arch/x86/kvm/svm/sev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel-based Virtual Machine driver for Linux
   4  *
   5  * AMD SVM-SEV support
   6  *
   7  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   8  */
   9
  10 #include <linux/kvm_types.h>
  11 #include <linux/kvm_host.h>
  12 #include <linux/kernel.h>
  13 #include <linux/highmem.h>
  14 #include <linux/psp-sev.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/swap.h>
  17 #include <linux/processor.h>
  18
  19 #include "x86.h"
  20 #include "svm.h"
  21 #include "cpuid.h"
  22
  23 static int sev_flush_asids(void);
  24 static DECLARE_RWSEM(sev_deactivate_lock);
  25 static DEFINE_MUTEX(sev_bitmap_lock);
  26 unsigned int max_sev_asid;
  27 static unsigned int min_sev_asid;
  28 static unsigned long *sev_asid_bitmap;
  29 static unsigned long *sev_reclaim_asid_bitmap;
  30 #define __sme_page_pa(x) __sme_set(page_to_pfn(x) << PAGE_SHIFT)
  31
  32 struct enc_region {
  33         struct list_head list;
  34         unsigned long npages;
  35         struct page **pages;
  36         unsigned long uaddr;
  37         unsigned long size;
  38 };
  39
  40 static int sev_flush_asids(void)
  41 {
  42         int ret, error = 0;
  43
  44         /*
  45          * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
  46          * so it must be guarded.
  47          */
  48         down_write(&sev_deactivate_lock);
  49
  50         wbinvd_on_all_cpus();
  51         ret = sev_guest_df_flush(&error);
  52
  53         up_write(&sev_deactivate_lock);
  54
  55         if (ret)
  56                 pr_err("SEV: DF_FLUSH failed, ret=%d, error=%#x\n", ret, error);
  57
  58         return ret;
  59 }
  60
  61 /* Must be called with the sev_bitmap_lock held */
  62 static bool __sev_recycle_asids(void)
  63 {
  64         int pos;
  65
  66         /* Check if there are any ASIDs to reclaim before performing a flush */
  67         pos = find_next_bit(sev_reclaim_asid_bitmap,
  68                             max_sev_asid, min_sev_asid - 1);
  69         if (pos >= max_sev_asid)
  70                 return false;
  71
  72         if (sev_flush_asids())
  73                 return false;
  74
  75         bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
  76                    max_sev_asid);
  77         bitmap_zero(sev_reclaim_asid_bitmap, max_sev_asid);
  78
  79         return true;
  80 }
  81
  82 static int sev_asid_new(void)
  83 {
  84         bool retry = true;
  85         int pos;
  86
  87         mutex_lock(&sev_bitmap_lock);
  88
  89         /*
  90          * SEV-enabled guest must use asid from min_sev_asid to max_sev_asid.
  91          */
  92 again:
  93         pos = find_next_zero_bit(sev_asid_bitmap, max_sev_asid, min_sev_asid - 1);
  94         if (pos >= max_sev_asid) {
  95                 if (retry && __sev_recycle_asids()) {
  96                         retry = false;
  97                         goto again;
  98                 }
  99                 mutex_unlock(&sev_bitmap_lock);
 100                 return -EBUSY;
 101         }
 102
 103         __set_bit(pos, sev_asid_bitmap);
 104
 105         mutex_unlock(&sev_bitmap_lock);
 106
 107         return pos + 1;
 108 }
 109
 110 static int sev_get_asid(struct kvm *kvm)
 111 {
 112         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 113
 114         return sev->asid;
 115 }
 116
 117 static void sev_asid_free(int asid)
 118 {
 119         struct svm_cpu_data *sd;
 120         int cpu, pos;
 121
 122         mutex_lock(&sev_bitmap_lock);
 123
 124         pos = asid - 1;
 125         __set_bit(pos, sev_reclaim_asid_bitmap);
 126
 127         for_each_possible_cpu(cpu) {
 128                 sd = per_cpu(svm_data, cpu);
 129                 sd->sev_vmcbs[pos] = NULL;
 130         }
 131
 132         mutex_unlock(&sev_bitmap_lock);
 133 }
 134
 135 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
 136 {
 137         struct sev_data_decommission *decommission;
 138         struct sev_data_deactivate *data;
 139
 140         if (!handle)
 141                 return;
 142
 143         data = kzalloc(sizeof(*data), GFP_KERNEL);
 144         if (!data)
 145                 return;
 146
 147         /* deactivate handle */
 148         data->handle = handle;
 149
 150         /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
 151         down_read(&sev_deactivate_lock);
