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 #include <asm/fpu/internal.h>
  21
  22 #include <asm/trapnr.h>
  23
  24 #include "x86.h"
  25 #include "svm.h"
  26 #include "svm_ops.h"
  27 #include "cpuid.h"
  28 #include "trace.h"
  29
  30 #define __ex(x) __kvm_handle_fault_on_reboot(x)
  31
  32 #ifndef CONFIG_KVM_AMD_SEV
  33 /*
  34  * When this config is not defined, SEV feature is not supported and APIs in
  35  * this file are not used but this file still gets compiled into the KVM AMD
  36  * module.
  37  *
  38  * We will not have MISC_CG_RES_SEV and MISC_CG_RES_SEV_ES entries in the enum
  39  * misc_res_type {} defined in linux/misc_cgroup.h.
  40  *
  41  * Below macros allow compilation to succeed.
  42  */
  43 #define MISC_CG_RES_SEV MISC_CG_RES_TYPES
  44 #define MISC_CG_RES_SEV_ES MISC_CG_RES_TYPES
  45 #endif
  46
  47 static u8 sev_enc_bit;
  48 static int sev_flush_asids(void);
  49 static DECLARE_RWSEM(sev_deactivate_lock);
  50 static DEFINE_MUTEX(sev_bitmap_lock);
  51 unsigned int max_sev_asid;
  52 static unsigned int min_sev_asid;
  53 static unsigned long *sev_asid_bitmap;
  54 static unsigned long *sev_reclaim_asid_bitmap;
  55
  56 struct enc_region {
  57         struct list_head list;
  58         unsigned long npages;
  59         struct page **pages;
  60         unsigned long uaddr;
  61         unsigned long size;
  62 };
  63
  64 static int sev_flush_asids(void)
  65 {
  66         int ret, error = 0;
  67
  68         /*
  69          * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
  70          * so it must be guarded.
  71          */
  72         down_write(&sev_deactivate_lock);
  73
  74         wbinvd_on_all_cpus();
  75         ret = sev_guest_df_flush(&error);
  76
  77         up_write(&sev_deactivate_lock);
  78
  79         if (ret)
  80                 pr_err("SEV: DF_FLUSH failed, ret=%d, error=%#x\n", ret, error);
  81
  82         return ret;
  83 }
  84
  85 /* Must be called with the sev_bitmap_lock held */
  86 static bool __sev_recycle_asids(int min_asid, int max_asid)
  87 {
  88         int pos;
  89
  90         /* Check if there are any ASIDs to reclaim before performing a flush */
  91         pos = find_next_bit(sev_reclaim_asid_bitmap, max_sev_asid, min_asid);
  92         if (pos >= max_asid)
  93                 return false;
  94
  95         if (sev_flush_asids())
  96                 return false;
  97
  98         /* The flush process will flush all reclaimable SEV and SEV-ES ASIDs */
  99         bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
 100                    max_sev_asid);
 101         bitmap_zero(sev_reclaim_asid_bitmap, max_sev_asid);
 102
 103         return true;
 104 }
 105
 106 static int sev_asid_new(struct kvm_sev_info *sev)
 107 {
 108         int pos, min_asid, max_asid, ret;
 109         bool retry = true;
 110         enum misc_res_type type;
 111
 112         type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 113         WARN_ON(sev->misc_cg);
 114         sev->misc_cg = get_current_misc_cg();
 115         ret = misc_cg_try_charge(type, sev->misc_cg, 1);
 116         if (ret) {
 117                 put_misc_cg(sev->misc_cg);
 118                 sev->misc_cg = NULL;
 119                 return ret;
 120         }
 121
 122         mutex_lock(&sev_bitmap_lock);
 123
 124         /*
 125          * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
 126          * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.
 127          */
 128         min_asid = sev->es_active ? 0 : min_sev_asid - 1;
 129         max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
 130 again:
 131         pos = find_next_zero_bit(sev_asid_bitmap, max_sev_asid, min_asid);
 132         if (pos >= max_asid) {
 133                 if (retry && __sev_recycle_asids(min_asid, max_asid)) {
 134                         retry = false;
 135                         goto again;
 136                 }
 137                 mutex_unlock(&sev_bitmap_lock);
 138                 ret = -EBUSY;
 139                 goto e_uncharge;
 140         }
 141
 142         __set_bit(pos, sev_asid_bitmap);
 143
 144         mutex_unlock(&sev_bitmap_lock);
 145
 146         return pos + 1;
 147 e_uncharge:
 148         misc_cg_uncharge(type, sev->misc_cg, 1);
 149         put_misc_cg(sev->misc_cg);
 150         sev->misc_cg = NULL;
 151         return ret;
 152 }
 153
 154 static int sev_get_asid(struct kvm *kvm)
 155 {
 156         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 157
 158         return sev->asid;
 159 }
 160
 161 static void sev_asid_free(struct kvm_sev_info *sev)
 162 {
 163         struct svm_cpu_data *sd;
 164         int cpu, pos;
 165         enum misc_res_type type;
 166
 167         mutex_lock(&sev_bitmap_lock);
 168
 169         pos = sev->asid - 1;
 170         __set_bit(pos, sev_reclaim_asid_bitmap);
 171
 172         for_each_possible_cpu(cpu) {
 173                 sd = per_cpu(svm_data, cpu);
 174                 sd->sev_vmcbs[pos] = NULL;
 175         }
 176
 177         mutex_unlock(&sev_bitmap_lock);
 178
 179         type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 180         misc_cg_uncharge(type, sev->misc_cg, 1);
 181         put_misc_cg(sev->misc_cg);
 182         sev->misc_cg = NULL;
 183 }
 184
 185 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
 186 {
 187         struct sev_data_decommission *decommission;
 188         struct sev_data_deactivate *data;
 189
 190         if (!handle)
 191                 return;
 192
 193         data = kzalloc(sizeof(*data), GFP_KERNEL);
 194         if (!data)
 195                 return;
 196
 197         /* deactivate handle */
 198         data->handle = handle;
 199
 200         /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
 201         down_read(&sev_deactivate_lock);
 202         sev_guest_deactivate(data, NULL);
 203         up_read(&sev_deactivate_lock);
 204
 205         kfree(data);
 206
 207         decommission = kzalloc(sizeof(*decommission), GFP_KERNEL);
 208         if (!decommission)
 209                 return;
 210
 211         /* decommission handle */
 212         decommission->handle = handle;
 213         sev_guest_decommission(decommission, NULL);
 214
 215         kfree(decommission);
 216 }
 217
 218 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 219 {
 220         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 221         int asid, ret;
 222
 223         ret = -EBUSY;
 224         if (unlikely(sev->active))
 225                 return ret;
 226
 227         asid = sev_asid_new(sev);
 228         if (asid < 0)
 229                 return ret;
 230         sev->asid = asid;
 231
 232         ret = sev_platform_init(&argp->error);
 233         if (ret)
 234                 goto e_free;
 235
 236         sev->active = true;
 237         INIT_LIST_HEAD(&sev->regions_list);
 238
 239         return 0;
 240
 241 e_free:
 242         sev_asid_free(sev);
 243         sev->asid = 0;
 244         return ret;
 245 }
 246
 247 static int sev_es_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 248 {
 249         if (!sev_es)
 250                 return -ENOTTY;
 251
 252         to_kvm_svm(kvm)->sev_info.es_active = true;
 253
 254         return sev_guest_init(kvm, argp);
 255 }
 256
 257 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
 258 {
 259         struct sev_data_activate *data;
 260         int asid = sev_get_asid(kvm);
 261         int ret;
 262
 263         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 264         if (!data)
 265                 return -ENOMEM;
 266
 267         /* activate ASID on the given handle */
 268         data->handle = handle;
 269         data->asid   = asid;
 270         ret = sev_guest_activate(data, error);
 271         kfree(data);
 272
 273         return ret;
 274 }
 275
 276 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
 277 {
 278         struct fd f;
 279         int ret;
 280
 281         f = fdget(fd);
 282         if (!f.file)
 283                 return -EBADF;
 284
 285         ret = sev_issue_cmd_external_user(f.file, id, data, error);
 286
 287         fdput(f);
 288         return ret;
 289 }
 290
 291 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
 292 {
 293         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 294
 295         return __sev_issue_cmd(sev->fd, id, data, error);
 296 }
 297
 298 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
 299 {
 300         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 301         struct sev_data_launch_start *start;
 302         struct kvm_sev_launch_start params;
 303         void *dh_blob, *session_blob;
 304         int *error = &argp->error;
 305         int ret;
 306
 307         if (!sev_guest(kvm))
 308                 return -ENOTTY;
 309
 310         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 311                 return -EFAULT;
 312
 313         start = kzalloc(sizeof(*start), GFP_KERNEL_ACCOUNT);
 314         if (!start)
 315                 return -ENOMEM;
 316
 317         dh_blob = NULL;
 318         if (params.dh_uaddr) {
 319                 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
 320                 if (IS_ERR(dh_blob)) {
 321                         ret = PTR_ERR(dh_blob);
 322                         goto e_free;
 323                 }
 324
 325                 start->dh_cert_address = __sme_set(__pa(dh_blob));
 326                 start->dh_cert_len = params.dh_len;
 327         }
 328
 329         session_blob = NULL;
 330         if (params.session_uaddr) {
 331                 session_blob = psp_copy_user_blob(params.session_uaddr, params.session_len);
 332                 if (IS_ERR(session_blob)) {
 333                         ret = PTR_ERR(session_blob);
 334                         goto e_free_dh;
 335                 }
 336
