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/pkru.h>
  23 #include <asm/trapnr.h>
  24
  25 #include "x86.h"
  26 #include "svm.h"
  27 #include "svm_ops.h"
  28 #include "cpuid.h"
  29 #include "trace.h"
  30
  31 #ifndef CONFIG_KVM_AMD_SEV
  32 /*
  33  * When this config is not defined, SEV feature is not supported and APIs in
  34  * this file are not used but this file still gets compiled into the KVM AMD
  35  * module.
  36  *
  37  * We will not have MISC_CG_RES_SEV and MISC_CG_RES_SEV_ES entries in the enum
  38  * misc_res_type {} defined in linux/misc_cgroup.h.
  39  *
  40  * Below macros allow compilation to succeed.
  41  */
  42 #define MISC_CG_RES_SEV MISC_CG_RES_TYPES
  43 #define MISC_CG_RES_SEV_ES MISC_CG_RES_TYPES
  44 #endif
  45
  46 #ifdef CONFIG_KVM_AMD_SEV
  47 /* enable/disable SEV support */
  48 static bool sev_enabled = true;
  49 module_param_named(sev, sev_enabled, bool, 0444);
  50
  51 /* enable/disable SEV-ES support */
  52 static bool sev_es_enabled = true;
  53 module_param_named(sev_es, sev_es_enabled, bool, 0444);
  54 #else
  55 #define sev_enabled false
  56 #define sev_es_enabled false
  57 #endif /* CONFIG_KVM_AMD_SEV */
  58
  59 static u8 sev_enc_bit;
  60 static DECLARE_RWSEM(sev_deactivate_lock);
  61 static DEFINE_MUTEX(sev_bitmap_lock);
  62 unsigned int max_sev_asid;
  63 static unsigned int min_sev_asid;
  64 static unsigned long sev_me_mask;
  65 static unsigned int nr_asids;
  66 static unsigned long *sev_asid_bitmap;
  67 static unsigned long *sev_reclaim_asid_bitmap;
  68
  69 struct enc_region {
  70         struct list_head list;
  71         unsigned long npages;
  72         struct page **pages;
  73         unsigned long uaddr;
  74         unsigned long size;
  75 };
  76
  77 /* Called with the sev_bitmap_lock held, or on shutdown  */
  78 static int sev_flush_asids(int min_asid, int max_asid)
  79 {
  80         int ret, asid, error = 0;
  81
  82         /* Check if there are any ASIDs to reclaim before performing a flush */
  83         asid = find_next_bit(sev_reclaim_asid_bitmap, nr_asids, min_asid);
  84         if (asid > max_asid)
  85                 return -EBUSY;
  86
  87         /*
  88          * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
  89          * so it must be guarded.
  90          */
  91         down_write(&sev_deactivate_lock);
  92
  93         wbinvd_on_all_cpus();
  94         ret = sev_guest_df_flush(&error);
  95
  96         up_write(&sev_deactivate_lock);
  97
  98         if (ret)
  99                 pr_err("SEV: DF_FLUSH failed, ret=%d, error=%#x\n", ret, error);
 100
 101         return ret;
 102 }
 103
 104 static inline bool is_mirroring_enc_context(struct kvm *kvm)
 105 {
 106         return !!to_kvm_svm(kvm)->sev_info.enc_context_owner;
 107 }
 108
 109 /* Must be called with the sev_bitmap_lock held */
 110 static bool __sev_recycle_asids(int min_asid, int max_asid)
 111 {
 112         if (sev_flush_asids(min_asid, max_asid))
 113                 return false;
 114
 115         /* The flush process will flush all reclaimable SEV and SEV-ES ASIDs */
 116         bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
 117                    nr_asids);
 118         bitmap_zero(sev_reclaim_asid_bitmap, nr_asids);
 119
 120         return true;
 121 }
 122
 123 static int sev_asid_new(struct kvm_sev_info *sev)
 124 {
 125         int asid, min_asid, max_asid, ret;
 126         bool retry = true;
 127         enum misc_res_type type;
 128
 129         type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 130         WARN_ON(sev->misc_cg);
 131         sev->misc_cg = get_current_misc_cg();
 132         ret = misc_cg_try_charge(type, sev->misc_cg, 1);
 133         if (ret) {
 134                 put_misc_cg(sev->misc_cg);
 135                 sev->misc_cg = NULL;
 136                 return ret;
 137         }
 138
 139         mutex_lock(&sev_bitmap_lock);
 140
 141         /*
 142          * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
 143          * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.
 144          */
 145         min_asid = sev->es_active ? 1 : min_sev_asid;
 146         max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
 147 again:
 148         asid = find_next_zero_bit(sev_asid_bitmap, max_asid + 1, min_asid);
 149         if (asid > max_asid) {
 150                 if (retry && __sev_recycle_asids(min_asid, max_asid)) {
 151                         retry = false;
 152                         goto again;
 153                 }
 154                 mutex_unlock(&sev_bitmap_lock);
 155                 ret = -EBUSY;
 156                 goto e_uncharge;
 157         }
 158
 159         __set_bit(asid, sev_asid_bitmap);
 160
 161         mutex_unlock(&sev_bitmap_lock);
 162
 163         return asid;
 164 e_uncharge:
 165         misc_cg_uncharge(type, sev->misc_cg, 1);
 166         put_misc_cg(sev->misc_cg);
 167         sev->misc_cg = NULL;
 168         return ret;
 169 }
 170
 171 static int sev_get_asid(struct kvm *kvm)
 172 {
 173         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 174
 175         return sev->asid;
 176 }
 177
 178 static void sev_asid_free(struct kvm_sev_info *sev)
 179 {
 180         struct svm_cpu_data *sd;
 181         int cpu;
 182         enum misc_res_type type;
 183
 184         mutex_lock(&sev_bitmap_lock);
 185
 186         __set_bit(sev->asid, sev_reclaim_asid_bitmap);
 187
 188         for_each_possible_cpu(cpu) {
 189                 sd = per_cpu(svm_data, cpu);
 190                 sd->sev_vmcbs[sev->asid] = NULL;
 191         }
 192
 193         mutex_unlock(&sev_bitmap_lock);
 194
 195         type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 196         misc_cg_uncharge(type, sev->misc_cg, 1);
 197         put_misc_cg(sev->misc_cg);
 198         sev->misc_cg = NULL;
 199 }
 200
 201 static void sev_decommission(unsigned int handle)
 202 {
 203         struct sev_data_decommission decommission;
 204
 205         if (!handle)
 206                 return;
 207
 208         decommission.handle = handle;
 209         sev_guest_decommission(&decommission, NULL);
 210 }
 211
 212 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
 213 {
 214         struct sev_data_deactivate deactivate;
 215
 216         if (!handle)
 217                 return;
 218
 219         deactivate.handle = handle;
 220
 221         /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
 222         down_read(&sev_deactivate_lock);
 223         sev_guest_deactivate(&deactivate, NULL);
 224         up_read(&sev_deactivate_lock);
 225
 226         sev_decommission(handle);
 227 }
 228
 229 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 230 {
 231         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 232         bool es_active = argp->id == KVM_SEV_ES_INIT;
 233         int asid, ret;
 234
 235         if (kvm->created_vcpus)
 236                 return -EINVAL;
 237
 238         ret = -EBUSY;
 239         if (unlikely(sev->active))
 240                 return ret;
 241
 242         sev->es_active = es_active;
 243         asid = sev_asid_new(sev);
 244         if (asid < 0)
 245                 goto e_no_asid;
 246         sev->asid = asid;
 247
 248         ret = sev_platform_init(&argp->error);
 249         if (ret)
 250                 goto e_free;
 251
 252         sev->active = true;
 253         sev->asid = asid;
 254         INIT_LIST_HEAD(&sev->regions_list);
 255
 256         return 0;
 257
 258 e_free:
 259         sev_asid_free(sev);
 260         sev->asid = 0;
 261 e_no_asid:
 262         sev->es_active = false;
 263         return ret;
 264 }
 265
 266 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
 267 {
 268         struct sev_data_activate activate;
 269         int asid = sev_get_asid(kvm);
 270         int ret;
 271
 272         /* activate ASID on the given handle */
 273         activate.handle = handle;
 274         activate.asid   = asid;
 275         ret = sev_guest_activate(&activate, error);
 276
 277         return ret;
 278 }
 279
 280 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
 281 {
 282         struct fd f;
 283         int ret;
 284
 285         f = fdget(fd);
 286         if (!f.file)
 287                 return -EBADF;
 288
 289         ret = sev_issue_cmd_external_user(f.file, id, data, error);
 290
 291         fdput(f);
 292         return ret;
 293 }
 294
 295 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
 296 {
 297         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 298
 299         return __sev_issue_cmd(sev->fd, id, data, error);
 300 }
 301
 302 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
 303 {
 304         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 305         struct sev_data_launch_start start;
 306         struct kvm_sev_launch_start params;
 307         void *dh_blob, *session_blob;
 308         int *error = &argp->error;
 309         int ret;
 310
 311         if (!sev_guest(kvm))
 312                 return -ENOTTY;
 313
 314         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 315                 return -EFAULT;
 316
 317         memset(&start, 0, sizeof(start));
 318
 319         dh_blob = NULL;
 320         if (params.dh_uaddr) {
 321                 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
 322                 if (IS_ERR(dh_blob))
 323                         return PTR_ERR(dh_blob);
 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                 sev_decommission(start.handle);
 353                 goto e_free_session;
 354         }
 355
 356         /* return handle to userspace */
 357         params.handle = start.handle;
 358         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params))) {
 359                 sev_unbind_asid(kvm, start.handle);
 360                 ret = -EFAULT;
 361                 goto e_free_session;
 362         }
 363
 364         sev->handle = start.handle;
 365         sev->fd = argp->sev_fd;
 366
 367 e_free_session:
 368         kfree(session_blob);
 369 e_free_dh:
 370         kfree(dh_blob);
 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         vaddr = params.uaddr;
 500         size = params.len;
 501         vaddr_end = vaddr + size;
 502
 503         /* Lock the user memory. */
 504         inpages = sev_pin_memory(kvm, vaddr, size, &npages, 1);
 505         if (IS_ERR(inpages))
 506                 return PTR_ERR(inpages);
 507
 508         /*
 509          * Flush (on non-coherent CPUs) before LAUNCH_UPDATE encrypts pages in
 510          * place; the cache may contain the data that was written unencrypted.
 511          */
 512         sev_clflush_pages(inpages, npages);
 513
 514         data.reserved = 0;
 515         data.handle = sev->handle;
 516
 517         for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
 518                 int offset, len;
 519
 520                 /*
 521                  * If the user buffer is not page-aligned, calculate the offset
 522                  * within the page.