 152         sev_guest_deactivate(data, NULL);
 153         up_read(&sev_deactivate_lock);
 154
 155         kfree(data);
 156
 157         decommission = kzalloc(sizeof(*decommission), GFP_KERNEL);
 158         if (!decommission)
 159                 return;
 160
 161         /* decommission handle */
 162         decommission->handle = handle;
 163         sev_guest_decommission(decommission, NULL);
 164
 165         kfree(decommission);
 166 }
 167
 168 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 169 {
 170         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 171         int asid, ret;
 172
 173         ret = -EBUSY;
 174         if (unlikely(sev->active))
 175                 return ret;
 176
 177         asid = sev_asid_new();
 178         if (asid < 0)
 179                 return ret;
 180
 181         ret = sev_platform_init(&argp->error);
 182         if (ret)
 183                 goto e_free;
 184
 185         sev->active = true;
 186         sev->asid = asid;
 187         INIT_LIST_HEAD(&sev->regions_list);
 188
 189         return 0;
 190
 191 e_free:
 192         sev_asid_free(asid);
 193         return ret;
 194 }
 195
 196 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
 197 {
 198         struct sev_data_activate *data;
 199         int asid = sev_get_asid(kvm);
 200         int ret;
 201
 202         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 203         if (!data)
 204                 return -ENOMEM;
 205
 206         /* activate ASID on the given handle */
 207         data->handle = handle;
 208         data->asid   = asid;
 209         ret = sev_guest_activate(data, error);
 210         kfree(data);
 211
 212         return ret;
 213 }
 214
 215 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
 216 {
 217         struct fd f;
 218         int ret;
 219
 220         f = fdget(fd);
 221         if (!f.file)
 222                 return -EBADF;
 223
 224         ret = sev_issue_cmd_external_user(f.file, id, data, error);
 225
 226         fdput(f);
 227         return ret;
 228 }
 229
 230 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
 231 {
 232         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 233
 234         return __sev_issue_cmd(sev->fd, id, data, error);
 235 }
 236
 237 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
 238 {
 239         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 240         struct sev_data_launch_start *start;
 241         struct kvm_sev_launch_start params;
 242         void *dh_blob, *session_blob;
 243         int *error = &argp->error;
 244         int ret;
 245
 246         if (!sev_guest(kvm))
 247                 return -ENOTTY;
 248
 249         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 250                 return -EFAULT;
 251
 252         start = kzalloc(sizeof(*start), GFP_KERNEL_ACCOUNT);
 253         if (!start)
 254                 return -ENOMEM;
 255
 256         dh_blob = NULL;
 257         if (params.dh_uaddr) {
 258                 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
 259                 if (IS_ERR(dh_blob)) {
 260                         ret = PTR_ERR(dh_blob);
 261                         goto e_free;
 262                 }
 263
 264                 start->dh_cert_address = __sme_set(__pa(dh_blob));
 265                 start->dh_cert_len = params.dh_len;
 266         }
 267
 268         session_blob = NULL;
 269         if (params.session_uaddr) {
 270                 session_blob = psp_copy_user_blob(params.session_uaddr, params.session_len);
 271                 if (IS_ERR(session_blob)) {
 272                         ret = PTR_ERR(session_blob);
 273                         goto e_free_dh;
 274                 }
 275
 276                 start->session_address = __sme_set(__pa(session_blob));
 277                 start->session_len = params.session_len;
 278         }
 279
 280         start->handle = params.handle;
 281         start->policy = params.policy;
 282
 283         /* create memory encryption context */
 284         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_LAUNCH_START, start, error);
 285         if (ret)
 286                 goto e_free_session;
 287
 288         /* Bind ASID to this guest */
 289         ret = sev_bind_asid(kvm, start->handle, error);
 290         if (ret)
 291                 goto e_free_session;
 292
 293         /* return handle to userspace */