 337                 start->session_address = __sme_set(__pa(session_blob));
 338                 start->session_len = params.session_len;
 339         }
 340
 341         start->handle = params.handle;
 342         start->policy = params.policy;
 343
 344         /* create memory encryption context */
 345         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_LAUNCH_START, start, error);
 346         if (ret)
 347                 goto e_free_session;
 348
 349         /* Bind ASID to this guest */
 350         ret = sev_bind_asid(kvm, start->handle, error);
 351         if (ret)
 352                 goto e_free_session;
 353
 354         /* return handle to userspace */
 355         params.handle = start->handle;
 356         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params))) {
 357                 sev_unbind_asid(kvm, start->handle);
 358                 ret = -EFAULT;
 359                 goto e_free_session;
 360         }
 361
 362         sev->handle = start->handle;
 363         sev->fd = argp->sev_fd;
 364
 365 e_free_session:
 366         kfree(session_blob);
 367 e_free_dh:
 368         kfree(dh_blob);
 369 e_free:
 370         kfree(start);
 371         return ret;
 372 }
 373
 374 static struct page **sev_pin_memory(struct kvm *kvm, unsigned long uaddr,
 375                                     unsigned long ulen, unsigned long *n,
 376                                     int write)
 377 {
 378         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 379         unsigned long npages, size;
 380         int npinned;
 381         unsigned long locked, lock_limit;
 382         struct page **pages;
 383         unsigned long first, last;
 384         int ret;
 385
 386         lockdep_assert_held(&kvm->lock);
 387
 388         if (ulen == 0 || uaddr + ulen < uaddr)
 389                 return ERR_PTR(-EINVAL);
 390
 391         /* Calculate number of pages. */
 392         first = (uaddr & PAGE_MASK) >> PAGE_SHIFT;
 393         last = ((uaddr + ulen - 1) & PAGE_MASK) >> PAGE_SHIFT;
 394         npages = (last - first + 1);
 395
 396         locked = sev->pages_locked + npages;
 397         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 398         if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
 399                 pr_err("SEV: %lu locked pages exceed the lock limit of %lu.\n", locked, lock_limit);
 400                 return ERR_PTR(-ENOMEM);
 401         }
 402
 403         if (WARN_ON_ONCE(npages > INT_MAX))
 404                 return ERR_PTR(-EINVAL);
 405
 406         /* Avoid using vmalloc for smaller buffers. */
 407         size = npages * sizeof(struct page *);
 408         if (size > PAGE_SIZE)
 409                 pages = __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 410         else
 411                 pages = kmalloc(size, GFP_KERNEL_ACCOUNT);
 412
 413         if (!pages)
 414                 return ERR_PTR(-ENOMEM);
 415
 416         /* Pin the user virtual address. */
 417         npinned = pin_user_pages_fast(uaddr, npages, write ? FOLL_WRITE : 0, pages);
 418         if (npinned != npages) {
 419                 pr_err("SEV: Failure locking %lu pages.\n", npages);
 420                 ret = -ENOMEM;
 421                 goto err;
 422         }
 423
 424         *n = npages;
 425         sev->pages_locked = locked;
 426
 427         return pages;
 428
 429 err:
 430         if (npinned > 0)
 431                 unpin_user_pages(pages, npinned);
 432
 433         kvfree(pages);
 434         return ERR_PTR(ret);
 435 }
 436
 437 static void sev_unpin_memory(struct kvm *kvm, struct page **pages,
 438                              unsigned long npages)
 439 {
 440         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 441
 442         unpin_user_pages(pages, npages);
 443         kvfree(pages);
 444         sev->pages_locked -= npages;
 445 }
 446
 447 static void sev_clflush_pages(struct page *pages[], unsigned long npages)
 448 {
 449         uint8_t *page_virtual;
 450         unsigned long i;
 451
 452         if (this_cpu_has(X86_FEATURE_SME_COHERENT) || npages == 0 ||
 453             pages == NULL)
 454                 return;
 455
 456         for (i = 0; i < npages; i++) {
 457                 page_virtual = kmap_atomic(pages[i]);
 458                 clflush_cache_range(page_virtual, PAGE_SIZE);
 459                 kunmap_atomic(page_virtual);
 460         }
 461 }
 462
 463 static unsigned long get_num_contig_pages(unsigned long idx,
 464                                 struct page **inpages, unsigned long npages)
 465 {
 466         unsigned long paddr, next_paddr;
 467         unsigned long i = idx + 1, pages = 1;
 468
 469         /* find the number of contiguous pages starting from idx */
 470         paddr = __sme_page_pa(inpages[idx]);
 471         while (i < npages) {
 472                 next_paddr = __sme_page_pa(inpages[i++]);
 473                 if ((paddr + PAGE_SIZE) == next_paddr) {
 474                         pages++;
 475                         paddr = next_paddr;
 476                         continue;
 477                 }
 478                 break;
 479         }
 480
 481         return pages;
 482 }
 483
 484 static int sev_launch_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
 485 {
 486         unsigned long vaddr, vaddr_end, next_vaddr, npages, pages, size, i;
 487         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 488         struct kvm_sev_launch_update_data params;
 489         struct sev_data_launch_update_data *data;
 490         struct page **inpages;
 491         int ret;
 492
 493         if (!sev_guest(kvm))
 494                 return -ENOTTY;
 495
 496         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 497                 return -EFAULT;
 498
 499         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 500         if (!data)
 501                 return -ENOMEM;
 502
 503         vaddr = params.uaddr;
 504         size = params.len;
 505         vaddr_end = vaddr + size;
 506
 507         /* Lock the user memory. */
 508         inpages = sev_pin_memory(kvm, vaddr, size, &npages, 1);
 509         if (IS_ERR(inpages)) {
 510                 ret = PTR_ERR(inpages);
 511                 goto e_free;
 512         }
 513
 514         /*
 515          * Flush (on non-coherent CPUs) before LAUNCH_UPDATE encrypts pages in
 516          * place; the cache may contain the data that was written unencrypted.
 517          */
 518         sev_clflush_pages(inpages, npages);
 519
 520         for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
 521                 int offset, len;
 522
 523                 /*
 524                  * If the user buffer is not page-aligned, calculate the offset
 525                  * within the page.
 526                  */
 527                 offset = vaddr & (PAGE_SIZE - 1);
 528
 529                 /* Calculate the number of pages that can be encrypted in one go. */
 530                 pages = get_num_contig_pages(i, inpages, npages);
 531
 532                 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
 533
 534                 data->handle = sev->handle;
 535                 data->len = len;
 536                 data->address = __sme_page_pa(inpages[i]) + offset;
 537                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, data, &argp->error);
 538                 if (ret)
 539                         goto e_unpin;
 540
 541                 size -= len;
 542                 next_vaddr = vaddr + len;
 543         }
 544
 545 e_unpin:
 546         /* content of memory is updated, mark pages dirty */
 547         for (i = 0; i < npages; i++) {
 548                 set_page_dirty_lock(inpages[i]);
 549                 mark_page_accessed(inpages[i]);
 550         }
 551         /* unlock the user pages */
 552         sev_unpin_memory(kvm, inpages, npages);
 553 e_free:
 554         kfree(data);
 555         return ret;
 556 }
 557
 558 static int sev_es_sync_vmsa(struct vcpu_svm *svm)
 559 {
 560         struct vmcb_save_area *save = &svm->vmcb->save;
 561
 562         /* Check some debug related fields before encrypting the VMSA */
 563         if (svm->vcpu.guest_debug || (save->dr7 & ~DR7_FIXED_1))
 564                 return -EINVAL;
 565
 566         /* Sync registgers */
 567         save->rax = svm->vcpu.arch.regs[VCPU_REGS_RAX];
 568         save->rbx = svm->vcpu.arch.regs[VCPU_REGS_RBX];
 569         save->rcx = svm->vcpu.arch.regs[VCPU_REGS_RCX];
 570         save->rdx = svm->vcpu.arch.regs[VCPU_REGS_RDX];
 571         save->rsp = svm->vcpu.arch.regs[VCPU_REGS_RSP];
 572         save->rbp = svm->vcpu.arch.regs[VCPU_REGS_RBP];
 573         save->rsi = svm->vcpu.arch.regs[VCPU_REGS_RSI];
 574         save->rdi = svm->vcpu.arch.regs[VCPU_REGS_RDI];
 575 #ifdef CONFIG_X86_64
 576         save->r8  = svm->vcpu.arch.regs[VCPU_REGS_R8];
 577         save->r9  = svm->vcpu.arch.regs[VCPU_REGS_R9];
 578         save->r10 = svm->vcpu.arch.regs[VCPU_REGS_R10];
 579         save->r11 = svm->vcpu.arch.regs[VCPU_REGS_R11];
 580         save->r12 = svm->vcpu.arch.regs[VCPU_REGS_R12];
 581         save->r13 = svm->vcpu.arch.regs[VCPU_REGS_R13];
 582         save->r14 = svm->vcpu.arch.regs[VCPU_REGS_R14];
 583         save->r15 = svm->vcpu.arch.regs[VCPU_REGS_R15];
 584 #endif
 585         save->rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
 586
 587         /* Sync some non-GPR registers before encrypting */
 588         save->xcr0 = svm->vcpu.arch.xcr0;
 589         save->pkru = svm->vcpu.arch.pkru;
 590         save->xss  = svm->vcpu.arch.ia32_xss;
 591
 592         /*
 593          * SEV-ES will use a VMSA that is pointed to by the VMCB, not
 594          * the traditional VMSA that is part of the VMCB. Copy the
 595          * traditional VMSA as it has been built so far (in prep
 596          * for LAUNCH_UPDATE_VMSA) to be the initial SEV-ES state.