 523                  */
 524                 offset = vaddr & (PAGE_SIZE - 1);
 525
 526                 /* Calculate the number of pages that can be encrypted in one go. */
 527                 pages = get_num_contig_pages(i, inpages, npages);
 528
 529                 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
 530
 531                 data.len = len;
 532                 data.address = __sme_page_pa(inpages[i]) + offset;
 533                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, &data, &argp->error);
 534                 if (ret)
 535                         goto e_unpin;
 536
 537                 size -= len;
 538                 next_vaddr = vaddr + len;
 539         }
 540
 541 e_unpin:
 542         /* content of memory is updated, mark pages dirty */
 543         for (i = 0; i < npages; i++) {
 544                 set_page_dirty_lock(inpages[i]);
 545                 mark_page_accessed(inpages[i]);
 546         }
 547         /* unlock the user pages */
 548         sev_unpin_memory(kvm, inpages, npages);
 549         return ret;
 550 }
 551
 552 static int sev_es_sync_vmsa(struct vcpu_svm *svm)
 553 {
 554         struct vmcb_save_area *save = &svm->vmcb->save;
 555
 556         /* Check some debug related fields before encrypting the VMSA */
 557         if (svm->vcpu.guest_debug || (save->dr7 & ~DR7_FIXED_1))
 558                 return -EINVAL;
 559
 560         /* Sync registgers */
 561         save->rax = svm->vcpu.arch.regs[VCPU_REGS_RAX];
 562         save->rbx = svm->vcpu.arch.regs[VCPU_REGS_RBX];
 563         save->rcx = svm->vcpu.arch.regs[VCPU_REGS_RCX];
 564         save->rdx = svm->vcpu.arch.regs[VCPU_REGS_RDX];
 565         save->rsp = svm->vcpu.arch.regs[VCPU_REGS_RSP];
 566         save->rbp = svm->vcpu.arch.regs[VCPU_REGS_RBP];
 567         save->rsi = svm->vcpu.arch.regs[VCPU_REGS_RSI];
 568         save->rdi = svm->vcpu.arch.regs[VCPU_REGS_RDI];
 569 #ifdef CONFIG_X86_64
 570         save->r8  = svm->vcpu.arch.regs[VCPU_REGS_R8];
 571         save->r9  = svm->vcpu.arch.regs[VCPU_REGS_R9];
 572         save->r10 = svm->vcpu.arch.regs[VCPU_REGS_R10];
 573         save->r11 = svm->vcpu.arch.regs[VCPU_REGS_R11];
 574         save->r12 = svm->vcpu.arch.regs[VCPU_REGS_R12];
 575         save->r13 = svm->vcpu.arch.regs[VCPU_REGS_R13];
 576         save->r14 = svm->vcpu.arch.regs[VCPU_REGS_R14];
 577         save->r15 = svm->vcpu.arch.regs[VCPU_REGS_R15];
 578 #endif
 579         save->rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
 580
 581         /* Sync some non-GPR registers before encrypting */
 582         save->xcr0 = svm->vcpu.arch.xcr0;
 583         save->pkru = svm->vcpu.arch.pkru;
 584         save->xss  = svm->vcpu.arch.ia32_xss;
 585         save->dr6  = svm->vcpu.arch.dr6;
 586
 587         /*
 588          * SEV-ES will use a VMSA that is pointed to by the VMCB, not
 589          * the traditional VMSA that is part of the VMCB. Copy the
 590          * traditional VMSA as it has been built so far (in prep
 591          * for LAUNCH_UPDATE_VMSA) to be the initial SEV-ES state.
 592          */
 593         memcpy(svm->vmsa, save, sizeof(*save));
 594
 595         return 0;
 596 }
 597
 598 static int __sev_launch_update_vmsa(struct kvm *kvm, struct kvm_vcpu *vcpu,
 599                                     int *error)
 600 {
 601         struct sev_data_launch_update_vmsa vmsa;
 602         struct vcpu_svm *svm = to_svm(vcpu);
 603         int ret;
 604
 605         /* Perform some pre-encryption checks against the VMSA */
 606         ret = sev_es_sync_vmsa(svm);
 607         if (ret)
 608                 return ret;
 609
 610         /*
 611          * The LAUNCH_UPDATE_VMSA command will perform in-place encryption of
 612          * the VMSA memory content (i.e it will write the same memory region
 613          * with the guest's key), so invalidate it first.
 614          */
 615         clflush_cache_range(svm->vmsa, PAGE_SIZE);
 616
 617         vmsa.reserved = 0;
 618         vmsa.handle = to_kvm_svm(kvm)->sev_info.handle;
 619         vmsa.address = __sme_pa(svm->vmsa);
 620         vmsa.len = PAGE_SIZE;
 621         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_VMSA, &vmsa, error);
 622         if (ret)
 623           return ret;
 624
 625         vcpu->arch.guest_state_protected = true;
 626         return 0;
 627 }
 628
 629 static int sev_launch_update_vmsa(struct kvm *kvm, struct kvm_sev_cmd *argp)
 630 {
 631         struct kvm_vcpu *vcpu;
 632         int i, ret;
 633
 634         if (!sev_es_guest(kvm))
 635                 return -ENOTTY;
 636
 637         kvm_for_each_vcpu(i, vcpu, kvm) {
 638                 ret = mutex_lock_killable(&vcpu->mutex);
 639                 if (ret)
 640                         return ret;
 641
 642                 ret = __sev_launch_update_vmsa(kvm, vcpu, &argp->error);
 643
 644                 mutex_unlock(&vcpu->mutex);
 645                 if (ret)
 646                         return ret;
 647         }
 648
 649         return 0;
 650 }
 651
 652 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 653 {
 654         void __user *measure = (void __user *)(uintptr_t)argp->data;
 655         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 656         struct sev_data_launch_measure data;
 657         struct kvm_sev_launch_measure params;
 658         void __user *p = NULL;
 659         void *blob = NULL;
 660         int ret;
 661
 662         if (!sev_guest(kvm))
 663                 return -ENOTTY;
 664
 665         if (copy_from_user(&params, measure, sizeof(params)))
 666                 return -EFAULT;
 667
 668         memset(&data, 0, sizeof(data));
 669
 670         /* User wants to query the blob length */
 671         if (!params.len)
 672                 goto cmd;
 673
 674         p = (void __user *)(uintptr_t)params.uaddr;
 675         if (p) {
 676                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
 677                         return -EINVAL;
 678
 679                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
 680                 if (!blob)
 681                         return -ENOMEM;
 682
 683                 data.address = __psp_pa(blob);
 684                 data.len = params.len;
 685         }
 686
 687 cmd:
 688         data.handle = sev->handle;
 689         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, &data, &argp->error);
 690
 691         /*
 692          * If we query the session length, FW responded with expected data.
 693          */
 694         if (!params.len)
 695                 goto done;
 696
 697         if (ret)
 698                 goto e_free_blob;
 699
 700         if (blob) {
 701                 if (copy_to_user(p, blob, params.len))
 702                         ret = -EFAULT;
 703         }
 704
 705 done:
 706         params.len = data.len;
 707         if (copy_to_user(measure, &params, sizeof(params)))
 708                 ret = -EFAULT;
 709 e_free_blob:
 710         kfree(blob);
 711         return ret;
 712 }
 713
 714 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 715 {
 716         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 717         struct sev_data_launch_finish data;
 718
 719         if (!sev_guest(kvm))
 720                 return -ENOTTY;
 721
 722         data.handle = sev->handle;
 723         return sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, &data, &argp->error);
 724 }
 725
 726 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 727 {
 728         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 729         struct kvm_sev_guest_status params;
 730         struct sev_data_guest_status data;
 731         int ret;
 732
 733         if (!sev_guest(kvm))
 734                 return -ENOTTY;
 735
 736         memset(&data, 0, sizeof(data));
 737
 738         data.handle = sev->handle;
 739         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, &data, &argp->error);
 740         if (ret)
 741                 return ret;
 742
 743         params.policy = data.policy;
 744         params.state = data.state;
 745         params.handle = data.handle;
 746
 747         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 748                 ret = -EFAULT;
 749
 750         return ret;
 751 }
 752
 753 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 754                                unsigned long dst, int size,
 755                                int *error, bool enc)
 756 {
 757         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 758         struct sev_data_dbg data;
 759
 760         data.reserved = 0;
 761         data.handle = sev->handle;
 762         data.dst_addr = dst;
 763         data.src_addr = src;
 764         data.len = size;
 765
 766         return sev_issue_cmd(kvm,
 767                              enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 768                              &data, error);
 769 }
 770
 771 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 772                              unsigned long dst_paddr, int sz, int *err)
 773 {
 774         int offset;
 775
 776         /*
 777          * Its safe to read more than we are asked, caller should ensure that
 778          * destination has enough space.
 779          */
 780         offset = src_paddr & 15;
 781         src_paddr = round_down(src_paddr, 16);
 782         sz = round_up(sz + offset, 16);
 783
 784         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 785 }
 786
 787 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 788                                   void __user *dst_uaddr,
 789                                   unsigned long dst_paddr,
 790                                   int size, int *err)
 791 {
 792         struct page *tpage = NULL;
 793         int ret, offset;
 794
 795         /* if inputs are not 16-byte then use intermediate buffer */
 796         if (!IS_ALIGNED(dst_paddr, 16) ||
 797             !IS_ALIGNED(paddr,     16) ||
 798             !IS_ALIGNED(size,      16)) {
 799                 tpage = (void *)alloc_page(GFP_KERNEL);
 800                 if (!tpage)
 801                         return -ENOMEM;
 802
 803                 dst_paddr = __sme_page_pa(tpage);
 804         }
 805
 806         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 807         if (ret)
 808                 goto e_free;
 809
 810         if (tpage) {
 811                 offset = paddr & 15;
 812                 if (copy_to_user(dst_uaddr, page_address(tpage) + offset, size))
 813                         ret = -EFAULT;
 814         }
 815
 816 e_free:
 817         if (tpage)
 818                 __free_page(tpage);
 819
 820         return ret;
 821 }
 822
 823 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 824                                   void __user *vaddr,
 825                                   unsigned long dst_paddr,
 826                                   void __user *dst_vaddr,
 827                                   int size, int *error)
 828 {
 829         struct page *src_tpage = NULL;
 830         struct page *dst_tpage = NULL;
 831         int ret, len = size;
 832
 833         /* If source buffer is not aligned then use an intermediate buffer */
 834         if (!IS_ALIGNED((unsigned long)vaddr, 16)) {
 835                 src_tpage = alloc_page(GFP_KERNEL);
 836                 if (!src_tpage)
 837                         return -ENOMEM;
 838
 839                 if (copy_from_user(page_address(src_tpage), vaddr, size)) {
 840                         __free_page(src_tpage);
 841                         return -EFAULT;
 842                 }
 843
 844                 paddr = __sme_page_pa(src_tpage);
 845         }
 846
 847         /*
 848          *  If destination buffer or length is not aligned then do read-modify-write:
 849          *   - decrypt destination in an intermediate buffer
 850          *   - copy the source buffer in an intermediate buffer
 851          *   - use the intermediate buffer as source buffer
 852          */
 853         if (!IS_ALIGNED((unsigned long)dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 854                 int dst_offset;
 855
 856                 dst_tpage = alloc_page(GFP_KERNEL);
 857                 if (!dst_tpage) {
 858                         ret = -ENOMEM;
 859                         goto e_free;
 860                 }
 861
 862                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 863                                         __sme_page_pa(dst_tpage), size, error);
 864                 if (ret)
 865                         goto e_free;
 866
 867                 /*
 868                  *  If source is kernel buffer then use memcpy() otherwise
 869                  *  copy_from_user().
