are restored on VMEXIT, only save ones that aren't
2008          * restored.
2009          */
2010         for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++) {
2011                 if (host_save_user_msrs[i].sev_es_restored)
2012                         continue;
2013
2014                 rdmsrl(host_save_user_msrs[i].index, svm->host_user_msrs[i]);
2015         }
2016
2017         /* XCR0 is restored on VMEXIT, save the current host value */
2018         hostsa = (struct vmcb_save_area *)(page_address(sd->save_area) + 0x400);
2019         hostsa->xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
2020
2021         /* PKRU is restored on VMEXIT, save the curent host value */
2022         hostsa->pkru = read_pkru();
2023
2024         /* MSR_IA32_XSS is restored on VMEXIT, save the currnet host value */
2025         hostsa->xss = host_xss;
2026 }
2027
2028 void sev_es_vcpu_put(struct vcpu_svm *svm)
2029 {
2030         unsigned int i;
2031
2032         /*
2033          * Certain MSRs are restored on VMEXIT and were saved with vmsave in
2034          * sev_es_vcpu_load() above. Only restore ones that weren't.
2035          */
2036         for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++) {
2037                 if (host_save_user_msrs[i].sev_es_restored)
2038                         continue;
2039
2040                 wrmsrl(host_save_user_msrs[i].index, svm->host_user_msrs[i]);
2041         }
2042 }