 597          */
 598         memcpy(svm->vmsa, save, sizeof(*save));
 599
 600         return 0;
 601 }
 602
 603 static int sev_launch_update_vmsa(struct kvm *kvm, struct kvm_sev_cmd *argp)
 604 {
 605         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 606         struct sev_data_launch_update_vmsa *vmsa;
 607         int i, ret;
 608
 609         if (!sev_es_guest(kvm))
 610                 return -ENOTTY;
 611
 612         vmsa = kzalloc(sizeof(*vmsa), GFP_KERNEL);
 613         if (!vmsa)
 614                 return -ENOMEM;
 615
 616         for (i = 0; i < kvm->created_vcpus; i++) {
 617                 struct vcpu_svm *svm = to_svm(kvm->vcpus[i]);
 618
 619                 /* Perform some pre-encryption checks against the VMSA */
 620                 ret = sev_es_sync_vmsa(svm);
 621                 if (ret)
 622                         goto e_free;
 623
 624                 /*
 625                  * The LAUNCH_UPDATE_VMSA command will perform in-place
 626                  * encryption of the VMSA memory content (i.e it will write
 627                  * the same memory region with the guest's key), so invalidate
 628                  * it first.
 629                  */
 630                 clflush_cache_range(svm->vmsa, PAGE_SIZE);
 631
 632                 vmsa->handle = sev->handle;
 633                 vmsa->address = __sme_pa(svm->vmsa);
 634                 vmsa->len = PAGE_SIZE;
 635                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_VMSA, vmsa,
 636                                     &argp->error);
 637                 if (ret)
 638                         goto e_free;
 639
 640                 svm->vcpu.arch.guest_state_protected = true;
 641         }
 642
 643 e_free:
 644         kfree(vmsa);
 645         return ret;
 646 }
 647
 648 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 649 {
 650         void __user *measure = (void __user *)(uintptr_t)argp->data;
 651         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 652         struct sev_data_launch_measure *data;
 653         struct kvm_sev_launch_measure params;
 654         void __user *p = NULL;
 655         void *blob = NULL;
 656         int ret;
 657
 658         if (!sev_guest(kvm))
 659                 return -ENOTTY;
 660
 661         if (copy_from_user(&params, measure, sizeof(params)))
 662                 return -EFAULT;
 663
 664         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 665         if (!data)
 666                 return -ENOMEM;
 667
 668         /* User wants to query the blob length */
 669         if (!params.len)
 670                 goto cmd;
 671
 672         p = (void __user *)(uintptr_t)params.uaddr;
 673         if (p) {
 674                 if (params.len > SEV_FW_BLOB_MAX_SIZE) {
 675                         ret = -EINVAL;
 676                         goto e_free;
 677                 }
 678
 679                 ret = -ENOMEM;
 680                 blob = kmalloc(params.len, GFP_KERNEL);
 681                 if (!blob)
 682                         goto e_free;
 683
 684                 data->address = __psp_pa(blob);
 685                 data->len = params.len;
 686         }
 687
 688 cmd:
 689         data->handle = sev->handle;
 690         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, data, &argp->error);
 691
 692         /*
 693          * If we query the session length, FW responded with expected data.
 694          */
 695         if (!params.len)
 696                 goto done;
 697
 698         if (ret)
 699                 goto e_free_blob;
 700
 701         if (blob) {
 702                 if (copy_to_user(p, blob, params.len))
 703                         ret = -EFAULT;
 704         }
 705
 706 done:
 707         params.len = data->len;
 708         if (copy_to_user(measure, &params, sizeof(params)))
 709                 ret = -EFAULT;
 710 e_free_blob:
 711         kfree(blob);
 712 e_free:
 713         kfree(data);
 714         return ret;
 715 }
 716
 717 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 718 {
 719         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 720         struct sev_data_launch_finish *data;
 721         int ret;
 722
 723         if (!sev_guest(kvm))
 724                 return -ENOTTY;
 725
 726         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 727         if (!data)
 728                 return -ENOMEM;
 729
 730         data->handle = sev->handle;
 731         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, data, &argp->error);
 732
 733         kfree(data);
 734         return ret;
 735 }
 736
 737 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 738 {
 739         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 740         struct kvm_sev_guest_status params;
 741         struct sev_data_guest_status *data;
 742         int ret;
 743
 744         if (!sev_guest(kvm))
 745                 return -ENOTTY;
 746
 747         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 748         if (!data)
 749                 return -ENOMEM;
 750
 751         data->handle = sev->handle;
 752         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, data, &argp->error);
 753         if (ret)
 754                 goto e_free;
 755
 756         params.policy = data->policy;
 757         params.state = data->state;
 758         params.handle = data->handle;
 759
 760         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 761                 ret = -EFAULT;
 762 e_free:
 763         kfree(data);
 764         return ret;
 765 }
 766
 767 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 768                                unsigned long dst, int size,
 769                                int *error, bool enc)
 770 {
 771         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 772         struct sev_data_dbg *data;
 773         int ret;
 774
 775         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 776         if (!data)
 777                 return -ENOMEM;
 778
 779         data->handle = sev->handle;
 780         data->dst_addr = dst;
 781         data->src_addr = src;
 782         data->len = size;
 783
 784         ret = sev_issue_cmd(kvm,
 785                             enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 786                             data, error);
 787         kfree(data);
 788         return ret;
 789 }
 790
 791 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 792                              unsigned long dst_paddr, int sz, int *err)
 793 {
 794         int offset;
 795
 796         /*
 797          * Its safe to read more than we are asked, caller should ensure that
 798          * destination has enough space.
 799          */
 800         offset = src_paddr & 15;
 801         src_paddr = round_down(src_paddr, 16);
 802         sz = round_up(sz + offset, 16);
 803
 804         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 805 }
 806
 807 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 808                                   unsigned long __user dst_uaddr,
 809                                   unsigned long dst_paddr,
 810                                   int size, int *err)
 811 {
 812         struct page *tpage = NULL;
 813         int ret, offset;
 814
 815         /* if inputs are not 16-byte then use intermediate buffer */
 816         if (!IS_ALIGNED(dst_paddr, 16) ||
 817             !IS_ALIGNED(paddr,     16) ||
 818             !IS_ALIGNED(size,      16)) {
 819                 tpage = (void *)alloc_page(GFP_KERNEL);
 820                 if (!tpage)
 821                         return -ENOMEM;
 822
 823                 dst_paddr = __sme_page_pa(tpage);
 824         }
 825
 826         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 827         if (ret)
 828                 goto e_free;
 829
 830         if (tpage) {
 831                 offset = paddr & 15;
 832                 if (copy_to_user((void __user *)(uintptr_t)dst_uaddr,
 833                                  page_address(tpage) + offset, size))
 834                         ret = -EFAULT;
 835         }
 836
 837 e_free:
 838         if (tpage)
 839                 __free_page(tpage);
 840
 841         return ret;
 842 }
 843
 844 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 845                                   unsigned long __user vaddr,
 846                                   unsigned long dst_paddr,
 847                                   unsigned long __user dst_vaddr,
 848                                   int size, int *error)
 849 {
 850         struct page *src_tpage = NULL;
 851         struct page *dst_tpage = NULL;
 852         int ret, len = size;
 853
 854         /* If source buffer is not aligned then use an intermediate buffer */
 855         if (!IS_ALIGNED(vaddr, 16)) {
 856                 src_tpage = alloc_page(GFP_KERNEL);
 857                 if (!src_tpage)
 858                         return -ENOMEM;
 859
 860                 if (copy_from_user(page_address(src_tpage),
 861                                 (void __user *)(uintptr_t)vaddr, size)) {
 862                         __free_page(src_tpage);
 863                         return -EFAULT;
 864                 }
 865
 866                 paddr = __sme_page_pa(src_tpage);
 867         }
 868
 869         /*
 870          *  If destination buffer or length is not aligned then do read-modify-write:
 871          *   - decrypt destination in an intermediate buffer
 872          *   - copy the source buffer in an intermediate buffer
 873          *   - use the intermediate buffer as source buffer
 874          */
 875         if (!IS_ALIGNED(dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 876                 int dst_offset;
 877
 878                 dst_tpage = alloc_page(GFP_KERNEL);
 879                 if (!dst_tpage) {
 880                         ret = -ENOMEM;
 881                         goto e_free;
 882                 }
 883
 884                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 885                                         __sme_page_pa(dst_tpage), size, error);
 886                 if (ret)
 887                         goto e_free;
 888
 889                 /*
 890                  *  If source is kernel buffer then use memcpy() otherwise
 891                  *  copy_from_user().