 870                  */
 871                 dst_offset = dst_paddr & 15;
 872
 873                 if (src_tpage)
 874                         memcpy(page_address(dst_tpage) + dst_offset,
 875                                page_address(src_tpage), size);
 876                 else {
 877                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 878                                            vaddr, size)) {
 879                                 ret = -EFAULT;
 880                                 goto e_free;
 881                         }
 882                 }
 883
 884                 paddr = __sme_page_pa(dst_tpage);
 885                 dst_paddr = round_down(dst_paddr, 16);
 886                 len = round_up(size, 16);
 887         }
 888
 889         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 890
 891 e_free:
 892         if (src_tpage)
 893                 __free_page(src_tpage);
 894         if (dst_tpage)
 895                 __free_page(dst_tpage);
 896         return ret;
 897 }
 898
 899 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 900 {
 901         unsigned long vaddr, vaddr_end, next_vaddr;
 902         unsigned long dst_vaddr;
 903         struct page **src_p, **dst_p;
 904         struct kvm_sev_dbg debug;
 905         unsigned long n;
 906         unsigned int size;
 907         int ret;
 908
 909         if (!sev_guest(kvm))
 910                 return -ENOTTY;
 911
 912         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 913                 return -EFAULT;
 914
 915         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 916                 return -EINVAL;
 917         if (!debug.dst_uaddr)
 918                 return -EINVAL;
 919
 920         vaddr = debug.src_uaddr;
 921         size = debug.len;
 922         vaddr_end = vaddr + size;
 923         dst_vaddr = debug.dst_uaddr;
 924
 925         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 926                 int len, s_off, d_off;
 927
 928                 /* lock userspace source and destination page */
 929                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 930                 if (IS_ERR(src_p))
 931                         return PTR_ERR(src_p);
 932
 933                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 934                 if (IS_ERR(dst_p)) {
 935                         sev_unpin_memory(kvm, src_p, n);
 936                         return PTR_ERR(dst_p);
 937                 }
 938
 939                 /*
 940                  * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
 941                  * the pages; flush the destination too so that future accesses do not
 942                  * see stale data.
 943                  */
 944                 sev_clflush_pages(src_p, 1);
 945                 sev_clflush_pages(dst_p, 1);
 946
 947                 /*
 948                  * Since user buffer may not be page aligned, calculate the
 949                  * offset within the page.
 950                  */
 951                 s_off = vaddr & ~PAGE_MASK;
 952                 d_off = dst_vaddr & ~PAGE_MASK;
 953                 len = min_t(size_t, (PAGE_SIZE - s_off), size);
 954
 955                 if (dec)
 956                         ret = __sev_dbg_decrypt_user(kvm,
 957                                                      __sme_page_pa(src_p[0]) + s_off,
 958                                                      (void __user *)dst_vaddr,
 959                                                      __sme_page_pa(dst_p[0]) + d_off,
 960                                                      len, &argp->error);
 961                 else
 962                         ret = __sev_dbg_encrypt_user(kvm,
 963                                                      __sme_page_pa(src_p[0]) + s_off,
 964                                                      (void __user *)vaddr,
 965                                                      __sme_page_pa(dst_p[0]) + d_off,
 966                                                      (void __user *)dst_vaddr,
 967                                                      len, &argp->error);
 968
 969                 sev_unpin_memory(kvm, src_p, n);
 970                 sev_unpin_memory(kvm, dst_p, n);
 971
 972                 if (ret)
 973                         goto err;
 974
 975                 next_vaddr = vaddr + len;
 976                 dst_vaddr = dst_vaddr + len;
 977                 size -= len;
 978         }
 979 err:
 980         return ret;
 981 }
 982
 983 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
 984 {
 985         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 986         struct sev_data_launch_secret data;
 987         struct kvm_sev_launch_secret params;
 988         struct page **pages;
 989         void *blob, *hdr;
 990         unsigned long n, i;
 991         int ret, offset;
 992
 993         if (!sev_guest(kvm))
 994                 return -ENOTTY;
 995
 996         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 997                 return -EFAULT;
 998
 999         pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, 1);
1000         if (IS_ERR(pages))
1001                 return PTR_ERR(pages);
1002
1003         /*
1004          * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
1005          * place; the cache may contain the data that was written unencrypted.
1006          */
1007         sev_clflush_pages(pages, n);
1008
1009         /*
1010          * The secret must be copied into contiguous memory region, lets verify
1011          * that userspace memory pages are contiguous before we issue command.
1012          */
1013         if (get_num_contig_pages(0, pages, n) != n) {
1014                 ret = -EINVAL;
1015                 goto e_unpin_memory;
1016         }
1017
1018         memset(&data, 0, sizeof(data));
1019
1020         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1021         data.guest_address = __sme_page_pa(pages[0]) + offset;
1022         data.guest_len = params.guest_len;
1023
1024         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1025         if (IS_ERR(blob)) {
1026                 ret = PTR_ERR(blob);
1027                 goto e_unpin_memory;
1028         }
1029
1030         data.trans_address = __psp_pa(blob);
1031         data.trans_len = params.trans_len;
1032
1033         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1034         if (IS_ERR(hdr)) {
1035                 ret = PTR_ERR(hdr);
1036                 goto e_free_blob;
1037         }
1038         data.hdr_address = __psp_pa(hdr);
1039         data.hdr_len = params.hdr_len;
1040
1041         data.handle = sev->handle;
1042         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, &data, &argp->error);
1043
1044         kfree(hdr);
1045
1046 e_free_blob:
1047         kfree(blob);
1048 e_unpin_memory:
1049         /* content of memory is updated, mark pages dirty */
1050         for (i = 0; i < n; i++) {
1051                 set_page_dirty_lock(pages[i]);
1052                 mark_page_accessed(pages[i]);
1053         }
1054         sev_unpin_memory(kvm, pages, n);
1055         return ret;
1056 }
1057
1058 static int sev_get_attestation_report(struct kvm *kvm, struct kvm_sev_cmd *argp)
1059 {
1060         void __user *report = (void __user *)(uintptr_t)argp->data;
1061         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1062         struct sev_data_attestation_report data;
1063         struct kvm_sev_attestation_report params;
1064         void __user *p;
1065         void *blob = NULL;
1066         int ret;
1067
1068         if (!sev_guest(kvm))
1069                 return -ENOTTY;
1070
1071         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1072                 return -EFAULT;
1073
1074         memset(&data, 0, sizeof(data));
1075
1076         /* User wants to query the blob length */
1077         if (!params.len)
1078                 goto cmd;
1079
1080         p = (void __user *)(uintptr_t)params.uaddr;
1081         if (p) {
1082                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
1083                         return -EINVAL;
1084
1085                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
1086                 if (!blob)
1087                         return -ENOMEM;
1088
1089                 data.address = __psp_pa(blob);
1090                 data.len = params.len;
1091                 memcpy(data.mnonce, params.mnonce, sizeof(params.mnonce));
1092         }
1093 cmd:
1094         data.handle = sev->handle;
1095         ret = sev_issue_cmd(kvm, SEV_CMD_ATTESTATION_REPORT, &data, &argp->error);
1096         /*
1097          * If we query the session length, FW responded with expected data.
1098          */
1099         if (!params.len)
1100                 goto done;
1101
1102         if (ret)
1103                 goto e_free_blob;
1104
1105         if (blob) {
1106                 if (copy_to_user(p, blob, params.len))
1107                         ret = -EFAULT;
1108         }
1109
1110 done:
1111         params.len = data.len;
1112         if (copy_to_user(report, &params, sizeof(params)))
1113                 ret = -EFAULT;
1114 e_free_blob:
1115         kfree(blob);
1116         return ret;
1117 }
1118
1119 /* Userspace wants to query session length. */
1120 static int
1121 __sev_send_start_query_session_length(struct kvm *kvm, struct kvm_sev_cmd *argp,
1122                                       struct kvm_sev_send_start *params)
1123 {
1124         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1125         struct sev_data_send_start data;
1126         int ret;
1127
1128         memset(&data, 0, sizeof(data));
1129         data.handle = sev->handle;
1130         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1131
1132         params->session_len = data.session_len;
1133         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1134                                 sizeof(struct kvm_sev_send_start)))
1135                 ret = -EFAULT;
1136
1137         return ret;
1138 }
1139
1140 static int sev_send_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1141 {
1142         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1143         struct sev_data_send_start data;
1144         struct kvm_sev_send_start params;
1145         void *amd_certs, *session_data;
1146         void *pdh_cert, *plat_certs;
1147         int ret;
1148
1149         if (!sev_guest(kvm))
1150                 return -ENOTTY;
1151
1152         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1153                                 sizeof(struct kvm_sev_send_start)))
1154                 return -EFAULT;
1155
1156         /* if session_len is zero, userspace wants to query the session length */
1157         if (!params.session_len)
1158                 return __sev_send_start_query_session_length(kvm, argp,
1159                                 &params);
1160
1161         /* some sanity checks */
1162         if (!params.pdh_cert_uaddr || !params.pdh_cert_len ||
1163             !params.session_uaddr || params.session_len > SEV_FW_BLOB_MAX_SIZE)
1164                 return -EINVAL;
1165
1166         /* allocate the memory to hold the session data blob */
1167         session_data = kmalloc(params.session_len, GFP_KERNEL_ACCOUNT);
1168         if (!session_data)
1169                 return -ENOMEM;
1170
1171         /* copy the certificate blobs from userspace */
1172         pdh_cert = psp_copy_user_blob(params.pdh_cert_uaddr,
1173                                 params.pdh_cert_len);
1174         if (IS_ERR(pdh_cert)) {
1175                 ret = PTR_ERR(pdh_cert);
1176                 goto e_free_session;
1177         }
1178
1179         plat_certs = psp_copy_user_blob(params.plat_certs_uaddr,
1180                                 params.plat_certs_len);
1181         if (IS_ERR(plat_certs)) {
1182                 ret = PTR_ERR(plat_certs);
1183                 goto e_free_pdh;
1184         }
1185
1186         amd_certs = psp_copy_user_blob(params.amd_certs_uaddr,
1187                                 params.amd_certs_len);
1188         if (IS_ERR(amd_certs)) {
1189                 ret = PTR_ERR(amd_certs);
1190                 goto e_free_plat_cert;
1191         }
1192
1193         /* populate the FW SEND_START field with system physical address */
1194         memset(&data, 0, sizeof(data));
1195         data.pdh_cert_address = __psp_pa(pdh_cert);
1196         data.pdh_cert_len = params.pdh_cert_len;