 294         params.handle = start->handle;
 295         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params))) {
 296                 sev_unbind_asid(kvm, start->handle);
 297                 ret = -EFAULT;
 298                 goto e_free_session;
 299         }
 300
 301         sev->handle = start->handle;
 302         sev->fd = argp->sev_fd;
 303
 304 e_free_session:
 305         kfree(session_blob);
 306 e_free_dh:
 307         kfree(dh_blob);
 308 e_free:
 309         kfree(start);
 310         return ret;
 311 }
 312
 313 static struct page **sev_pin_memory(struct kvm *kvm, unsigned long uaddr,
 314                                     unsigned long ulen, unsigned long *n,
 315                                     int write)
 316 {
 317         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 318         unsigned long npages, size;
 319         int npinned;
 320         unsigned long locked, lock_limit;
 321         struct page **pages;
 322         unsigned long first, last;
 323         int ret;
 324
 325         if (ulen == 0 || uaddr + ulen < uaddr)
 326                 return ERR_PTR(-EINVAL);
 327
 328         /* Calculate number of pages. */
 329         first = (uaddr & PAGE_MASK) >> PAGE_SHIFT;
 330         last = ((uaddr + ulen - 1) & PAGE_MASK) >> PAGE_SHIFT;
 331         npages = (last - first + 1);
 332
 333         locked = sev->pages_locked + npages;
 334         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 335         if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
 336                 pr_err("SEV: %lu locked pages exceed the lock limit of %lu.\n", locked, lock_limit);
 337                 return ERR_PTR(-ENOMEM);
 338         }
 339
 340         if (WARN_ON_ONCE(npages > INT_MAX))
 341                 return ERR_PTR(-EINVAL);
 342
 343         /* Avoid using vmalloc for smaller buffers. */
 344         size = npages * sizeof(struct page *);
 345         if (size > PAGE_SIZE)
 346                 pages = __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 347         else
 348                 pages = kmalloc(size, GFP_KERNEL_ACCOUNT);
 349
 350         if (!pages)
 351                 return ERR_PTR(-ENOMEM);
 352
 353         /* Pin the user virtual address. */
 354         npinned = pin_user_pages_fast(uaddr, npages, write ? FOLL_WRITE : 0, pages);
 355         if (npinned != npages) {
 356                 pr_err("SEV: Failure locking %lu pages.\n", npages);
 357                 ret = -ENOMEM;
 358                 goto err;
 359         }
 360
 361         *n = npages;
 362         sev->pages_locked = locked;
 363
 364         return pages;
 365
 366 err:
 367         if (npinned > 0)
 368                 unpin_user_pages(pages, npinned);
 369
 370         kvfree(pages);
 371         return ERR_PTR(ret);
 372 }
 373
 374 static void sev_unpin_memory(struct kvm *kvm, struct page **pages,
 375                              unsigned long npages)
 376 {
 377         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 378
 379         unpin_user_pages(pages, npages);
 380         kvfree(pages);
 381         sev->pages_locked -= npages;
 382 }
 383
 384 static void sev_clflush_pages(struct page *pages[], unsigned long npages)
 385 {
 386         uint8_t *page_virtual;
 387         unsigned long i;
 388
 389         if (this_cpu_has(X86_FEATURE_SME_COHERENT) || npages == 0 ||
 390             pages == NULL)
 391                 return;
 392
 393         for (i = 0; i < npages; i++) {
 394                 page_virtual = kmap_atomic(pages[i]);
 395                 clflush_cache_range(page_virtual, PAGE_SIZE);
 396                 kunmap_atomic(page_virtual);
 397         }
 398 }
 399
 400 static unsigned long get_num_contig_pages(unsigned long idx,
 401                                 struct page **inpages, unsigned long npages)
 402 {
 403         unsigned long paddr, next_paddr;
 404         unsigned long i = idx + 1, pages = 1;
 405
 406         /* find the number of contiguous pages starting from idx */
 407         paddr = __sme_page_pa(inpages[idx]);
 408         while (i < npages) {
 409                 next_paddr = __sme_page_pa(inpages[i++]);
 410                 if ((paddr + PAGE_SIZE) == next_paddr) {
 411                         pages++;
 412                         paddr = next_paddr;
 413                         continue;
 414                 }
 415                 break;
 416         }
 417