 892                  */
 893                 dst_offset = dst_paddr & 15;
 894
 895                 if (src_tpage)
 896                         memcpy(page_address(dst_tpage) + dst_offset,
 897                                page_address(src_tpage), size);
 898                 else {
 899                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 900                                            (void __user *)(uintptr_t)vaddr, size)) {
 901                                 ret = -EFAULT;
 902                                 goto e_free;
 903                         }
 904                 }
 905
 906                 paddr = __sme_page_pa(dst_tpage);
 907                 dst_paddr = round_down(dst_paddr, 16);
 908                 len = round_up(size, 16);
 909         }
 910
 911         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 912
 913 e_free:
 914         if (src_tpage)
 915                 __free_page(src_tpage);
 916         if (dst_tpage)
 917                 __free_page(dst_tpage);
 918         return ret;
 919 }
 920
 921 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 922 {
 923         unsigned long vaddr, vaddr_end, next_vaddr;
 924         unsigned long dst_vaddr;
 925         struct page **src_p, **dst_p;
 926         struct kvm_sev_dbg debug;
 927         unsigned long n;
 928         unsigned int size;
 929         int ret;
 930
 931         if (!sev_guest(kvm))
 932                 return -ENOTTY;
 933
 934         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 935                 return -EFAULT;
 936
 937         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 938                 return -EINVAL;
 939         if (!debug.dst_uaddr)
 940                 return -EINVAL;
 941
 942         vaddr = debug.src_uaddr;
 943         size = debug.len;
 944         vaddr_end = vaddr + size;
 945         dst_vaddr = debug.dst_uaddr;
 946
 947         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 948                 int len, s_off, d_off;
 949
 950                 /* lock userspace source and destination page */
 951                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 952                 if (IS_ERR(src_p))
 953                         return PTR_ERR(src_p);
 954
 955                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 956                 if (IS_ERR(dst_p)) {
 957                         sev_unpin_memory(kvm, src_p, n);
 958                         return PTR_ERR(dst_p);
 959                 }
 960
 961                 /*
 962                  * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
 963                  * the pages; flush the destination too so that future accesses do not
 964                  * see stale data.
 965                  */
 966                 sev_clflush_pages(src_p, 1);
 967                 sev_clflush_pages(dst_p, 1);
 968
 969                 /*
 970                  * Since user buffer may not be page aligned, calculate the
 971                  * offset within the page.
 972                  */
 973                 s_off = vaddr & ~PAGE_MASK;
 974                 d_off = dst_vaddr & ~PAGE_MASK;
 975                 len = min_t(size_t, (PAGE_SIZE - s_off), size);
 976
 977                 if (dec)
 978                         ret = __sev_dbg_decrypt_user(kvm,
 979                                                      __sme_page_pa(src_p[0]) + s_off,
 980                                                      dst_vaddr,
 981                                                      __sme_page_pa(dst_p[0]) + d_off,
 982                                                      len, &argp->error);
 983                 else
 984                         ret = __sev_dbg_encrypt_user(kvm,
 985                                                      __sme_page_pa(src_p[0]) + s_off,
 986                                                      vaddr,
 987                                                      __sme_page_pa(dst_p[0]) + d_off,
 988                                                      dst_vaddr,
 989                                                      len, &argp->error);
 990
 991                 sev_unpin_memory(kvm, src_p, n);
 992                 sev_unpin_memory(kvm, dst_p, n);
 993
 994                 if (ret)
 995                         goto err;
 996
 997                 next_vaddr = vaddr + len;
 998                 dst_vaddr = dst_vaddr + len;
 999                 size -= len;
1000         }
1001 err:
1002         return ret;
1003 }
1004
1005 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
1006 {
1007         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1008         struct sev_data_launch_secret *data;
1009         struct kvm_sev_launch_secret params;
1010         struct page **pages;
1011         void *blob, *hdr;
1012         unsigned long n, i;
1013         int ret, offset;
1014
1015         if (!sev_guest(kvm))
1016                 return -ENOTTY;
1017
1018         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1019                 return -EFAULT;
1020
1021         pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, 1);
1022         if (IS_ERR(pages))
1023                 return PTR_ERR(pages);
1024
1025         /*
1026          * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
1027          * place; the cache may contain the data that was written unencrypted.
1028          */
1029         sev_clflush_pages(pages, n);
1030
1031         /*
1032          * The secret must be copied into contiguous memory region, lets verify
1033          * that userspace memory pages are contiguous before we issue command.
1034          */
1035         if (get_num_contig_pages(0, pages, n) != n) {
1036                 ret = -EINVAL;
1037                 goto e_unpin_memory;
1038         }
1039
1040         ret = -ENOMEM;
1041         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
1042         if (!data)
1043                 goto e_unpin_memory;
1044
1045         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1046         data->guest_address = __sme_page_pa(pages[0]) + offset;
1047         data->guest_len = params.guest_len;
1048
1049         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1050         if (IS_ERR(blob)) {
1051                 ret = PTR_ERR(blob);
1052                 goto e_free;
1053         }
1054
1055         data->trans_address = __psp_pa(blob);
1056         data->trans_len = params.trans_len;
1057
1058         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1059         if (IS_ERR(hdr)) {
1060                 ret = PTR_ERR(hdr);
1061                 goto e_free_blob;
1062         }
1063         data->hdr_address = __psp_pa(hdr);
1064         data->hdr_len = params.hdr_len;
1065
1066         data->handle = sev->handle;
1067         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, data, &argp->error);
1068
1069         kfree(hdr);
1070
1071 e_free_blob:
1072         kfree(blob);
1073 e_free:
1074         kfree(data);
1075 e_unpin_memory:
1076         /* content of memory is updated, mark pages dirty */
1077         for (i = 0; i < n; i++) {
1078                 set_page_dirty_lock(pages[i]);
1079                 mark_page_accessed(pages[i]);
1080         }
1081         sev_unpin_memory(kvm, pages, n);
1082         return ret;
1083 }
1084
1085 static int sev_get_attestation_report(struct kvm *kvm, struct kvm_sev_cmd *argp)
1086 {
1087         void __user *report = (void __user *)(uintptr_t)argp->data;
1088         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1089         struct sev_data_attestation_report *data;
1090         struct kvm_sev_attestation_report params;
1091         void __user *p;
1092         void *blob = NULL;
1093         int ret;
1094
1095         if (!sev_guest(kvm))
1096                 return -ENOTTY;
1097
1098         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1099                 return -EFAULT;
1100
1101         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
1102         if (!data)
1103                 return -ENOMEM;
1104
1105         /* User wants to query the blob length */
1106         if (!params.len)
1107                 goto cmd;
1108
1109         p = (void __user *)(uintptr_t)params.uaddr;
1110         if (p) {
1111                 if (params.len > SEV_FW_BLOB_MAX_SIZE) {
1112                         ret = -EINVAL;
1113                         goto e_free;
1114                 }
1115
1116                 ret = -ENOMEM;
1117                 blob = kmalloc(params.len, GFP_KERNEL);
1118                 if (!blob)
1119                         goto e_free;
1120
1121                 data->address = __psp_pa(blob);
1122                 data->len = params.len;
1123                 memcpy(data->mnonce, params.mnonce, sizeof(params.mnonce));
1124         }
1125 cmd:
1126         data->handle = sev->handle;
1127         ret = sev_issue_cmd(kvm, SEV_CMD_ATTESTATION_REPORT, data, &argp->error);
1128         /*
1129          * If we query the session length, FW responded with expected data.