1197         data.plat_certs_address = __psp_pa(plat_certs);
1198         data.plat_certs_len = params.plat_certs_len;
1199         data.amd_certs_address = __psp_pa(amd_certs);
1200         data.amd_certs_len = params.amd_certs_len;
1201         data.session_address = __psp_pa(session_data);
1202         data.session_len = params.session_len;
1203         data.handle = sev->handle;
1204
1205         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1206
1207         if (!ret && copy_to_user((void __user *)(uintptr_t)params.session_uaddr,
1208                         session_data, params.session_len)) {
1209                 ret = -EFAULT;
1210                 goto e_free_amd_cert;
1211         }
1212
1213         params.policy = data.policy;
1214         params.session_len = data.session_len;
1215         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params,
1216                                 sizeof(struct kvm_sev_send_start)))
1217                 ret = -EFAULT;
1218
1219 e_free_amd_cert:
1220         kfree(amd_certs);
1221 e_free_plat_cert:
1222         kfree(plat_certs);
1223 e_free_pdh:
1224         kfree(pdh_cert);
1225 e_free_session:
1226         kfree(session_data);
1227         return ret;
1228 }
1229
1230 /* Userspace wants to query either header or trans length. */
1231 static int
1232 __sev_send_update_data_query_lengths(struct kvm *kvm, struct kvm_sev_cmd *argp,
1233                                      struct kvm_sev_send_update_data *params)
1234 {
1235         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1236         struct sev_data_send_update_data data;
1237         int ret;
1238
1239         memset(&data, 0, sizeof(data));
1240         data.handle = sev->handle;
1241         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1242
1243         params->hdr_len = data.hdr_len;
1244         params->trans_len = data.trans_len;
1245
1246         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1247                          sizeof(struct kvm_sev_send_update_data)))
1248                 ret = -EFAULT;
1249
1250         return ret;
1251 }
1252
1253 static int sev_send_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1254 {
1255         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1256         struct sev_data_send_update_data data;
1257         struct kvm_sev_send_update_data params;
1258         void *hdr, *trans_data;
1259         struct page **guest_page;
1260         unsigned long n;
1261         int ret, offset;
1262
1263         if (!sev_guest(kvm))
1264                 return -ENOTTY;
1265
1266         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1267                         sizeof(struct kvm_sev_send_update_data)))
1268                 return -EFAULT;
1269
1270         /* userspace wants to query either header or trans length */
1271         if (!params.trans_len || !params.hdr_len)
1272                 return __sev_send_update_data_query_lengths(kvm, argp, &params);
1273
1274         if (!params.trans_uaddr || !params.guest_uaddr ||
1275             !params.guest_len || !params.hdr_uaddr)
1276                 return -EINVAL;
1277
1278         /* Check if we are crossing the page boundary */
1279         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1280         if ((params.guest_len + offset > PAGE_SIZE))
1281                 return -EINVAL;
1282
1283         /* Pin guest memory */
1284         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1285                                     PAGE_SIZE, &n, 0);
1286         if (IS_ERR(guest_page))
1287                 return PTR_ERR(guest_page);
1288
1289         /* allocate memory for header and transport buffer */
1290         ret = -ENOMEM;
1291         hdr = kmalloc(params.hdr_len, GFP_KERNEL_ACCOUNT);
1292         if (!hdr)
1293                 goto e_unpin;
1294
1295         trans_data = kmalloc(params.trans_len, GFP_KERNEL_ACCOUNT);
1296         if (!trans_data)
1297                 goto e_free_hdr;
1298
1299         memset(&data, 0, sizeof(data));
1300         data.hdr_address = __psp_pa(hdr);
1301         data.hdr_len = params.hdr_len;
1302         data.trans_address = __psp_pa(trans_data);
1303         data.trans_len = params.trans_len;
1304
1305         /* The SEND_UPDATE_DATA command requires C-bit to be always set. */
1306         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1307         data.guest_address |= sev_me_mask;
1308         data.guest_len = params.guest_len;
1309         data.handle = sev->handle;
1310
1311         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1312
1313         if (ret)
1314                 goto e_free_trans_data;
1315
1316         /* copy transport buffer to user space */
1317         if (copy_to_user((void __user *)(uintptr_t)params.trans_uaddr,
1318                          trans_data, params.trans_len)) {
1319                 ret = -EFAULT;
1320                 goto e_free_trans_data;
1321         }
1322
1323         /* Copy packet header to userspace. */
1324         if (copy_to_user((void __user *)(uintptr_t)params.hdr_uaddr, hdr,
1325                          params.hdr_len))
1326                 ret = -EFAULT;
1327
1328 e_free_trans_data:
1329         kfree(trans_data);
1330 e_free_hdr:
1331         kfree(hdr);
1332 e_unpin:
1333         sev_unpin_memory(kvm, guest_page, n);
1334
1335         return ret;
1336 }
1337
1338 static int sev_send_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1339 {
1340         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1341         struct sev_data_send_finish data;
1342
1343         if (!sev_guest(kvm))
1344                 return -ENOTTY;
1345
1346         data.handle = sev->handle;
1347         return sev_issue_cmd(kvm, SEV_CMD_SEND_FINISH, &data, &argp->error);
1348 }
1349
1350 static int sev_send_cancel(struct kvm *kvm, struct kvm_sev_cmd *argp)
1351 {
1352         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1353         struct sev_data_send_cancel data;
1354
1355         if (!sev_guest(kvm))
1356                 return -ENOTTY;
1357
1358         data.handle = sev->handle;
1359         return sev_issue_cmd(kvm, SEV_CMD_SEND_CANCEL, &data, &argp->error);
1360 }
1361
1362 static int sev_receive_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1363 {
1364         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1365         struct sev_data_receive_start start;
1366         struct kvm_sev_receive_start params;
1367         int *error = &argp->error;
1368         void *session_data;
1369         void *pdh_data;
1370         int ret;
1371
1372         if (!sev_guest(kvm))
1373                 return -ENOTTY;
1374
1375         /* Get parameter from the userspace */
1376         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1377                         sizeof(struct kvm_sev_receive_start)))
1378                 return -EFAULT;
1379
1380         /* some sanity checks */
1381         if (!params.pdh_uaddr || !params.pdh_len ||
1382             !params.session_uaddr || !params.session_len)
1383                 return -EINVAL;
1384
1385         pdh_data = psp_copy_user_blob(params.pdh_uaddr, params.pdh_len);
1386         if (IS_ERR(pdh_data))
1387                 return PTR_ERR(pdh_data);
1388
1389         session_data = psp_copy_user_blob(params.session_uaddr,
1390                         params.session_len);
1391         if (IS_ERR(session_data)) {
1392                 ret = PTR_ERR(session_data);
1393                 goto e_free_pdh;
1394         }
1395
1396         memset(&start, 0, sizeof(start));
1397         start.handle = params.handle;
1398         start.policy = params.policy;
1399         start.pdh_cert_address = __psp_pa(pdh_data);
1400         start.pdh_cert_len = params.pdh_len;
1401         start.session_address = __psp_pa(session_data);
1402         start.session_len = params.session_len;
1403
1404         /* create memory encryption context */
1405         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_RECEIVE_START, &start,
1406                                 error);
1407         if (ret)
1408                 goto e_free_session;
1409
1410         /* Bind ASID to this guest */
1411         ret = sev_bind_asid(kvm, start.handle, error);
1412         if (ret) {
1413                 sev_decommission(start.handle);
1414                 goto e_free_session;
1415         }
1416
1417         params.handle = start.handle;
1418         if (copy_to_user((void __user *)(uintptr_t)argp->data,
1419                          &params, sizeof(struct kvm_sev_receive_start))) {
1420                 ret = -EFAULT;
1421                 sev_unbind_asid(kvm, start.handle);
1422                 goto e_free_session;
1423         }
1424
1425         sev->handle = start.handle;
1426         sev->fd = argp->sev_fd;
1427
1428 e_free_session:
1429         kfree(session_data);
1430 e_free_pdh:
1431         kfree(pdh_data);
1432
1433         return ret;
1434 }
1435
1436 static int sev_receive_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1437 {
1438         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1439         struct kvm_sev_receive_update_data params;
1440         struct sev_data_receive_update_data data;
1441         void *hdr = NULL, *trans = NULL;
1442         struct page **guest_page;
1443         unsigned long n;
1444         int ret, offset;
1445
1446         if (!sev_guest(kvm))
1447                 return -EINVAL;
1448
1449         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1450                         sizeof(struct kvm_sev_receive_update_data)))
1451                 return -EFAULT;
1452
1453         if (!params.hdr_uaddr || !params.hdr_len ||
1454             !params.guest_uaddr || !params.guest_len ||
1455             !params.trans_uaddr || !params.trans_len)
1456                 return -EINVAL;
1457
1458         /* Check if we are crossing the page boundary */
1459         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1460         if ((params.guest_len + offset > PAGE_SIZE))
1461                 return -EINVAL;
1462
1463         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1464         if (IS_ERR(hdr))
1465                 return PTR_ERR(hdr);
1466
1467         trans = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1468         if (IS_ERR(trans)) {
1469                 ret = PTR_ERR(trans);
1470                 goto e_free_hdr;
1471         }
1472
1473         memset(&data, 0, sizeof(data));
1474         data.hdr_address = __psp_pa(hdr);
1475         data.hdr_len = params.hdr_len;
1476         data.trans_address = __psp_pa(trans);
1477         data.trans_len = params.trans_len;
1478
1479         /* Pin guest memory */
1480         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1481                                     PAGE_SIZE, &n, 1);
1482         if (IS_ERR(guest_page)) {
1483                 ret = PTR_ERR(guest_page);
1484                 goto e_free_trans;
1485         }
1486
1487         /*
1488          * Flush (on non-coherent CPUs) before RECEIVE_UPDATE_DATA, the PSP
1489          * encrypts the written data with the guest's key, and the cache may
1490          * contain dirty, unencrypted data.
1491          */
1492         sev_clflush_pages(guest_page, n);
1493
1494         /* The RECEIVE_UPDATE_DATA command requires C-bit to be always set. */
1495         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1496         data.guest_address |= sev_me_mask;
1497         data.guest_len = params.guest_len;
1498         data.handle = sev->handle;
1499
1500         ret = sev_issue_cmd(kvm, SEV_CMD_RECEIVE_UPDATE_DATA, &data,
1501                                 &argp->error);
1502
1503         sev_unpin_memory(kvm, guest_page, n);
1504
1505 e_free_trans:
1506         kfree(trans);
1507 e_free_hdr:
1508         kfree(hdr);
1509
1510         return ret;
1511 }
1512
1513 static int sev_receive_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1514 {
1515         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1516         struct sev_data_receive_finish data;
1517
1518         if (!sev_guest(kvm))
1519                 return -ENOTTY;
1520
1521         data.handle = sev->handle;
1522         return sev_issue_cmd(kvm, SEV_CMD_RECEIVE_FINISH, &data, &argp->error);
1523 }
1524
1525 static bool cmd_allowed_from_miror(u32 cmd_id)
1526 {
1527         /*
1528          * Allow mirrors VM to call KVM_SEV_LAUNCH_UPDATE_VMSA to enable SEV-ES
1529          * active mirror VMs. Also allow the debugging and status commands.