 418         return pages;
 419 }
 420
 421 static int sev_launch_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
 422 {
 423         unsigned long vaddr, vaddr_end, next_vaddr, npages, pages, size, i;
 424         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 425         struct kvm_sev_launch_update_data params;
 426         struct sev_data_launch_update_data *data;
 427         struct page **inpages;
 428         int ret;
 429
 430         if (!sev_guest(kvm))
 431                 return -ENOTTY;
 432
 433         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 434                 return -EFAULT;
 435
 436         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 437         if (!data)
 438                 return -ENOMEM;
 439
 440         vaddr = params.uaddr;
 441         size = params.len;
 442         vaddr_end = vaddr + size;
 443
 444         /* Lock the user memory. */
 445         inpages = sev_pin_memory(kvm, vaddr, size, &npages, 1);
 446         if (IS_ERR(inpages)) {
 447                 ret = PTR_ERR(inpages);
 448                 goto e_free;
 449         }
 450
 451         /*
 452          * Flush (on non-coherent CPUs) before LAUNCH_UPDATE encrypts pages in
 453          * place; the cache may contain the data that was written unencrypted.
 454          */
 455         sev_clflush_pages(inpages, npages);
 456
 457         for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
 458                 int offset, len;
 459
 460                 /*
 461                  * If the user buffer is not page-aligned, calculate the offset
 462                  * within the page.
 463                  */
 464                 offset = vaddr & (PAGE_SIZE - 1);
 465
 466                 /* Calculate the number of pages that can be encrypted in one go. */
 467                 pages = get_num_contig_pages(i, inpages, npages);
 468
 469                 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
 470
 471                 data->handle = sev->handle;
 472                 data->len = len;
 473                 data->address = __sme_page_pa(inpages[i]) + offset;
 474                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, data, &argp->error);
 475                 if (ret)
 476                         goto e_unpin;
 477
 478                 size -= len;
 479                 next_vaddr = vaddr + len;
 480         }
 481
 482 e_unpin:
 483         /* content of memory is updated, mark pages dirty */
 484         for (i = 0; i < npages; i++) {
 485                 set_page_dirty_lock(inpages[i]);
 486                 mark_page_accessed(inpages[i]);
 487         }
 488         /* unlock the user pages */
 489         sev_unpin_memory(kvm, inpages, npages);
 490 e_free:
 491         kfree(data);
 492         return ret;
 493 }
 494
 495 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 496 {
 497         void __user *measure = (void __user *)(uintptr_t)argp->data;
 498         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 499         struct sev_data_launch_measure *data;
 500         struct kvm_sev_launch_measure params;
 501         void __user *p = NULL;
 502         void *blob = NULL;
 503         int ret;
 504
 505         if (!sev_guest(kvm))
 506                 return -ENOTTY;
 507
 508         if (copy_from_user(&params, measure, sizeof(params)))
 509                 return -EFAULT;
 510
 511         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 512         if (!data)
 513                 return -ENOMEM;
 514
 515         /* User wants to query the blob length */
 516         if (!params.len)
 517                 goto cmd;
 518
 519         p = (void __user *)(uintptr_t)params.uaddr;
 520         if (p) {
 521                 if (params.len > SEV_FW_BLOB_MAX_SIZE) {
 522                         ret = -EINVAL;
 523                         goto e_free;
 524                 }
 525
 526                 ret = -ENOMEM;
 527                 blob = kmalloc(params.len, GFP_KERNEL);
 528                 if (!blob)
 529                         goto e_free;
 530
 531                 data->address = __psp_pa(blob);
 532                 data->len = params.len;
 533         }
 534
 535 cmd:
 536         data->handle = sev->handle;
 537         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, data, &argp->error);
 538
 539         /*
 540          * If we query the session length, FW responded with expected data.