1130          */
1131         if (!params.len)
1132                 goto done;
1133
1134         if (ret)
1135                 goto e_free_blob;
1136
1137         if (blob) {
1138                 if (copy_to_user(p, blob, params.len))
1139                         ret = -EFAULT;
1140         }
1141
1142 done:
1143         params.len = data->len;
1144         if (copy_to_user(report, &params, sizeof(params)))
1145                 ret = -EFAULT;
1146 e_free_blob:
1147         kfree(blob);
1148 e_free:
1149         kfree(data);
1150         return ret;
1151 }
1152
1153 int svm_mem_enc_op(struct kvm *kvm, void __user *argp)
1154 {
1155         struct kvm_sev_cmd sev_cmd;
1156         int r;
1157
1158         if (!svm_sev_enabled() || !sev)
1159                 return -ENOTTY;
1160
1161         if (!argp)
1162                 return 0;
1163
1164         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
1165                 return -EFAULT;
1166
1167         mutex_lock(&kvm->lock);
1168
1169         switch (sev_cmd.id) {
1170         case KVM_SEV_INIT:
1171                 r = sev_guest_init(kvm, &sev_cmd);
1172                 break;
1173         case KVM_SEV_ES_INIT:
1174                 r = sev_es_guest_init(kvm, &sev_cmd);
1175                 break;
1176         case KVM_SEV_LAUNCH_START:
1177                 r = sev_launch_start(kvm, &sev_cmd);
1178                 break;
1179         case KVM_SEV_LAUNCH_UPDATE_DATA:
1180                 r = sev_launch_update_data(kvm, &sev_cmd);
1181                 break;
1182         case KVM_SEV_LAUNCH_UPDATE_VMSA:
1183                 r = sev_launch_update_vmsa(kvm, &sev_cmd);
1184                 break;
1185         case KVM_SEV_LAUNCH_MEASURE:
1186                 r = sev_launch_measure(kvm, &sev_cmd);
1187                 break;
1188         case KVM_SEV_LAUNCH_FINISH:
1189                 r = sev_launch_finish(kvm, &sev_cmd);
1190                 break;
1191         case KVM_SEV_GUEST_STATUS:
1192                 r = sev_guest_status(kvm, &sev_cmd);
1193                 break;
1194         case KVM_SEV_DBG_DECRYPT:
1195                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
1196                 break;
1197         case KVM_SEV_DBG_ENCRYPT:
1198                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
1199                 break;
1200         case KVM_SEV_LAUNCH_SECRET:
1201                 r = sev_launch_secret(kvm, &sev_cmd);
1202                 break;
1203         case KVM_SEV_GET_ATTESTATION_REPORT:
1204                 r = sev_get_attestation_report(kvm, &sev_cmd);
1205                 break;
1206         default:
1207                 r = -EINVAL;
1208                 goto out;
1209         }
1210
1211         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
1212                 r = -EFAULT;
1213
1214 out:
1215         mutex_unlock(&kvm->lock);
1216         return r;
1217 }
1218
1219 int svm_register_enc_region(struct kvm *kvm,
1220                             struct kvm_enc_region *range)
1221 {
1222         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1223         struct enc_region *region;
1224         int ret = 0;
1225
1226         if (!sev_guest(kvm))
1227                 return -ENOTTY;
1228
1229         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
1230                 return -EINVAL;
1231
1232         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
1233         if (!region)
1234                 return -ENOMEM;
1235
1236         mutex_lock(&kvm->lock);
1237         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
1238         if (IS_ERR(region->pages)) {
1239                 ret = PTR_ERR(region->pages);
1240                 mutex_unlock(&kvm->lock);
1241                 goto e_free;
1242         }
1243
1244         region->uaddr = range->addr;
1245         region->size = range->size;
1246
1247         list_add_tail(&region->list, &sev->regions_list);
1248         mutex_unlock(&kvm->lock);
1249
1250         /*
1251          * The guest may change the memory encryption attribute from C=0 -> C=1
1252          * or vice versa for this memory range. Lets make sure caches are
1253          * flushed to ensure that guest data gets written into memory with
1254          * correct C-bit.
1255          */
1256         sev_clflush_pages(region->pages, region->npages);
1257
1258         return ret;
1259
1260 e_free:
1261         kfree(region);
1262         return ret;
1263 }
1264
1265 static struct enc_region *
1266 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
1267 {
1268         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1269         struct list_head *head = &sev->regions_list;
1270         struct enc_region *i;
1271
1272         list_for_each_entry(i, head, list) {
1273                 if (i->uaddr == range->addr &&
1274                     i->size == range->size)
1275                         return i;
1276         }
1277
1278         return NULL;
1279 }
1280
1281 static void __unregister_enc_region_locked(struct kvm *kvm,
1282                                            struct enc_region *region)
1283 {
1284         sev_unpin_memory(kvm, region->pages, region->npages);
1285         list_del(&region->list);
1286         kfree(region);
1287 }
1288
1289 int svm_unregister_enc_region(struct kvm *kvm,
1290                               struct kvm_enc_region *range)
1291 {
1292         struct enc_region *region;
1293         int ret;
1294
1295         mutex_lock(&kvm->lock);
1296
1297         if (!sev_guest(kvm)) {
1298                 ret = -ENOTTY;
1299                 goto failed;
1300         }
1301
1302         region = find_enc_region(kvm, range);
1303         if (!region) {
1304                 ret = -EINVAL;
1305                 goto failed;
1306         }
1307
1308         /*
1309          * Ensure that all guest tagged cache entries are flushed before
1310          * releasing the pages back to the system for use. CLFLUSH will
1311          * not do this, so issue a WBINVD.
1312          */
1313         wbinvd_on_all_cpus();
1314
1315         __unregister_enc_region_locked(kvm, region);
1316
1317         mutex_unlock(&kvm->lock);
1318         return 0;
1319
1320 failed:
1321         mutex_unlock(&kvm->lock);
1322         return ret;
1323 }
1324
1325 void sev_vm_destroy(struct kvm *kvm)
1326 {
1327         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1328         struct list_head *head = &sev->regions_list;
1329         struct list_head *pos, *q;
1330
1331         if (!sev_guest(kvm))
1332                 return;
1333
1334         mutex_lock(&kvm->lock);
1335
1336         /*
1337          * Ensure that all guest tagged cache entries are flushed before
1338          * releasing the pages back to the system for use. CLFLUSH will
1339          * not do this, so issue a WBINVD.
1340          */
1341         wbinvd_on_all_cpus();
1342
1343         /*
1344          * if userspace was terminated before unregistering the memory regions
1345          * then lets unpin all the registered memory.
1346          */
1347         if (!list_empty(head)) {
1348                 list_for_each_safe(pos, q, head) {
1349                         __unregister_enc_region_locked(kvm,
1350                                 list_entry(pos, struct enc_region, list));
1351                         cond_resched();
1352                 }
1353         }
1354
1355         mutex_unlock(&kvm->lock);
1356
1357         sev_unbind_asid(kvm, sev->handle);
1358         sev_asid_free(sev);
1359 }
1360
1361 void __init sev_hardware_setup(void)
1362 {
1363         unsigned int eax, ebx, ecx, edx, sev_asid_count, sev_es_asid_count;
1364         bool sev_es_supported = false;
1365         bool sev_supported = false;
1366
1367         /* Does the CPU support SEV? */
1368         if (!boot_cpu_has(X86_FEATURE_SEV))
1369                 goto out;
1370
1371         /* Retrieve SEV CPUID information */
1372         cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
1373
1374         /* Set encryption bit location for SEV-ES guests */
1375         sev_enc_bit = ebx & 0x3f;
1376
1377         /* Maximum number of encrypted guests supported simultaneously */
1378         max_sev_asid = ecx;
1379
1380         if (!svm_sev_enabled())
1381                 goto out;
1382
1383         /* Minimum ASID value that should be used for SEV guest */
1384         min_sev_asid = edx;
1385
1386         /* Initialize SEV ASID bitmaps */
1387         sev_asid_bitmap = bitmap_zalloc(max_sev_asid, GFP_KERNEL);
1388         if (!sev_asid_bitmap)
1389                 goto out;
1390
1391         sev_reclaim_asid_bitmap = bitmap_zalloc(max_sev_asid, GFP_KERNEL);
1392         if (!sev_reclaim_asid_bitmap)
1393                 goto out;
1394
1395         sev_asid_count = max_sev_asid - min_sev_asid + 1;
1396         if (misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count))
1397                 goto out;
1398
1399         pr_info("SEV supported: %u ASIDs\n", sev_asid_count);
1400         sev_supported = true;
1401
1402         /* SEV-ES support requested? */
1403         if (!sev_es)
1404                 goto out;
1405
1406         /* Does the CPU support SEV-ES? */
1407         if (!boot_cpu_has(X86_FEATURE_SEV_ES))
1408                 goto out;
1409
1410         /* Has the system been allocated ASIDs for SEV-ES? */
1411         if (min_sev_asid == 1)
1412                 goto out;
1413
1414         sev_es_asid_count = min_sev_asid - 1;
1415         if (misc_cg_set_capacity(MISC_CG_RES_SEV_ES, sev_es_asid_count))
1416                 goto out;
1417
1418         pr_info("SEV-ES supported: %u ASIDs\n", sev_es_asid_count);
1419         sev_es_supported = true;
1420
1421 out:
1422         sev = sev_supported;
1423         sev_es = sev_es_supported;
1424 }
1425
1426 void sev_hardware_teardown(void)
1427 {
1428         if (!svm_sev_enabled())
1429                 return;
1430
1431         bitmap_free(sev_asid_bitmap);
1432         bitmap_free(sev_reclaim_asid_bitmap);
1433         misc_cg_set_capacity(MISC_CG_RES_SEV, 0);
1434         misc_cg_set_capacity(MISC_CG_RES_SEV_ES, 0);
1435
1436         sev_flush_asids();
1437 }
1438
1439 /*
1440  * Pages used by hardware to hold guest encrypted state must be flushed before
1441  * returning them to the system.
1442  */
1443 static void sev_flush_guest_memory(struct vcpu_svm *svm, void *va,
1444                                    unsigned long len)
1445 {
1446         /*
1447          * If hardware enforced cache coherency for encrypted mappings of the
1448          * same physical page is supported, nothing to do.
1449          */
1450         if (boot_cpu_has(X86_FEATURE_SME_COHERENT))
1451                 return;
1452
1453         /*
1454          * If the VM Page Flush MSR is supported, use it to flush the page
1455          * (using the page virtual address and the guest ASID).
1456          */
1457         if (boot_cpu_has(X86_FEATURE_VM_PAGE_FLUSH)) {
1458                 struct kvm_sev_info *sev;
1459                 unsigned long va_start;
1460                 u64 start, stop;
1461
1462                 /* Align start and stop to page boundaries. */
1463                 va_start = (unsigned long)va;
1464                 start = (u64)va_start & PAGE_MASK;
1465                 stop = PAGE_ALIGN((u64)va_start + len);
1466
1467                 if (start < stop) {
1468                         sev = &to_kvm_svm(svm->vcpu.kvm)->sev_info;
1469
1470                         while (start < stop) {
1471                                 wrmsrl(MSR_AMD64_VM_PAGE_FLUSH,
1472                                        start | sev->asid);
1473
1474                                 start += PAGE_SIZE;
1475                         }
1476
1477                         return;
1478                 }
1479
1480                 WARN(1, "Address overflow, using WBINVD\n");
1481         }
1482
1483         /*
1484          * Hardware should always have one of the above features,
1485          * but if not, use WBINVD and issue a warning.