1530          */
1531         if (cmd_id == KVM_SEV_LAUNCH_UPDATE_VMSA ||
1532             cmd_id == KVM_SEV_GUEST_STATUS || cmd_id == KVM_SEV_DBG_DECRYPT ||
1533             cmd_id == KVM_SEV_DBG_ENCRYPT)
1534                 return true;
1535
1536         return false;
1537 }
1538
1539 int svm_mem_enc_op(struct kvm *kvm, void __user *argp)
1540 {
1541         struct kvm_sev_cmd sev_cmd;
1542         int r;
1543
1544         if (!sev_enabled)
1545                 return -ENOTTY;
1546
1547         if (!argp)
1548                 return 0;
1549
1550         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
1551                 return -EFAULT;
1552
1553         mutex_lock(&kvm->lock);
1554
1555         /* Only the enc_context_owner handles some memory enc operations. */
1556         if (is_mirroring_enc_context(kvm) &&
1557             !cmd_allowed_from_miror(sev_cmd.id)) {
1558                 r = -EINVAL;
1559                 goto out;
1560         }
1561
1562         switch (sev_cmd.id) {
1563         case KVM_SEV_ES_INIT:
1564                 if (!sev_es_enabled) {
1565                         r = -ENOTTY;
1566                         goto out;
1567                 }
1568                 fallthrough;
1569         case KVM_SEV_INIT:
1570                 r = sev_guest_init(kvm, &sev_cmd);
1571                 break;
1572         case KVM_SEV_LAUNCH_START:
1573                 r = sev_launch_start(kvm, &sev_cmd);
1574                 break;
1575         case KVM_SEV_LAUNCH_UPDATE_DATA:
1576                 r = sev_launch_update_data(kvm, &sev_cmd);
1577                 break;
1578         case KVM_SEV_LAUNCH_UPDATE_VMSA:
1579                 r = sev_launch_update_vmsa(kvm, &sev_cmd);
1580                 break;
1581         case KVM_SEV_LAUNCH_MEASURE:
1582                 r = sev_launch_measure(kvm, &sev_cmd);
1583                 break;
1584         case KVM_SEV_LAUNCH_FINISH:
1585                 r = sev_launch_finish(kvm, &sev_cmd);
1586                 break;
1587         case KVM_SEV_GUEST_STATUS:
1588                 r = sev_guest_status(kvm, &sev_cmd);
1589                 break;
1590         case KVM_SEV_DBG_DECRYPT:
1591                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
1592                 break;
1593         case KVM_SEV_DBG_ENCRYPT:
1594                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
1595                 break;
1596         case KVM_SEV_LAUNCH_SECRET:
1597                 r = sev_launch_secret(kvm, &sev_cmd);
1598                 break;
1599         case KVM_SEV_GET_ATTESTATION_REPORT:
1600                 r = sev_get_attestation_report(kvm, &sev_cmd);
1601                 break;
1602         case KVM_SEV_SEND_START:
1603                 r = sev_send_start(kvm, &sev_cmd);
1604                 break;
1605         case KVM_SEV_SEND_UPDATE_DATA:
1606                 r = sev_send_update_data(kvm, &sev_cmd);
1607                 break;
1608         case KVM_SEV_SEND_FINISH:
1609                 r = sev_send_finish(kvm, &sev_cmd);
1610                 break;
1611         case KVM_SEV_SEND_CANCEL:
1612                 r = sev_send_cancel(kvm, &sev_cmd);
1613                 break;
1614         case KVM_SEV_RECEIVE_START:
1615                 r = sev_receive_start(kvm, &sev_cmd);
1616                 break;
1617         case KVM_SEV_RECEIVE_UPDATE_DATA:
1618                 r = sev_receive_update_data(kvm, &sev_cmd);
1619                 break;
1620         case KVM_SEV_RECEIVE_FINISH:
1621                 r = sev_receive_finish(kvm, &sev_cmd);
1622                 break;
1623         default:
1624                 r = -EINVAL;
1625                 goto out;
1626         }
1627
1628         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
1629                 r = -EFAULT;
1630
1631 out:
1632         mutex_unlock(&kvm->lock);
1633         return r;
1634 }
1635
1636 int svm_register_enc_region(struct kvm *kvm,
1637                             struct kvm_enc_region *range)
1638 {
1639         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1640         struct enc_region *region;
1641         int ret = 0;
1642
1643         if (!sev_guest(kvm))
1644                 return -ENOTTY;
1645
1646         /* If kvm is mirroring encryption context it isn't responsible for it */
1647         if (is_mirroring_enc_context(kvm))
1648                 return -EINVAL;
1649
1650         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
1651                 return -EINVAL;
1652
1653         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
1654         if (!region)
1655                 return -ENOMEM;
1656
1657         mutex_lock(&kvm->lock);
1658         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
1659         if (IS_ERR(region->pages)) {
1660                 ret = PTR_ERR(region->pages);
1661                 mutex_unlock(&kvm->lock);
1662                 goto e_free;
1663         }
1664
1665         region->uaddr = range->addr;
1666         region->size = range->size;
1667
1668         list_add_tail(&region->list, &sev->regions_list);
1669         mutex_unlock(&kvm->lock);
1670
1671         /*
1672          * The guest may change the memory encryption attribute from C=0 -> C=1
1673          * or vice versa for this memory range. Lets make sure caches are
1674          * flushed to ensure that guest data gets written into memory with
1675          * correct C-bit.
1676          */
1677         sev_clflush_pages(region->pages, region->npages);
1678
1679         return ret;
1680
1681 e_free:
1682         kfree(region);
1683         return ret;
1684 }
1685
1686 static struct enc_region *
1687 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
1688 {
1689         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1690         struct list_head *head = &sev->regions_list;
1691         struct enc_region *i;
1692
1693         list_for_each_entry(i, head, list) {
1694                 if (i->uaddr == range->addr &&
1695                     i->size == range->size)
1696                         return i;
1697         }
1698
1699         return NULL;
1700 }
1701
1702 static void __unregister_enc_region_locked(struct kvm *kvm,
1703                                            struct enc_region *region)
1704 {
1705         sev_unpin_memory(kvm, region->pages, region->npages);
1706         list_del(&region->list);
1707         kfree(region);
1708 }
1709
1710 int svm_unregister_enc_region(struct kvm *kvm,
1711                               struct kvm_enc_region *range)
1712 {
1713         struct enc_region *region;
1714         int ret;
1715
1716         /* If kvm is mirroring encryption context it isn't responsible for it */
1717         if (is_mirroring_enc_context(kvm))
1718                 return -EINVAL;
1719
1720         mutex_lock(&kvm->lock);
1721
1722         if (!sev_guest(kvm)) {
1723                 ret = -ENOTTY;
1724                 goto failed;
1725         }
1726
1727         region = find_enc_region(kvm, range);
1728         if (!region) {
1729                 ret = -EINVAL;
1730                 goto failed;
1731         }
1732
1733         /*
1734          * Ensure that all guest tagged cache entries are flushed before
1735          * releasing the pages back to the system for use. CLFLUSH will
1736          * not do this, so issue a WBINVD.
1737          */
1738         wbinvd_on_all_cpus();
1739
1740         __unregister_enc_region_locked(kvm, region);
1741
1742         mutex_unlock(&kvm->lock);
1743         return 0;
1744
1745 failed:
1746         mutex_unlock(&kvm->lock);
1747         return ret;
1748 }
1749
1750 int svm_vm_copy_asid_from(struct kvm *kvm, unsigned int source_fd)
1751 {
1752         struct file *source_kvm_file;
1753         struct kvm *source_kvm;
1754         struct kvm_sev_info source_sev, *mirror_sev;
1755         int ret;
1756
1757         source_kvm_file = fget(source_fd);
1758         if (!file_is_kvm(source_kvm_file)) {
1759                 ret = -EBADF;
1760                 goto e_source_put;
1761         }
1762
1763         source_kvm = source_kvm_file->private_data;
1764         mutex_lock(&source_kvm->lock);
1765
1766         if (!sev_guest(source_kvm)) {
1767                 ret = -EINVAL;
1768                 goto e_source_unlock;
1769         }
1770
1771         /* Mirrors of mirrors should work, but let's not get silly */
1772         if (is_mirroring_enc_context(source_kvm) || source_kvm == kvm) {
1773                 ret = -EINVAL;
1774                 goto e_source_unlock;
1775         }
1776
1777         memcpy(&source_sev, &to_kvm_svm(source_kvm)->sev_info,
1778                sizeof(source_sev));
1779
1780         /*
1781          * The mirror kvm holds an enc_context_owner ref so its asid can't
1782          * disappear until we're done with it
1783          */
1784         kvm_get_kvm(source_kvm);
1785
1786         fput(source_kvm_file);
1787         mutex_unlock(&source_kvm->lock);
1788         mutex_lock(&kvm->lock);
1789
1790         if (sev_guest(kvm)) {
1791                 ret = -EINVAL;
1792                 goto e_mirror_unlock;
1793         }
1794
1795         /* Set enc_context_owner and copy its encryption context over */
1796         mirror_sev = &to_kvm_svm(kvm)->sev_info;
1797         mirror_sev->enc_context_owner = source_kvm;
1798         mirror_sev->active = true;
1799         mirror_sev->asid = source_sev.asid;
1800         mirror_sev->fd = source_sev.fd;
1801         mirror_sev->es_active = source_sev.es_active;
1802         mirror_sev->handle = source_sev.handle;
1803         /*
1804          * Do not copy ap_jump_table. Since the mirror does not share the same
1805          * KVM contexts as the original, and they may have different
1806          * memory-views.
1807          */
1808
1809         mutex_unlock(&kvm->lock);
1810         return 0;
1811
1812 e_mirror_unlock:
1813         mutex_unlock(&kvm->lock);
1814         kvm_put_kvm(source_kvm);
1815         return ret;
1816 e_source_unlock:
1817         mutex_unlock(&source_kvm->lock);
1818 e_source_put:
1819         if (source_kvm_file)
1820                 fput(source_kvm_file);
1821         return ret;
1822 }
1823
1824 void sev_vm_destroy(struct kvm *kvm)
1825 {
1826         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1827         struct list_head *head = &sev->regions_list;
1828         struct list_head *pos, *q;
1829
1830         if (!sev_guest(kvm))
1831                 return;
1832
1833         /* If this is a mirror_kvm release the enc_context_owner and skip sev cleanup */
1834         if (is_mirroring_enc_context(kvm)) {
1835                 kvm_put_kvm(sev->enc_context_owner);
1836                 return;
1837         }
1838
1839         mutex_lock(&kvm->lock);
1840
1841         /*
1842          * Ensure that all guest tagged cache entries are flushed before
1843          * releasing the pages back to the system for use. CLFLUSH will
1844          * not do this, so issue a WBINVD.
1845          */
1846         wbinvd_on_all_cpus();
1847
1848         /*
1849          * if userspace was terminated before unregistering the memory regions
1850          * then lets unpin all the registered memory.