 541          */
 542         if (!params.len)
 543                 goto done;
 544
 545         if (ret)
 546                 goto e_free_blob;
 547
 548         if (blob) {
 549                 if (copy_to_user(p, blob, params.len))
 550                         ret = -EFAULT;
 551         }
 552
 553 done:
 554         params.len = data->len;
 555         if (copy_to_user(measure, &params, sizeof(params)))
 556                 ret = -EFAULT;
 557 e_free_blob:
 558         kfree(blob);
 559 e_free:
 560         kfree(data);
 561         return ret;
 562 }
 563
 564 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 565 {
 566         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 567         struct sev_data_launch_finish *data;
 568         int ret;
 569
 570         if (!sev_guest(kvm))
 571                 return -ENOTTY;
 572
 573         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 574         if (!data)
 575                 return -ENOMEM;
 576
 577         data->handle = sev->handle;
 578         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, data, &argp->error);
 579
 580         kfree(data);
 581         return ret;
 582 }
 583
 584 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 585 {
 586         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 587         struct kvm_sev_guest_status params;
 588         struct sev_data_guest_status *data;
 589         int ret;
 590
 591         if (!sev_guest(kvm))
 592                 return -ENOTTY;
 593
 594         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 595         if (!data)
 596                 return -ENOMEM;
 597
 598         data->handle = sev->handle;
 599         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, data, &argp->error);
 600         if (ret)
 601                 goto e_free;
 602
 603         params.policy = data->policy;
 604         params.state = data->state;
 605         params.handle = data->handle;
 606
 607         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 608                 ret = -EFAULT;
 609 e_free:
 610         kfree(data);
 611         return ret;
 612 }
 613
 614 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 615                                unsigned long dst, int size,
 616                                int *error, bool enc)
 617 {
 618         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 619         struct sev_data_dbg *data;
 620         int ret;
 621
 622         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 623         if (!data)
 624                 return -ENOMEM;
 625
 626         data->handle = sev->handle;
 627         data->dst_addr = dst;
 628         data->src_addr = src;
 629         data->len = size;
 630
 631         ret = sev_issue_cmd(kvm,
 632                             enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 633                             data, error);
 634         kfree(data);
 635         return ret;
 636 }
 637
 638 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 639                              unsigned long dst_paddr, int sz, int *err)
 640 {
 641         int offset;
 642
 643         /*
 644          * Its safe to read more than we are asked, caller should ensure that
 645          * destination has enough space.
 646          */
 647         src_paddr = round_down(src_paddr, 16);
 648         offset = src_paddr & 15;
 649         sz = round_up(sz + offset, 16);
 650
 651         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 652 }
 653
 654 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 655                                   unsigned long __user dst_uaddr,
 656                                   unsigned long dst_paddr,
 657                                   int size, int *err)
 658 {
 659         struct page *tpage = NULL;
 660         int ret, offset;
 661
 662         /* if inputs are not 16-byte then use intermediate buffer */
 663         if (!IS_ALIGNED(dst_paddr, 16) ||
 664             !IS_ALIGNED(paddr,     16) ||
 665             !IS_ALIGNED(size,      16)) {
 666                 tpage = (void *)alloc_page(GFP_KERNEL);
 667                 if (!tpage)
 668                         return -ENOMEM;
 669
 670                 dst_paddr = __sme_page_pa(tpage);
 671         }
 672
 673         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 674         if (ret)
 675                 goto e_free;
 676
 677         if (tpage) {
 678                 offset = paddr & 15;
 679                 if (copy_to_user((void __user *)(uintptr_t)dst_uaddr,
 680                                  page_address(tpage) + offset, size))
 681                         ret = -EFAULT;
 682         }
 683
 684 e_free:
 685         if (tpage)
 686                 __free_page(tpage);
 687
 688         return ret;
 689 }
 690
 691 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 692                                   unsigned long __user vaddr,
 693                                   unsigned long dst_paddr,