1486          */
1487         WARN_ONCE(1, "Using WBINVD to flush guest memory\n");
1488         wbinvd_on_all_cpus();
1489 }
1490
1491 void sev_free_vcpu(struct kvm_vcpu *vcpu)
1492 {
1493         struct vcpu_svm *svm;
1494
1495         if (!sev_es_guest(vcpu->kvm))
1496                 return;
1497
1498         svm = to_svm(vcpu);
1499
1500         if (vcpu->arch.guest_state_protected)
1501                 sev_flush_guest_memory(svm, svm->vmsa, PAGE_SIZE);
1502         __free_page(virt_to_page(svm->vmsa));
1503
1504         if (svm->ghcb_sa_free)
1505                 kfree(svm->ghcb_sa);
1506 }
1507
1508 static void dump_ghcb(struct vcpu_svm *svm)
1509 {
1510         struct ghcb *ghcb = svm->ghcb;
1511         unsigned int nbits;
1512
1513         /* Re-use the dump_invalid_vmcb module parameter */
1514         if (!dump_invalid_vmcb) {
1515                 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
1516                 return;
1517         }
1518
1519         nbits = sizeof(ghcb->save.valid_bitmap) * 8;
1520
1521         pr_err("GHCB (GPA=%016llx):\n", svm->vmcb->control.ghcb_gpa);
1522         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
1523                ghcb->save.sw_exit_code, ghcb_sw_exit_code_is_valid(ghcb));
1524         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
1525                ghcb->save.sw_exit_info_1, ghcb_sw_exit_info_1_is_valid(ghcb));
1526         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
1527                ghcb->save.sw_exit_info_2, ghcb_sw_exit_info_2_is_valid(ghcb));
1528         pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
1529                ghcb->save.sw_scratch, ghcb_sw_scratch_is_valid(ghcb));
1530         pr_err("%-20s%*pb\n", "valid_bitmap", nbits, ghcb->save.valid_bitmap);
1531 }
1532
1533 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
1534 {
1535         struct kvm_vcpu *vcpu = &svm->vcpu;
1536         struct ghcb *ghcb = svm->ghcb;
1537
1538         /*
1539          * The GHCB protocol so far allows for the following data
1540          * to be returned:
1541          *   GPRs RAX, RBX, RCX, RDX
1542          *
1543          * Copy their values, even if they may not have been written during the
1544          * VM-Exit.  It's the guest's responsibility to not consume random data.
1545          */
1546         ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
1547         ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
1548         ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
1549         ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
1550 }
1551
1552 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
1553 {
1554         struct vmcb_control_area *control = &svm->vmcb->control;
1555         struct kvm_vcpu *vcpu = &svm->vcpu;
1556         struct ghcb *ghcb = svm->ghcb;
1557         u64 exit_code;
1558
1559         /*
1560          * The GHCB protocol so far allows for the following data
1561          * to be supplied:
1562          *   GPRs RAX, RBX, RCX, RDX
1563          *   XCR0
1564          *   CPL
1565          *
1566          * VMMCALL allows the guest to provide extra registers. KVM also
1567          * expects RSI for hypercalls, so include that, too.
1568          *
1569          * Copy their values to the appropriate location if supplied.
1570          */
1571         memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
1572
1573         vcpu->arch.regs[VCPU_REGS_RAX] = ghcb_get_rax_if_valid(ghcb);
1574         vcpu->arch.regs[VCPU_REGS_RBX] = ghcb_get_rbx_if_valid(ghcb);
1575         vcpu->arch.regs[VCPU_REGS_RCX] = ghcb_get_rcx_if_valid(ghcb);
1576         vcpu->arch.regs[VCPU_REGS_RDX] = ghcb_get_rdx_if_valid(ghcb);
1577         vcpu->arch.regs[VCPU_REGS_RSI] = ghcb_get_rsi_if_valid(ghcb);
1578
1579         svm->vmcb->save.cpl = ghcb_get_cpl_if_valid(ghcb);
1580
1581         if (ghcb_xcr0_is_valid(ghcb)) {
1582                 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
1583                 kvm_update_cpuid_runtime(vcpu);
1584         }
1585
1586         /* Copy the GHCB exit information into the VMCB fields */
1587         exit_code = ghcb_get_sw_exit_code(ghcb);
1588         control->exit_code = lower_32_bits(exit_code);
1589         control->exit_code_hi = upper_32_bits(exit_code);
1590         control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
1591         control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
1592
1593         /* Clear the valid entries fields */
1594         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
1595 }
1596
1597 static int sev_es_validate_vmgexit(struct vcpu_svm *svm)
1598 {
1599         struct kvm_vcpu *vcpu;
1600         struct ghcb *ghcb;
1601         u64 exit_code = 0;
1602
1603         ghcb = svm->ghcb;
1604
1605         /* Only GHCB Usage code 0 is supported */
1606         if (ghcb->ghcb_usage)
1607                 goto vmgexit_err;
1608
1609         /*
1610          * Retrieve the exit code now even though is may not be marked valid
1611          * as it could help with debugging.
1612          */
1613         exit_code = ghcb_get_sw_exit_code(ghcb);
1614
1615         if (!ghcb_sw_exit_code_is_valid(ghcb) ||
1616             !ghcb_sw_exit_info_1_is_valid(ghcb) ||
1617             !ghcb_sw_exit_info_2_is_valid(ghcb))
1618                 goto vmgexit_err;
1619
1620         switch (ghcb_get_sw_exit_code(ghcb)) {
1621         case SVM_EXIT_READ_DR7:
1622                 break;
1623         case SVM_EXIT_WRITE_DR7:
1624                 if (!ghcb_rax_is_valid(ghcb))
1625                         goto vmgexit_err;
1626                 break;
1627         case SVM_EXIT_RDTSC:
1628                 break;
1629         case SVM_EXIT_RDPMC:
1630                 if (!ghcb_rcx_is_valid(ghcb))
1631                         goto vmgexit_err;
1632                 break;
1633         case SVM_EXIT_CPUID:
1634                 if (!ghcb_rax_is_valid(ghcb) ||
1635                     !ghcb_rcx_is_valid(ghcb))
1636                         goto vmgexit_err;
1637                 if (ghcb_get_rax(ghcb) == 0xd)
1638                         if (!ghcb_xcr0_is_valid(ghcb))
1639                                 goto vmgexit_err;
1640                 break;
1641         case SVM_EXIT_INVD:
1642                 break;
1643         case SVM_EXIT_IOIO:
1644                 if (ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_STR_MASK) {
1645                         if (!ghcb_sw_scratch_is_valid(ghcb))
1646                                 goto vmgexit_err;
1647                 } else {
1648                         if (!(ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_TYPE_MASK))
1649                                 if (!ghcb_rax_is_valid(ghcb))
1650                                         goto vmgexit_err;
1651                 }
1652                 break;
1653         case SVM_EXIT_MSR:
1654                 if (!ghcb_rcx_is_valid(ghcb))
1655                         goto vmgexit_err;
1656                 if (ghcb_get_sw_exit_info_1(ghcb)) {
1657                         if (!ghcb_rax_is_valid(ghcb) ||
1658                             !ghcb_rdx_is_valid(ghcb))
1659                                 goto vmgexit_err;
1660                 }
1661                 break;
1662         case SVM_EXIT_VMMCALL:
1663                 if (!ghcb_rax_is_valid(ghcb) ||
1664                     !ghcb_cpl_is_valid(ghcb))
1665                         goto vmgexit_err;
1666                 break;
1667         case SVM_EXIT_RDTSCP:
1668                 break;
1669         case SVM_EXIT_WBINVD:
1670                 break;
1671         case SVM_EXIT_MONITOR:
1672                 if (!ghcb_rax_is_valid(ghcb) ||
1673                     !ghcb_rcx_is_valid(ghcb) ||
1674                     !ghcb_rdx_is_valid(ghcb))
1675                         goto vmgexit_err;
1676                 break;
1677         case SVM_EXIT_MWAIT:
1678                 if (!ghcb_rax_is_valid(ghcb) ||
1679                     !ghcb_rcx_is_valid(ghcb))
1680                         goto vmgexit_err;
1681                 break;
1682         case SVM_VMGEXIT_MMIO_READ:
1683         case SVM_VMGEXIT_MMIO_WRITE:
1684                 if (!ghcb_sw_scratch_is_valid(ghcb))
1685                         goto vmgexit_err;
1686                 break;
1687         case SVM_VMGEXIT_NMI_COMPLETE:
1688         case SVM_VMGEXIT_AP_HLT_LOOP:
1689         case SVM_VMGEXIT_AP_JUMP_TABLE:
1690         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
1691                 break;
1692         default:
1693                 goto vmgexit_err;
1694         }
1695
1696         return 0;
1697
1698 vmgexit_err:
1699         vcpu = &svm->vcpu;
1700
1701         if (ghcb->ghcb_usage) {
1702                 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
1703                             ghcb->ghcb_usage);
1704         } else {
1705                 vcpu_unimpl(vcpu, "vmgexit: exit reason %#llx is not valid\n",
1706                             exit_code);
1707                 dump_ghcb(svm);
1708         }
1709
1710         vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
1711         vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
1712         vcpu->run->internal.ndata = 2;
1713         vcpu->run->internal.data[0] = exit_code;
1714         vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
1715
1716         return -EINVAL;
1717 }
1718
1719 static void pre_sev_es_run(struct vcpu_svm *svm)
1720 {
1721         if (!svm->ghcb)
1722                 return;
1723
1724         if (svm->ghcb_sa_free) {
1725                 /*
1726                  * The scratch area lives outside the GHCB, so there is a
1727                  * buffer that, depending on the operation performed, may
1728                  * need to be synced, then freed.