1851          */
1852         if (!list_empty(head)) {
1853                 list_for_each_safe(pos, q, head) {
1854                         __unregister_enc_region_locked(kvm,
1855                                 list_entry(pos, struct enc_region, list));
1856                         cond_resched();
1857                 }
1858         }
1859
1860         mutex_unlock(&kvm->lock);
1861
1862         sev_unbind_asid(kvm, sev->handle);
1863         sev_asid_free(sev);
1864 }
1865
1866 void __init sev_set_cpu_caps(void)
1867 {
1868         if (!sev_enabled)
1869                 kvm_cpu_cap_clear(X86_FEATURE_SEV);
1870         if (!sev_es_enabled)
1871                 kvm_cpu_cap_clear(X86_FEATURE_SEV_ES);
1872 }
1873
1874 void __init sev_hardware_setup(void)
1875 {
1876 #ifdef CONFIG_KVM_AMD_SEV
1877         unsigned int eax, ebx, ecx, edx, sev_asid_count, sev_es_asid_count;
1878         bool sev_es_supported = false;
1879         bool sev_supported = false;
1880
1881         if (!sev_enabled || !npt_enabled)
1882                 goto out;
1883
1884         /* Does the CPU support SEV? */
1885         if (!boot_cpu_has(X86_FEATURE_SEV))
1886                 goto out;
1887
1888         /* Retrieve SEV CPUID information */
1889         cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
1890
1891         /* Set encryption bit location for SEV-ES guests */
1892         sev_enc_bit = ebx & 0x3f;
1893
1894         /* Maximum number of encrypted guests supported simultaneously */
1895         max_sev_asid = ecx;
1896         if (!max_sev_asid)
1897                 goto out;
1898
1899         /* Minimum ASID value that should be used for SEV guest */
1900         min_sev_asid = edx;
1901         sev_me_mask = 1UL << (ebx & 0x3f);
1902
1903         /*
1904          * Initialize SEV ASID bitmaps. Allocate space for ASID 0 in the bitmap,
1905          * even though it's never used, so that the bitmap is indexed by the
1906          * actual ASID.
1907          */
1908         nr_asids = max_sev_asid + 1;
1909         sev_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
1910         if (!sev_asid_bitmap)
1911                 goto out;
1912
1913         sev_reclaim_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
1914         if (!sev_reclaim_asid_bitmap) {
1915                 bitmap_free(sev_asid_bitmap);
1916                 sev_asid_bitmap = NULL;
1917                 goto out;
1918         }
1919
1920         sev_asid_count = max_sev_asid - min_sev_asid + 1;
1921         if (misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count))
1922                 goto out;
1923
1924         pr_info("SEV supported: %u ASIDs\n", sev_asid_count);
1925         sev_supported = true;
1926
1927         /* SEV-ES support requested? */
1928         if (!sev_es_enabled)
1929                 goto out;
1930
1931         /* Does the CPU support SEV-ES? */
1932         if (!boot_cpu_has(X86_FEATURE_SEV_ES))
1933                 goto out;
1934
1935         /* Has the system been allocated ASIDs for SEV-ES? */
1936         if (min_sev_asid == 1)
1937                 goto out;
1938
1939         sev_es_asid_count = min_sev_asid - 1;
1940         if (misc_cg_set_capacity(MISC_CG_RES_SEV_ES, sev_es_asid_count))
1941                 goto out;
1942
1943         pr_info("SEV-ES supported: %u ASIDs\n", sev_es_asid_count);
1944         sev_es_supported = true;
1945
1946 out:
1947         sev_enabled = sev_supported;
1948         sev_es_enabled = sev_es_supported;
1949 #endif
1950 }
1951
1952 void sev_hardware_teardown(void)
1953 {
1954         if (!sev_enabled)
1955                 return;
1956
1957         /* No need to take sev_bitmap_lock, all VMs have been destroyed. */
1958         sev_flush_asids(1, max_sev_asid);
1959
1960         bitmap_free(sev_asid_bitmap);
1961         bitmap_free(sev_reclaim_asid_bitmap);
1962
1963         misc_cg_set_capacity(MISC_CG_RES_SEV, 0);
1964         misc_cg_set_capacity(MISC_CG_RES_SEV_ES, 0);
1965 }
1966
1967 int sev_cpu_init(struct svm_cpu_data *sd)
1968 {
1969         if (!sev_enabled)
1970                 return 0;
1971
1972         sd->sev_vmcbs = kcalloc(nr_asids, sizeof(void *), GFP_KERNEL);
1973         if (!sd->sev_vmcbs)
1974                 return -ENOMEM;
1975
1976         return 0;
1977 }
1978
1979 /*
1980  * Pages used by hardware to hold guest encrypted state must be flushed before
1981  * returning them to the system.
1982  */
1983 static void sev_flush_guest_memory(struct vcpu_svm *svm, void *va,
1984                                    unsigned long len)
1985 {
1986         /*
1987          * If hardware enforced cache coherency for encrypted mappings of the
1988          * same physical page is supported, nothing to do.
1989          */
1990         if (boot_cpu_has(X86_FEATURE_SME_COHERENT))
1991                 return;
1992
1993         /*
1994          * If the VM Page Flush MSR is supported, use it to flush the page
1995          * (using the page virtual address and the guest ASID).
1996          */
1997         if (boot_cpu_has(X86_FEATURE_VM_PAGE_FLUSH)) {
1998                 struct kvm_sev_info *sev;
1999                 unsigned long va_start;
2000                 u64 start, stop;
2001
2002                 /* Align start and stop to page boundaries. */
2003                 va_start = (unsigned long)va;
2004                 start = (u64)va_start & PAGE_MASK;
2005                 stop = PAGE_ALIGN((u64)va_start + len);
2006
2007                 if (start < stop) {
2008                         sev = &to_kvm_svm(svm->vcpu.kvm)->sev_info;
2009
2010                         while (start < stop) {
2011                                 wrmsrl(MSR_AMD64_VM_PAGE_FLUSH,
2012                                        start | sev->asid);
2013
2014                                 start += PAGE_SIZE;
2015                         }
2016
2017                         return;
2018                 }
2019
2020                 WARN(1, "Address overflow, using WBINVD\n");
2021         }
2022
2023         /*
2024          * Hardware should always have one of the above features,
2025          * but if not, use WBINVD and issue a warning.
2026          */
2027         WARN_ONCE(1, "Using WBINVD to flush guest memory\n");
2028         wbinvd_on_all_cpus();
2029 }
2030
2031 void sev_free_vcpu(struct kvm_vcpu *vcpu)
2032 {
2033         struct vcpu_svm *svm;
2034
2035         if (!sev_es_guest(vcpu->kvm))
2036                 return;
2037
2038         svm = to_svm(vcpu);
2039
2040         if (vcpu->arch.guest_state_protected)
2041                 sev_flush_guest_memory(svm, svm->vmsa, PAGE_SIZE);
2042         __free_page(virt_to_page(svm->vmsa));
2043
2044         if (svm->ghcb_sa_free)
2045                 kfree(svm->ghcb_sa);
2046 }
2047
2048 static void dump_ghcb(struct vcpu_svm *svm)
2049 {
2050         struct ghcb *ghcb = svm->ghcb;
2051         unsigned int nbits;
2052
2053         /* Re-use the dump_invalid_vmcb module parameter */
2054         if (!dump_invalid_vmcb) {
2055                 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
2056                 return;
2057         }
2058
2059         nbits = sizeof(ghcb->save.valid_bitmap) * 8;
2060
2061         pr_err("GHCB (GPA=%016llx):\n", svm->vmcb->control.ghcb_gpa);
2062         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
2063                ghcb->save.sw_exit_code, ghcb_sw_exit_code_is_valid(ghcb));
2064         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
2065                ghcb->save.sw_exit_info_1, ghcb_sw_exit_info_1_is_valid(ghcb));
2066         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
2067                ghcb->save.sw_exit_info_2, ghcb_sw_exit_info_2_is_valid(ghcb));
2068         pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
2069                ghcb->save.sw_scratch, ghcb_sw_scratch_is_valid(ghcb));
2070         pr_err("%-20s%*pb\n", "valid_bitmap", nbits, ghcb->save.valid_bitmap);
2071 }
2072
2073 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
2074 {
2075         struct kvm_vcpu *vcpu = &svm->vcpu;
2076         struct ghcb *ghcb = svm->ghcb;
2077
2078         /*
2079          * The GHCB protocol so far allows for the following data
2080          * to be returned:
2081          *   GPRs RAX, RBX, RCX, RDX
2082          *
2083          * Copy their values, even if they may not have been written during the
2084          * VM-Exit.  It's the guest's responsibility to not consume random data.
2085          */
2086         ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
2087         ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
2088         ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
2089         ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
2090 }
2091
2092 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
2093 {
2094         struct vmcb_control_area *control = &svm->vmcb->control;
2095         struct kvm_vcpu *vcpu = &svm->vcpu;
2096         struct ghcb *ghcb = svm->ghcb;
2097         u64 exit_code;
2098
2099         /*
2100          * The GHCB protocol so far allows for the following data
2101          * to be supplied:
2102          *   GPRs RAX, RBX, RCX, RDX
2103          *   XCR0
2104          *   CPL
2105          *
2106          * VMMCALL allows the guest to provide extra registers. KVM also
2107          * expects RSI for hypercalls, so include that, too.
2108          *
2109          * Copy their values to the appropriate location if supplied.
2110          */
2111         memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
2112
2113         vcpu->arch.regs[VCPU_REGS_RAX] = ghcb_get_rax_if_valid(ghcb);
2114         vcpu->arch.regs[VCPU_REGS_RBX] = ghcb_get_rbx_if_valid(ghcb);
2115         vcpu->arch.regs[VCPU_REGS_RCX] = ghcb_get_rcx_if_valid(ghcb);
2116         vcpu->arch.regs[VCPU_REGS_RDX] = ghcb_get_rdx_if_valid(ghcb);
2117         vcpu->arch.regs[VCPU_REGS_RSI] = ghcb_get_rsi_if_valid(ghcb);
2118
2119         svm->vmcb->save.cpl = ghcb_get_cpl_if_valid(ghcb);
2120
2121         if (ghcb_xcr0_is_valid(ghcb)) {
2122                 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
2123                 kvm_update_cpuid_runtime(vcpu);
2124         }
2125
2126         /* Copy the GHCB exit information into the VMCB fields */
2127         exit_code = ghcb_get_sw_exit_code(ghcb);
2128         control->exit_code = lower_32_bits(exit_code);
2129         control->exit_code_hi = upper_32_bits(exit_code);
2130         control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
2131         control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
2132
2133         /* Clear the valid entries fields */
2134         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
2135 }
2136
2137 static int sev_es_validate_vmgexit(struct vcpu_svm *svm)
2138 {
2139         struct kvm_vcpu *vcpu;
2140         struct ghcb *ghcb;
2141         u64 exit_code = 0;
2142
2143         ghcb = svm->ghcb;
2144
2145         /* Only GHCB Usage code 0 is supported */
2146         if (ghcb->ghcb_usage)
2147                 goto vmgexit_err;
2148
2149         /*
2150          * Retrieve the exit code now even though is may not be marked valid
2151          * as it could help with debugging.