 694                                   unsigned long __user dst_vaddr,
 695                                   int size, int *error)
 696 {
 697         struct page *src_tpage = NULL;
 698         struct page *dst_tpage = NULL;
 699         int ret, len = size;
 700
 701         /* If source buffer is not aligned then use an intermediate buffer */
 702         if (!IS_ALIGNED(vaddr, 16)) {
 703                 src_tpage = alloc_page(GFP_KERNEL);
 704                 if (!src_tpage)
 705                         return -ENOMEM;
 706
 707                 if (copy_from_user(page_address(src_tpage),
 708                                 (void __user *)(uintptr_t)vaddr, size)) {
 709                         __free_page(src_tpage);
 710                         return -EFAULT;
 711                 }
 712
 713                 paddr = __sme_page_pa(src_tpage);
 714         }
 715
 716         /*
 717          *  If destination buffer or length is not aligned then do read-modify-write:
 718          *   - decrypt destination in an intermediate buffer
 719          *   - copy the source buffer in an intermediate buffer
 720          *   - use the intermediate buffer as source buffer
 721          */
 722         if (!IS_ALIGNED(dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 723                 int dst_offset;
 724
 725                 dst_tpage = alloc_page(GFP_KERNEL);
 726                 if (!dst_tpage) {
 727                         ret = -ENOMEM;
 728                         goto e_free;
 729                 }
 730
 731                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 732                                         __sme_page_pa(dst_tpage), size, error);
 733                 if (ret)
 734                         goto e_free;
 735
 736                 /*
 737                  *  If source is kernel buffer then use memcpy() otherwise
 738                  *  copy_from_user().
 739                  */
 740                 dst_offset = dst_paddr & 15;
 741
 742                 if (src_tpage)
 743                         memcpy(page_address(dst_tpage) + dst_offset,
 744                                page_address(src_tpage), size);
 745                 else {
 746                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 747                                            (void __user *)(uintptr_t)vaddr, size)) {
 748                                 ret = -EFAULT;
 749                                 goto e_free;
 750                         }
 751                 }
 752
 753                 paddr = __sme_page_pa(dst_tpage);
 754                 dst_paddr = round_down(dst_paddr, 16);
 755                 len = round_up(size, 16);
 756         }
 757
 758         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 759
 760 e_free:
 761         if (src_tpage)
 762                 __free_page(src_tpage);
 763         if (dst_tpage)
 764                 __free_page(dst_tpage);
 765         return ret;
 766 }
 767
 768 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 769 {
 770         unsigned long vaddr, vaddr_end, next_vaddr;
 771         unsigned long dst_vaddr;
 772         struct page **src_p, **dst_p;
 773         struct kvm_sev_dbg debug;
 774         unsigned long n;
 775         unsigned int size;
 776         int ret;
 777
 778         if (!sev_guest(kvm))
 779                 return -ENOTTY;
 780
 781         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 782                 return -EFAULT;
 783
 784         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 785                 return -EINVAL;
 786         if (!debug.dst_uaddr)
 787                 return -EINVAL;
 788
 789         vaddr = debug.src_uaddr;
 790         size = debug.len;
 791         vaddr_end = vaddr + size;
 792         dst_vaddr = debug.dst_uaddr;
 793
 794         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 795                 int len, s_off, d_off;
 796
 797                 /* lock userspace source and destination page */
 798                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 799                 if (IS_ERR(src_p))
 800                         return PTR_ERR(src_p);
 801
 802                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 803                 if (IS_ERR(dst_p)) {
 804                         sev_unpin_memory(kvm, src_p, n);
 805                         return PTR_ERR(dst_p);
 806                 }
 807
 808                 /*
 809                  * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
 810                  * the pages; flush the destination too so that future accesses do not
 811                  * see stale data.
 812                  */
 813                 sev_clflush_pages(src_p, 1);
 814                 sev_clflush_pages(dst_p, 1);
 815
 816                 /*
 817                  * Since user buffer may not be page aligned, calculate the
 818                  * offset within the page.