1729                  */
1730                 if (svm->ghcb_sa_sync) {
1731                         kvm_write_guest(svm->vcpu.kvm,
1732                                         ghcb_get_sw_scratch(svm->ghcb),
1733                                         svm->ghcb_sa, svm->ghcb_sa_len);
1734                         svm->ghcb_sa_sync = false;
1735                 }
1736
1737                 kfree(svm->ghcb_sa);
1738                 svm->ghcb_sa = NULL;
1739                 svm->ghcb_sa_free = false;
1740         }
1741
1742         trace_kvm_vmgexit_exit(svm->vcpu.vcpu_id, svm->ghcb);
1743
1744         sev_es_sync_to_ghcb(svm);
1745
1746         kvm_vcpu_unmap(&svm->vcpu, &svm->ghcb_map, true);
1747         svm->ghcb = NULL;
1748 }
1749
1750 void pre_sev_run(struct vcpu_svm *svm, int cpu)
1751 {
1752         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1753         int asid = sev_get_asid(svm->vcpu.kvm);
1754
1755         /* Perform any SEV-ES pre-run actions */
1756         pre_sev_es_run(svm);
1757
1758         /* Assign the asid allocated with this SEV guest */
1759         svm->asid = asid;
1760
1761         /*
1762          * Flush guest TLB:
1763          *
1764          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
1765          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
1766          */
1767         if (sd->sev_vmcbs[asid] == svm->vmcb &&
1768             svm->vcpu.arch.last_vmentry_cpu == cpu)
1769                 return;
1770
1771         sd->sev_vmcbs[asid] = svm->vmcb;
1772         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
1773         vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1774 }
1775
1776 #define GHCB_SCRATCH_AREA_LIMIT         (16ULL * PAGE_SIZE)
1777 static bool setup_vmgexit_scratch(struct vcpu_svm *svm, bool sync, u64 len)
1778 {
1779         struct vmcb_control_area *control = &svm->vmcb->control;
1780         struct ghcb *ghcb = svm->ghcb;
1781         u64 ghcb_scratch_beg, ghcb_scratch_end;
1782         u64 scratch_gpa_beg, scratch_gpa_end;
1783         void *scratch_va;
1784
1785         scratch_gpa_beg = ghcb_get_sw_scratch(ghcb);
1786         if (!scratch_gpa_beg) {
1787                 pr_err("vmgexit: scratch gpa not provided\n");
1788                 return false;
1789         }
1790
1791         scratch_gpa_end = scratch_gpa_beg + len;
1792         if (scratch_gpa_end < scratch_gpa_beg) {
1793                 pr_err("vmgexit: scratch length (%#llx) not valid for scratch address (%#llx)\n",
1794                        len, scratch_gpa_beg);
1795                 return false;
1796         }
1797
1798         if ((scratch_gpa_beg & PAGE_MASK) == control->ghcb_gpa) {
1799                 /* Scratch area begins within GHCB */
1800                 ghcb_scratch_beg = control->ghcb_gpa +
1801                                    offsetof(struct ghcb, shared_buffer);
1802                 ghcb_scratch_end = control->ghcb_gpa +
1803                                    offsetof(struct ghcb, reserved_1);
1804
1805                 /*
1806                  * If the scratch area begins within the GHCB, it must be
1807                  * completely contained in the GHCB shared buffer area.
1808                  */
1809                 if (scratch_gpa_beg < ghcb_scratch_beg ||
1810                     scratch_gpa_end > ghcb_scratch_end) {
1811                         pr_err("vmgexit: scratch area is outside of GHCB shared buffer area (%#llx - %#llx)\n",
1812                                scratch_gpa_beg, scratch_gpa_end);
1813                         return false;
1814                 }
1815
1816                 scratch_va = (void *)svm->ghcb;
1817                 scratch_va += (scratch_gpa_beg - control->ghcb_gpa);
1818         } else {
1819                 /*
1820                  * The guest memory must be read into a kernel buffer, so
1821                  * limit the size
1822                  */
1823                 if (len > GHCB_SCRATCH_AREA_LIMIT) {
1824                         pr_err("vmgexit: scratch area exceeds KVM limits (%#llx requested, %#llx limit)\n",
1825                                len, GHCB_SCRATCH_AREA_LIMIT);
1826                         return false;
1827                 }
1828                 scratch_va = kzalloc(len, GFP_KERNEL);
1829                 if (!scratch_va)
1830                         return false;
1831
1832                 if (kvm_read_guest(svm->vcpu.kvm, scratch_gpa_beg, scratch_va, len)) {
1833                         /* Unable to copy scratch area from guest */
1834                         pr_err("vmgexit: kvm_read_guest for scratch area failed\n");
1835
1836                         kfree(scratch_va);
1837                         return false;
1838                 }
1839
1840                 /*
1841                  * The scratch area is outside the GHCB. The operation will
1842                  * dictate whether the buffer needs to be synced before running
1843                  * the vCPU next time (i.e. a read was requested so the data
1844                  * must be written back to the guest memory).
1845                  */
1846                 svm->ghcb_sa_sync = sync;
1847                 svm->ghcb_sa_free = true;
1848         }
1849
1850         svm->ghcb_sa = scratch_va;
1851         svm->ghcb_sa_len = len;
1852
1853         return true;
1854 }
1855
1856 static void set_ghcb_msr_bits(struct vcpu_svm *svm, u64 value, u64 mask,
1857                               unsigned int pos)
1858 {
1859         svm->vmcb->control.ghcb_gpa &= ~(mask << pos);
1860         svm->vmcb->control.ghcb_gpa |= (value & mask) << pos;
1861 }
1862
1863 static u64 get_ghcb_msr_bits(struct vcpu_svm *svm, u64 mask, unsigned int pos)
1864 {
1865         return (svm->vmcb->control.ghcb_gpa >> pos) & mask;
1866 }
1867
1868 static void set_ghcb_msr(struct vcpu_svm *svm, u64 value)
1869 {
1870         svm->vmcb->control.ghcb_gpa = value;
1871 }
1872
1873 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
1874 {
1875         struct vmcb_control_area *control = &svm->vmcb->control;
1876         struct kvm_vcpu *vcpu = &svm->vcpu;
1877         u64 ghcb_info;
1878         int ret = 1;
1879
1880         ghcb_info = control->ghcb_gpa & GHCB_MSR_INFO_MASK;
1881
1882         trace_kvm_vmgexit_msr_protocol_enter(svm->vcpu.vcpu_id,
1883                                              control->ghcb_gpa);
1884
1885         switch (ghcb_info) {
1886         case GHCB_MSR_SEV_INFO_REQ:
1887                 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
1888                                                     GHCB_VERSION_MIN,
1889                                                     sev_enc_bit));
1890                 break;
1891         case GHCB_MSR_CPUID_REQ: {
1892                 u64 cpuid_fn, cpuid_reg, cpuid_value;
1893
1894                 cpuid_fn = get_ghcb_msr_bits(svm,
1895                                              GHCB_MSR_CPUID_FUNC_MASK,
1896                                              GHCB_MSR_CPUID_FUNC_POS);
1897
1898                 /* Initialize the registers needed by the CPUID intercept */
1899                 vcpu->arch.regs[VCPU_REGS_RAX] = cpuid_fn;
1900                 vcpu->arch.regs[VCPU_REGS_RCX] = 0;
1901
1902                 ret = svm_invoke_exit_handler(svm, SVM_EXIT_CPUID);
1903                 if (!ret) {
1904                         ret = -EINVAL;
1905                         break;
1906                 }
1907
1908                 cpuid_reg = get_ghcb_msr_bits(svm,
1909                                               GHCB_MSR_CPUID_REG_MASK,
1910                                               GHCB_MSR_CPUID_REG_POS);
1911                 if (cpuid_reg == 0)
1912                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RAX];
1913                 else if (cpuid_reg == 1)
1914                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RBX];
1915                 else if (cpuid_reg == 2)
1916                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RCX];
1917                 else
1918                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RDX];
1919
1920                 set_ghcb_msr_bits(svm, cpuid_value,
1921                                   GHCB_MSR_CPUID_VALUE_MASK,
1922                                   GHCB_MSR_CPUID_VALUE_POS);
1923
1924                 set_ghcb_msr_bits(svm, GHCB_MSR_CPUID_RESP,
1925                                   GHCB_MSR_INFO_MASK,
1926                                   GHCB_MSR_INFO_POS);
1927                 break;
1928         }
1929         case GHCB_MSR_TERM_REQ: {
1930                 u64 reason_set, reason_code;
1931
1932                 reason_set = get_ghcb_msr_bits(svm,
1933                                                GHCB_MSR_TERM_REASON_SET_MASK,
1934                                                GHCB_MSR_TERM_REASON_SET_POS);
1935                 reason_code = get_ghcb_msr_bits(svm,
1936                                                 GHCB_MSR_TERM_REASON_MASK,
1937                                                 GHCB_MSR_TERM_REASON_POS);
1938                 pr_info("SEV-ES guest requested termination: %#llx:%#llx\n",
1939                         reason_set, reason_code);