2152          */
2153         exit_code = ghcb_get_sw_exit_code(ghcb);
2154
2155         if (!ghcb_sw_exit_code_is_valid(ghcb) ||
2156             !ghcb_sw_exit_info_1_is_valid(ghcb) ||
2157             !ghcb_sw_exit_info_2_is_valid(ghcb))
2158                 goto vmgexit_err;
2159
2160         switch (ghcb_get_sw_exit_code(ghcb)) {
2161         case SVM_EXIT_READ_DR7:
2162                 break;
2163         case SVM_EXIT_WRITE_DR7:
2164                 if (!ghcb_rax_is_valid(ghcb))
2165                         goto vmgexit_err;
2166                 break;
2167         case SVM_EXIT_RDTSC:
2168                 break;
2169         case SVM_EXIT_RDPMC:
2170                 if (!ghcb_rcx_is_valid(ghcb))
2171                         goto vmgexit_err;
2172                 break;
2173         case SVM_EXIT_CPUID:
2174                 if (!ghcb_rax_is_valid(ghcb) ||
2175                     !ghcb_rcx_is_valid(ghcb))
2176                         goto vmgexit_err;
2177                 if (ghcb_get_rax(ghcb) == 0xd)
2178                         if (!ghcb_xcr0_is_valid(ghcb))
2179                                 goto vmgexit_err;
2180                 break;
2181         case SVM_EXIT_INVD:
2182                 break;
2183         case SVM_EXIT_IOIO:
2184                 if (ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_STR_MASK) {
2185                         if (!ghcb_sw_scratch_is_valid(ghcb))
2186                                 goto vmgexit_err;
2187                 } else {
2188                         if (!(ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_TYPE_MASK))
2189                                 if (!ghcb_rax_is_valid(ghcb))
2190                                         goto vmgexit_err;
2191                 }
2192                 break;
2193         case SVM_EXIT_MSR:
2194                 if (!ghcb_rcx_is_valid(ghcb))
2195                         goto vmgexit_err;
2196                 if (ghcb_get_sw_exit_info_1(ghcb)) {
2197                         if (!ghcb_rax_is_valid(ghcb) ||
2198                             !ghcb_rdx_is_valid(ghcb))
2199                                 goto vmgexit_err;
2200                 }
2201                 break;
2202         case SVM_EXIT_VMMCALL:
2203                 if (!ghcb_rax_is_valid(ghcb) ||
2204                     !ghcb_cpl_is_valid(ghcb))
2205                         goto vmgexit_err;
2206                 break;
2207         case SVM_EXIT_RDTSCP:
2208                 break;
2209         case SVM_EXIT_WBINVD:
2210                 break;
2211         case SVM_EXIT_MONITOR:
2212                 if (!ghcb_rax_is_valid(ghcb) ||
2213                     !ghcb_rcx_is_valid(ghcb) ||
2214                     !ghcb_rdx_is_valid(ghcb))
2215                         goto vmgexit_err;
2216                 break;
2217         case SVM_EXIT_MWAIT:
2218                 if (!ghcb_rax_is_valid(ghcb) ||
2219                     !ghcb_rcx_is_valid(ghcb))
2220                         goto vmgexit_err;
2221                 break;
2222         case SVM_VMGEXIT_MMIO_READ:
2223         case SVM_VMGEXIT_MMIO_WRITE:
2224                 if (!ghcb_sw_scratch_is_valid(ghcb))
2225                         goto vmgexit_err;
2226                 break;
2227         case SVM_VMGEXIT_NMI_COMPLETE:
2228         case SVM_VMGEXIT_AP_HLT_LOOP:
2229         case SVM_VMGEXIT_AP_JUMP_TABLE:
2230         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2231                 break;
2232         default:
2233                 goto vmgexit_err;
2234         }
2235
2236         return 0;
2237
2238 vmgexit_err:
2239         vcpu = &svm->vcpu;
2240
2241         if (ghcb->ghcb_usage) {
2242                 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
2243                             ghcb->ghcb_usage);
2244         } else {
2245                 vcpu_unimpl(vcpu, "vmgexit: exit reason %#llx is not valid\n",
2246                             exit_code);
2247                 dump_ghcb(svm);
2248         }
2249
2250         vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2251         vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
2252         vcpu->run->internal.ndata = 2;
2253         vcpu->run->internal.data[0] = exit_code;
2254         vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
2255
2256         return -EINVAL;
2257 }
2258
2259 void sev_es_unmap_ghcb(struct vcpu_svm *svm)
2260 {
2261         if (!svm->ghcb)
2262                 return;
2263
2264         if (svm->ghcb_sa_free) {
2265                 /*
2266                  * The scratch area lives outside the GHCB, so there is a
2267                  * buffer that, depending on the operation performed, may
2268                  * need to be synced, then freed.
2269                  */
2270                 if (svm->ghcb_sa_sync) {
2271                         kvm_write_guest(svm->vcpu.kvm,
2272                                         ghcb_get_sw_scratch(svm->ghcb),
2273                                         svm->ghcb_sa, svm->ghcb_sa_len);
2274                         svm->ghcb_sa_sync = false;
2275                 }
2276
2277                 kfree(svm->ghcb_sa);
2278                 svm->ghcb_sa = NULL;
2279                 svm->ghcb_sa_free = false;
2280         }
2281
2282         trace_kvm_vmgexit_exit(svm->vcpu.vcpu_id, svm->ghcb);
2283
2284         sev_es_sync_to_ghcb(svm);
2285
2286         kvm_vcpu_unmap(&svm->vcpu, &svm->ghcb_map, true);
2287         svm->ghcb = NULL;
2288 }
2289
2290 void pre_sev_run(struct vcpu_svm *svm, int cpu)
2291 {
2292         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
2293         int asid = sev_get_asid(svm->vcpu.kvm);
2294
2295         /* Assign the asid allocated with this SEV guest */
2296         svm->asid = asid;
2297
2298         /*
2299          * Flush guest TLB:
2300          *
2301          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
2302          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
2303          */
2304         if (sd->sev_vmcbs[asid] == svm->vmcb &&
2305             svm->vcpu.arch.last_vmentry_cpu == cpu)
2306                 return;
2307
2308         sd->sev_vmcbs[asid] = svm->vmcb;
2309         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
2310         vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
2311 }
2312
2313 #define GHCB_SCRATCH_AREA_LIMIT         (16ULL * PAGE_SIZE)
2314 static bool setup_vmgexit_scratch(struct vcpu_svm *svm, bool sync, u64 len)
2315 {
2316         struct vmcb_control_area *control = &svm->vmcb->control;
2317         struct ghcb *ghcb = svm->ghcb;
2318         u64 ghcb_scratch_beg, ghcb_scratch_end;
2319         u64 scratch_gpa_beg, scratch_gpa_end;
2320         void *scratch_va;
2321
2322         scratch_gpa_beg = ghcb_get_sw_scratch(ghcb);
2323         if (!scratch_gpa_beg) {
2324                 pr_err("vmgexit: scratch gpa not provided\n");
2325                 return false;
2326         }
2327
2328         scratch_gpa_end = scratch_gpa_beg + len;
2329         if (scratch_gpa_end < scratch_gpa_beg) {
2330                 pr_err("vmgexit: scratch length (%#llx) not valid for scratch address (%#llx)\n",
2331                        len, scratch_gpa_beg);
2332                 return false;
2333         }
2334
2335         if ((scratch_gpa_beg & PAGE_MASK) == control->ghcb_gpa) {
2336                 /* Scratch area begins within GHCB */
2337                 ghcb_scratch_beg = control->ghcb_gpa +
2338                                    offsetof(struct ghcb, shared_buffer);
2339                 ghcb_scratch_end = control->ghcb_gpa +
2340                                    offsetof(struct ghcb, reserved_1);
2341
2342                 /*
2343                  * If the scratch area begins within the GHCB, it must be
2344                  * completely contained in the GHCB shared buffer area.
2345                  */
2346                 if (scratch_gpa_beg < ghcb_scratch_beg ||
2347                     scratch_gpa_end > ghcb_scratch_end) {
2348                         pr_err("vmgexit: scratch area is outside of GHCB shared buffer area (%#llx - %#llx)\n",
2349                                scratch_gpa_beg, scratch_gpa_end);
2350                         return false;
2351                 }
2352
2353                 scratch_va = (void *)svm->ghcb;
2354                 scratch_va += (scratch_gpa_beg - control->ghcb_gpa);
2355         } else {
2356                 /*
2357                  * The guest memory must be read into a kernel buffer, so
2358                  * limit the size
2359                  */
2360                 if (len > GHCB_SCRATCH_AREA_LIMIT) {
2361                         pr_err("vmgexit: scratch area exceeds KVM limits (%#llx requested, %#llx limit)\n",
2362                                len, GHCB_SCRATCH_AREA_LIMIT);
2363                         return false;
2364                 }
2365                 scratch_va = kzalloc(len, GFP_KERNEL_ACCOUNT);
2366                 if (!scratch_va)
2367                         return false;
2368
2369                 if (kvm_read_guest(svm->vcpu.kvm, scratch_gpa_beg, scratch_va, len)) {
2370                         /* Unable to copy scratch area from guest */
2371                         pr_err("vmgexit: kvm_read_guest for scratch area failed\n");
2372
2373                         kfree(scratch_va);
2374                         return false;
2375                 }
2376
2377                 /*
2378                  * The scratch area is outside the GHCB. The operation will
2379                  * dictate whether the buffer needs to be synced before running
2380                  * the vCPU next time (i.e. a read was requested so the data
2381                  * must be written back to the guest memory).