 819                  */
 820                 s_off = vaddr & ~PAGE_MASK;
 821                 d_off = dst_vaddr & ~PAGE_MASK;
 822                 len = min_t(size_t, (PAGE_SIZE - s_off), size);
 823
 824                 if (dec)
 825                         ret = __sev_dbg_decrypt_user(kvm,
 826                                                      __sme_page_pa(src_p[0]) + s_off,
 827                                                      dst_vaddr,
 828                                                      __sme_page_pa(dst_p[0]) + d_off,
 829                                                      len, &argp->error);
 830                 else
 831                         ret = __sev_dbg_encrypt_user(kvm,
 832                                                      __sme_page_pa(src_p[0]) + s_off,
 833                                                      vaddr,
 834                                                      __sme_page_pa(dst_p[0]) + d_off,
 835                                                      dst_vaddr,
 836                                                      len, &argp->error);
 837
 838                 sev_unpin_memory(kvm, src_p, n);
 839                 sev_unpin_memory(kvm, dst_p, n);
 840
 841                 if (ret)
 842                         goto err;
 843
 844                 next_vaddr = vaddr + len;
 845                 dst_vaddr = dst_vaddr + len;
 846                 size -= len;
 847         }
 848 err:
 849         return ret;
 850 }
 851
 852 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
 853 {
 854         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 855         struct sev_data_launch_secret *data;
 856         struct kvm_sev_launch_secret params;
 857         struct page **pages;
 858         void *blob, *hdr;
 859         unsigned long n, i;
 860         int ret, offset;
 861
 862         if (!sev_guest(kvm))
 863                 return -ENOTTY;
 864
 865         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 866                 return -EFAULT;
 867
 868         pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, 1);
 869         if (IS_ERR(pages))
 870                 return PTR_ERR(pages);
 871
 872         /*
 873          * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
 874          * place; the cache may contain the data that was written unencrypted.
 875          */
 876         sev_clflush_pages(pages, n);
 877
 878         /*
 879          * The secret must be copied into contiguous memory region, lets verify
 880          * that userspace memory pages are contiguous before we issue command.
 881          */
 882         if (get_num_contig_pages(0, pages, n) != n) {
 883                 ret = -EINVAL;
 884                 goto e_unpin_memory;
 885         }
 886
 887         ret = -ENOMEM;
 888         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 889         if (!data)
 890                 goto e_unpin_memory;
 891
 892         offset = params.guest_uaddr & (PAGE_SIZE - 1);
 893         data->guest_address = __sme_page_pa(pages[0]) + offset;
 894         data->guest_len = params.guest_len;
 895
 896         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
 897         if (IS_ERR(blob)) {
 898                 ret = PTR_ERR(blob);
 899                 goto e_free;
 900         }
 901
 902         data->trans_address = __psp_pa(blob);
 903         data->trans_len = params.trans_len;
 904
 905         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
 906         if (IS_ERR(hdr)) {
 907                 ret = PTR_ERR(hdr);
 908                 goto e_free_blob;
 909         }
 910         data->hdr_address = __psp_pa(hdr);
 911         data->hdr_len = params.hdr_len;
 912
 913         data->handle = sev->handle;
 914         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, data, &argp->error);
 915
 916         kfree(hdr);
 917
 918 e_free_blob:
 919         kfree(blob);
 920 e_free:
 921         kfree(data);
 922 e_unpin_memory:
 923         /* content of memory is updated, mark pages dirty */
 924         for (i = 0; i < n; i++) {
 925                 set_page_dirty_lock(pages[i]);
 926                 mark_page_accessed(pages[i]);
 927         }
 928         sev_unpin_memory(kvm, pages, n);
 929         return ret;
 930 }
 931
 932 int svm_mem_enc_op(struct kvm *kvm, void __user *argp)
 933 {
 934         struct kvm_sev_cmd sev_cmd;
 935         int r;
 936
 937         if (!svm_sev_enabled() || !sev)
 938                 return -ENOTTY;
 939
 940         if (!argp)
 941                 return 0;
 942
 943         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
 944                 return -EFAULT;
 945
 946         mutex_lock(&kvm->lock);
 947
 948         switch (sev_cmd.id) {
 949         case KVM_SEV_INIT:
 950                 r = sev_guest_init(kvm, &sev_cmd);
 951                 break;