1940                 fallthrough;
1941         }
1942         default:
1943                 ret = -EINVAL;
1944         }
1945
1946         trace_kvm_vmgexit_msr_protocol_exit(svm->vcpu.vcpu_id,
1947                                             control->ghcb_gpa, ret);
1948
1949         return ret;
1950 }
1951
1952 int sev_handle_vmgexit(struct vcpu_svm *svm)
1953 {
1954         struct vmcb_control_area *control = &svm->vmcb->control;
1955         u64 ghcb_gpa, exit_code;
1956         struct ghcb *ghcb;
1957         int ret;
1958
1959         /* Validate the GHCB */
1960         ghcb_gpa = control->ghcb_gpa;
1961         if (ghcb_gpa & GHCB_MSR_INFO_MASK)
1962                 return sev_handle_vmgexit_msr_protocol(svm);
1963
1964         if (!ghcb_gpa) {
1965                 vcpu_unimpl(&svm->vcpu, "vmgexit: GHCB gpa is not set\n");
1966                 return -EINVAL;
1967         }
1968
1969         if (kvm_vcpu_map(&svm->vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->ghcb_map)) {
1970                 /* Unable to map GHCB from guest */
1971                 vcpu_unimpl(&svm->vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
1972                             ghcb_gpa);
1973                 return -EINVAL;
1974         }
1975
1976         svm->ghcb = svm->ghcb_map.hva;
1977         ghcb = svm->ghcb_map.hva;
1978
1979         trace_kvm_vmgexit_enter(svm->vcpu.vcpu_id, ghcb);
1980
1981         exit_code = ghcb_get_sw_exit_code(ghcb);
1982
1983         ret = sev_es_validate_vmgexit(svm);
1984         if (ret)
1985                 return ret;
1986
1987         sev_es_sync_from_ghcb(svm);
1988         ghcb_set_sw_exit_info_1(ghcb, 0);
1989         ghcb_set_sw_exit_info_2(ghcb, 0);
1990
1991         ret = -EINVAL;
1992         switch (exit_code) {
1993         case SVM_VMGEXIT_MMIO_READ:
1994                 if (!setup_vmgexit_scratch(svm, true, control->exit_info_2))
1995                         break;
1996
1997                 ret = kvm_sev_es_mmio_read(&svm->vcpu,
1998                                            control->exit_info_1,
1999                                            control->exit_info_2,
2000                                            svm->ghcb_sa);
2001                 break;
2002         case SVM_VMGEXIT_MMIO_WRITE:
2003                 if (!setup_vmgexit_scratch(svm, false, control->exit_info_2))
2004                         break;
2005
2006                 ret = kvm_sev_es_mmio_write(&svm->vcpu,
2007                                             control->exit_info_1,
2008                                             control->exit_info_2,
2009                                             svm->ghcb_sa);
2010                 break;
2011         case SVM_VMGEXIT_NMI_COMPLETE:
2012                 ret = svm_invoke_exit_handler(svm, SVM_EXIT_IRET);
2013                 break;
2014         case SVM_VMGEXIT_AP_HLT_LOOP:
2015                 ret = kvm_emulate_ap_reset_hold(&svm->vcpu);
2016                 break;
2017         case SVM_VMGEXIT_AP_JUMP_TABLE: {
2018                 struct kvm_sev_info *sev = &to_kvm_svm(svm->vcpu.kvm)->sev_info;
2019
2020                 switch (control->exit_info_1) {
2021                 case 0:
2022                         /* Set AP jump table address */
2023                         sev->ap_jump_table = control->exit_info_2;
2024                         break;
2025                 case 1:
2026                         /* Get AP jump table address */
2027                         ghcb_set_sw_exit_info_2(ghcb, sev->ap_jump_table);
2028                         break;
2029                 default:
2030                         pr_err("svm: vmgexit: unsupported AP jump table request - exit_info_1=%#llx\n",
2031                                control->exit_info_1);
2032                         ghcb_set_sw_exit_info_1(ghcb, 1);
2033                         ghcb_set_sw_exit_info_2(ghcb,
2034                                                 X86_TRAP_UD |
2035                                                 SVM_EVTINJ_TYPE_EXEPT |
2036                                                 SVM_EVTINJ_VALID);
2037                 }
2038
2039                 ret = 1;
2040                 break;
2041         }
2042         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2043                 vcpu_unimpl(&svm->vcpu,
2044                             "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
2045                             control->exit_info_1, control->exit_info_2);
2046                 break;
2047         default:
2048                 ret = svm_invoke_exit_handler(svm, exit_code);
2049         }
2050
2051         return ret;
2052 }
2053
2054 int sev_es_string_io(struct vcpu_svm *svm, int size, unsigned int port, int in)
2055 {
2056         if (!setup_vmgexit_scratch(svm, in, svm->vmcb->control.exit_info_2))
2057                 return -EINVAL;
2058
2059         return kvm_sev_es_string_io(&svm->vcpu, size, port,
2060                                     svm->ghcb_sa, svm->ghcb_sa_len, in);
2061 }
2062
2063 void sev_es_init_vmcb(struct vcpu_svm *svm)
2064 {
2065         struct kvm_vcpu *vcpu = &svm->vcpu;
2066
2067         svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ES_ENABLE;
2068         svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
2069
2070         /*
2071          * An SEV-ES guest requires a VMSA area that is a separate from the
2072          * VMCB page. Do not include the encryption mask on the VMSA physical
2073          * address since hardware will access it using the guest key.
2074          */
2075         svm->vmcb->control.vmsa_pa = __pa(svm->vmsa);
2076
2077         /* Can't intercept CR register access, HV can't modify CR registers */
2078         svm_clr_intercept(svm, INTERCEPT_CR0_READ);
2079         svm_clr_intercept(svm, INTERCEPT_CR4_READ);
2080         svm_clr_intercept(svm, INTERCEPT_CR8_READ);
2081         svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
2082         svm_clr_intercept(svm, INTERCEPT_CR4_WRITE);
2083         svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
2084
2085         svm_clr_intercept(svm, INTERCEPT_SELECTIVE_CR0);
2086
2087         /* Track EFER/CR register changes */
2088         svm_set_intercept(svm, TRAP_EFER_WRITE);
2089         svm_set_intercept(svm, TRAP_CR0_WRITE);
2090         svm_set_intercept(svm, TRAP_CR4_WRITE);
2091         svm_set_intercept(svm, TRAP_CR8_WRITE);
2092
2093         /* No support for enable_vmware_backdoor */
2094         clr_exception_intercept(svm, GP_VECTOR);
2095
2096         /* Can't intercept XSETBV, HV can't modify XCR0 directly */
2097         svm_clr_intercept(svm, INTERCEPT_XSETBV);
2098
2099         /* Clear intercepts on selected MSRs */
2100         set_msr_interception(vcpu, svm->msrpm, MSR_EFER, 1, 1);
2101         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_CR_PAT, 1, 1);
2102         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
2103         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
2104         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
2105         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
2106 }
2107
2108 void sev_es_create_vcpu(struct vcpu_svm *svm)
2109 {
2110         /*
2111          * Set the GHCB MSR value as per the GHCB specification when creating
2112          * a vCPU for an SEV-ES guest.
2113          */
2114         set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2115                                             GHCB_VERSION_MIN,
2116                                             sev_enc_bit));
2117 }
2118
2119 void sev_es_prepare_guest_switch(struct vcpu_svm *svm, unsigned int cpu)
2120 {
2121         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
2122         struct vmcb_save_area *hostsa;
2123
2124         /*
2125          * As an SEV-ES guest, hardware will restore the host state on VMEXIT,
2126          * of which one step is to perform a VMLOAD. Since hardware does not
2127          * perform a VMSAVE on VMRUN, the host savearea must be updated.
2128          */
2129         vmsave(__sme_page_pa(sd->save_area));
2130
2131         /* XCR0 is restored on VMEXIT, save the current host value */
2132         hostsa = (struct vmcb_save_area *)(page_address(sd->save_area) + 0x400);
2133         hostsa->xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
2134
2135         /* PKRU is restored on VMEXIT, save the curent host value */
2136         hostsa->pkru = read_pkru();
2137
2138         /* MSR_IA32_XSS is restored on VMEXIT, save the currnet host value */
2139         hostsa->xss = host_xss;
2140 }
2141
2142 void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
2143 {
2144         struct vcpu_svm *svm = to_svm(vcpu);
2145
2146         /* First SIPI: Use the values as initially set by the VMM */
2147         if (!svm->received_first_sipi) {
2148                 svm->received_first_sipi = true;
2149                 return;
2150         }
2151
2152         /*
2153          * Subsequent SIPI: Return from an AP Reset Hold VMGEXIT, where
2154          * the guest will set the CS and RIP. Set SW_EXIT_INFO_2 to a
2155          * non-zero value.
2156          */
2157         ghcb_set_sw_exit_info_2(svm->ghcb, 1);
2158 }