2382                  */
2383                 svm->ghcb_sa_sync = sync;
2384                 svm->ghcb_sa_free = true;
2385         }
2386
2387         svm->ghcb_sa = scratch_va;
2388         svm->ghcb_sa_len = len;
2389
2390         return true;
2391 }
2392
2393 static void set_ghcb_msr_bits(struct vcpu_svm *svm, u64 value, u64 mask,
2394                               unsigned int pos)
2395 {
2396         svm->vmcb->control.ghcb_gpa &= ~(mask << pos);
2397         svm->vmcb->control.ghcb_gpa |= (value & mask) << pos;
2398 }
2399
2400 static u64 get_ghcb_msr_bits(struct vcpu_svm *svm, u64 mask, unsigned int pos)
2401 {
2402         return (svm->vmcb->control.ghcb_gpa >> pos) & mask;
2403 }
2404
2405 static void set_ghcb_msr(struct vcpu_svm *svm, u64 value)
2406 {
2407         svm->vmcb->control.ghcb_gpa = value;
2408 }
2409
2410 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
2411 {
2412         struct vmcb_control_area *control = &svm->vmcb->control;
2413         struct kvm_vcpu *vcpu = &svm->vcpu;
2414         u64 ghcb_info;
2415         int ret = 1;
2416
2417         ghcb_info = control->ghcb_gpa & GHCB_MSR_INFO_MASK;
2418
2419         trace_kvm_vmgexit_msr_protocol_enter(svm->vcpu.vcpu_id,
2420                                              control->ghcb_gpa);
2421
2422         switch (ghcb_info) {
2423         case GHCB_MSR_SEV_INFO_REQ:
2424                 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2425                                                     GHCB_VERSION_MIN,
2426                                                     sev_enc_bit));
2427                 break;
2428         case GHCB_MSR_CPUID_REQ: {
2429                 u64 cpuid_fn, cpuid_reg, cpuid_value;
2430
2431                 cpuid_fn = get_ghcb_msr_bits(svm,
2432                                              GHCB_MSR_CPUID_FUNC_MASK,
2433                                              GHCB_MSR_CPUID_FUNC_POS);
2434
2435                 /* Initialize the registers needed by the CPUID intercept */
2436                 vcpu->arch.regs[VCPU_REGS_RAX] = cpuid_fn;
2437                 vcpu->arch.regs[VCPU_REGS_RCX] = 0;
2438
2439                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_CPUID);
2440                 if (!ret) {
2441                         ret = -EINVAL;
2442                         break;
2443                 }
2444
2445                 cpuid_reg = get_ghcb_msr_bits(svm,
2446                                               GHCB_MSR_CPUID_REG_MASK,
2447                                               GHCB_MSR_CPUID_REG_POS);
2448                 if (cpuid_reg == 0)
2449                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RAX];
2450                 else if (cpuid_reg == 1)
2451                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RBX];
2452                 else if (cpuid_reg == 2)
2453                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RCX];
2454                 else
2455                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RDX];
2456
2457                 set_ghcb_msr_bits(svm, cpuid_value,
2458                                   GHCB_MSR_CPUID_VALUE_MASK,
2459                                   GHCB_MSR_CPUID_VALUE_POS);
2460
2461                 set_ghcb_msr_bits(svm, GHCB_MSR_CPUID_RESP,
2462                                   GHCB_MSR_INFO_MASK,
2463                                   GHCB_MSR_INFO_POS);
2464                 break;
2465         }
2466         case GHCB_MSR_TERM_REQ: {
2467                 u64 reason_set, reason_code;
2468
2469                 reason_set = get_ghcb_msr_bits(svm,
2470                                                GHCB_MSR_TERM_REASON_SET_MASK,
2471                                                GHCB_MSR_TERM_REASON_SET_POS);
2472                 reason_code = get_ghcb_msr_bits(svm,
2473                                                 GHCB_MSR_TERM_REASON_MASK,
2474                                                 GHCB_MSR_TERM_REASON_POS);
2475                 pr_info("SEV-ES guest requested termination: %#llx:%#llx\n",
2476                         reason_set, reason_code);
2477                 fallthrough;
2478         }
2479         default:
2480                 ret = -EINVAL;
2481         }
2482
2483         trace_kvm_vmgexit_msr_protocol_exit(svm->vcpu.vcpu_id,
2484                                             control->ghcb_gpa, ret);
2485
2486         return ret;
2487 }
2488
2489 int sev_handle_vmgexit(struct kvm_vcpu *vcpu)
2490 {
2491         struct vcpu_svm *svm = to_svm(vcpu);
2492         struct vmcb_control_area *control = &svm->vmcb->control;
2493         u64 ghcb_gpa, exit_code;
2494         struct ghcb *ghcb;
2495         int ret;
2496
2497         /* Validate the GHCB */
2498         ghcb_gpa = control->ghcb_gpa;
2499         if (ghcb_gpa & GHCB_MSR_INFO_MASK)
2500                 return sev_handle_vmgexit_msr_protocol(svm);
2501
2502         if (!ghcb_gpa) {
2503                 vcpu_unimpl(vcpu, "vmgexit: GHCB gpa is not set\n");
2504                 return -EINVAL;
2505         }
2506
2507         if (kvm_vcpu_map(vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->ghcb_map)) {
2508                 /* Unable to map GHCB from guest */
2509                 vcpu_unimpl(vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
2510                             ghcb_gpa);
2511                 return -EINVAL;
2512         }
2513
2514         svm->ghcb = svm->ghcb_map.hva;
2515         ghcb = svm->ghcb_map.hva;
2516
2517         trace_kvm_vmgexit_enter(vcpu->vcpu_id, ghcb);
2518
2519         exit_code = ghcb_get_sw_exit_code(ghcb);
2520
2521         ret = sev_es_validate_vmgexit(svm);
2522         if (ret)
2523                 return ret;
2524
2525         sev_es_sync_from_ghcb(svm);
2526         ghcb_set_sw_exit_info_1(ghcb, 0);
2527         ghcb_set_sw_exit_info_2(ghcb, 0);
2528
2529         ret = -EINVAL;
2530         switch (exit_code) {
2531         case SVM_VMGEXIT_MMIO_READ:
2532                 if (!setup_vmgexit_scratch(svm, true, control->exit_info_2))
2533                         break;
2534
2535                 ret = kvm_sev_es_mmio_read(vcpu,
2536                                            control->exit_info_1,
2537                                            control->exit_info_2,
2538                                            svm->ghcb_sa);
2539                 break;
2540         case SVM_VMGEXIT_MMIO_WRITE:
2541                 if (!setup_vmgexit_scratch(svm, false, control->exit_info_2))
2542                         break;
2543
2544                 ret = kvm_sev_es_mmio_write(vcpu,
2545                                             control->exit_info_1,
2546                                             control->exit_info_2,
2547                                             svm->ghcb_sa);
2548                 break;
2549         case SVM_VMGEXIT_NMI_COMPLETE:
2550                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_IRET);
2551                 break;
2552         case SVM_VMGEXIT_AP_HLT_LOOP:
2553                 ret = kvm_emulate_ap_reset_hold(vcpu);
2554                 break;
2555         case SVM_VMGEXIT_AP_JUMP_TABLE: {
2556                 struct kvm_sev_info *sev = &to_kvm_svm(vcpu->kvm)->sev_info;
2557
2558                 switch (control->exit_info_1) {
2559                 case 0:
2560                         /* Set AP jump table address */
2561                         sev->ap_jump_table = control->exit_info_2;
2562                         break;
2563                 case 1:
2564                         /* Get AP jump table address */
2565                         ghcb_set_sw_exit_info_2(ghcb, sev->ap_jump_table);
2566                         break;
2567                 default:
2568                         pr_err("svm: vmgexit: unsupported AP jump table request - exit_info_1=%#llx\n",
2569                                control->exit_info_1);
2570                         ghcb_set_sw_exit_info_1(ghcb, 1);
2571                         ghcb_set_sw_exit_info_2(ghcb,
2572                                                 X86_TRAP_UD |
2573                                                 SVM_EVTINJ_TYPE_EXEPT |
2574                                                 SVM_EVTINJ_VALID);
2575                 }
2576
2577                 ret = 1;
2578                 break;
2579         }
2580         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2581                 vcpu_unimpl(vcpu,
2582                             "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
2583                             control->exit_info_1, control->exit_info_2);
2584                 break;
2585         default:
2586                 ret = svm_invoke_exit_handler(vcpu, exit_code);
2587         }
2588
2589         return ret;
2590 }
2591
2592 int sev_es_string_io(struct vcpu_svm *svm, int size, unsigned int port, int in)
2593 {
2594         if (!setup_vmgexit_scratch(svm, in, svm->vmcb->control.exit_info_2))
2595                 return -EINVAL;
2596
2597         return kvm_sev_es_string_io(&svm->vcpu, size, port,
2598                                     svm->ghcb_sa, svm->ghcb_sa_len / size, in);
2599 }
2600
2601 void sev_es_init_vmcb(struct vcpu_svm *svm)
2602 {
2603         struct kvm_vcpu *vcpu = &svm->vcpu;
2604
2605         svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ES_ENABLE;
2606         svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
2607
2608         /*
2609          * An SEV-ES guest requires a VMSA area that is a separate from the
2610          * VMCB page. Do not include the encryption mask on the VMSA physical
2611          * address since hardware will access it using the guest key.
2612          */
2613         svm->vmcb->control.vmsa_pa = __pa(svm->vmsa);
2614
2615         /* Can't intercept CR register access, HV can't modify CR registers */
2616         svm_clr_intercept(svm, INTERCEPT_CR0_READ);
2617         svm_clr_intercept(svm, INTERCEPT_CR4_READ);
2618         svm_clr_intercept(svm, INTERCEPT_CR8_READ);
2619         svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
2620         svm_clr_intercept(svm, INTERCEPT_CR4_WRITE);
2621         svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
2622
2623         svm_clr_intercept(svm, INTERCEPT_SELECTIVE_CR0);
2624
2625         /* Track EFER/CR register changes */
2626         svm_set_intercept(svm, TRAP_EFER_WRITE);
2627         svm_set_intercept(svm, TRAP_CR0_WRITE);
2628         svm_set_intercept(svm, TRAP_CR4_WRITE);
2629         svm_set_intercept(svm, TRAP_CR8_WRITE);
2630
2631         /* No support for enable_vmware_backdoor */
2632         clr_exception_intercept(svm, GP_VECTOR);
2633
2634         /* Can't intercept XSETBV, HV can't modify XCR0 directly */
2635         svm_clr_intercept(svm, INTERCEPT_XSETBV);
2636
2637         /* Clear intercepts on selected MSRs */
2638         set_msr_interception(vcpu, svm->msrpm, MSR_EFER, 1, 1);
2639         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_CR_PAT, 1, 1);
2640         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
2641         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
2642         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
2643         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
2644 }
2645
2646 void sev_es_create_vcpu(struct vcpu_svm *svm)
2647 {
2648         /*
2649          * Set the GHCB MSR value as per the GHCB specification when creating
2650          * a vCPU for an SEV-ES guest.
2651          */
2652         set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2653                                             GHCB_VERSION_MIN,
2654                                             sev_enc_bit));
2655 }
2656
2657 void sev_es_prepare_guest_switch(struct vcpu_svm *svm, unsigned int cpu)
2658 {
2659         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
2660         struct vmcb_save_area *hostsa;
2661
2662         /*
2663          * As an SEV-ES guest, hardware will restore the host state on VMEXIT,
2664          * of which one step is to perform a VMLOAD. Since hardware does not
2665          * perform a VMSAVE on VMRUN, the host savearea must be updated.
2666          */
2667         vmsave(__sme_page_pa(sd->save_area));
2668
2669         /* XCR0 is restored on VMEXIT, save the current host value */
2670         hostsa = (struct vmcb_save_area *)(page_address(sd->save_area) + 0x400);
2671         hostsa->xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
2672
2673         /* PKRU is restored on VMEXIT, save the current host value */
2674         hostsa->pkru = read_pkru();
2675
2676         /* MSR_IA32_XSS is restored on VMEXIT, save the currnet host value */
2677         hostsa->xss = host_xss;
2678 }
2679
2680 void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
2681 {
2682         struct vcpu_svm *svm = to_svm(vcpu);
2683
2684         /* First SIPI: Use the values as initially set by the VMM */
2685         if (!svm->received_first_sipi) {
2686                 svm->received_first_sipi = true;
2687                 return;
2688         }
2689
2690         /*
2691          * Subsequent SIPI: Return from an AP Reset Hold VMGEXIT, where
2692          * the guest will set the CS and RIP. Set SW_EXIT_INFO_2 to a
2693          * non-zero value.
2694          */
2695         if (!svm->ghcb)
2696                 return;
2697
2698         ghcb_set_sw_exit_info_2(svm->ghcb, 1);
2699 }