 952         case KVM_SEV_LAUNCH_START:
 953                 r = sev_launch_start(kvm, &sev_cmd);
 954                 break;
 955         case KVM_SEV_LAUNCH_UPDATE_DATA:
 956                 r = sev_launch_update_data(kvm, &sev_cmd);
 957                 break;
 958         case KVM_SEV_LAUNCH_MEASURE:
 959                 r = sev_launch_measure(kvm, &sev_cmd);
 960                 break;
 961         case KVM_SEV_LAUNCH_FINISH:
 962                 r = sev_launch_finish(kvm, &sev_cmd);
 963                 break;
 964         case KVM_SEV_GUEST_STATUS:
 965                 r = sev_guest_status(kvm, &sev_cmd);
 966                 break;
 967         case KVM_SEV_DBG_DECRYPT:
 968                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
 969                 break;
 970         case KVM_SEV_DBG_ENCRYPT:
 971                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
 972                 break;
 973         case KVM_SEV_LAUNCH_SECRET:
 974                 r = sev_launch_secret(kvm, &sev_cmd);
 975                 break;
 976         default:
 977                 r = -EINVAL;
 978                 goto out;
 979         }
 980
 981         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
 982                 r = -EFAULT;
 983
 984 out:
 985         mutex_unlock(&kvm->lock);
 986         return r;
 987 }
 988
 989 int svm_register_enc_region(struct kvm *kvm,
 990                             struct kvm_enc_region *range)
 991 {
 992         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 993         struct enc_region *region;
 994         int ret = 0;
 995
 996         if (!sev_guest(kvm))
 997                 return -ENOTTY;
 998
 999         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
1000                 return -EINVAL;
1001
1002         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
1003         if (!region)
1004                 return -ENOMEM;
1005
1006         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
1007         if (IS_ERR(region->pages)) {
1008                 ret = PTR_ERR(region->pages);
1009                 goto e_free;
1010         }
1011
1012         /*
1013          * The guest may change the memory encryption attribute from C=0 -> C=1
1014          * or vice versa for this memory range. Lets make sure caches are
1015          * flushed to ensure that guest data gets written into memory with
1016          * correct C-bit.
1017          */
1018         sev_clflush_pages(region->pages, region->npages);
1019
1020         region->uaddr = range->addr;
1021         region->size = range->size;
1022
1023         mutex_lock(&kvm->lock);
1024         list_add_tail(&region->list, &sev->regions_list);
1025         mutex_unlock(&kvm->lock);
1026
1027         return ret;
1028
1029 e_free:
1030         kfree(region);
1031         return ret;
1032 }
1033
1034 static struct enc_region *
1035 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
1036 {
1037         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1038         struct list_head *head = &sev->regions_list;
1039         struct enc_region *i;
1040
1041         list_for_each_entry(i, head, list) {
1042                 if (i->uaddr == range->addr &&
1043                     i->size == range->size)
1044                         return i;
1045         }
1046
1047         return NULL;
1048 }
1049
1050 static void __unregister_enc_region_locked(struct kvm *kvm,
1051                                            struct enc_region *region)
1052 {
1053         sev_unpin_memory(kvm, region->pages, region->npages);
1054         list_del(&region->list);
1055         kfree(region);
1056 }
1057
1058 int svm_unregister_enc_region(struct kvm *kvm,
1059                               struct kvm_enc_region *range)
1060 {
1061         struct enc_region *region;
1062         int ret;
1063
1064         mutex_lock(&kvm->lock);
1065
1066         if (!sev_guest(kvm)) {
1067                 ret = -ENOTTY;
1068                 goto failed;
1069         }
1070
1071         region = find_enc_region(kvm, range);
1072         if (!region) {
1073                 ret = -EINVAL;
1074                 goto failed;
1075         }
1076
1077         /*
1078          * Ensure that all guest tagged cache entries are flushed before
1079          * releasing the pages back to the system for use. CLFLUSH will
1080          * not do this, so issue a WBINVD.
1081          */
1082         wbinvd_on_all_cpus();
1083
1084         __unregister_enc_region_locked(kvm, region);
1085
1086         mutex_unlock(&kvm->lock);
1087         return 0;
1088
1089 failed:
1090         mutex_unlock(&kvm->lock);
1091         return ret;
1092 }
1093
1094 void sev_vm_destroy(struct kvm *kvm)
1095 {
1096         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1097         struct list_head *head = &sev->regions_list;
1098         struct list_head *pos, *q;
1099
1100         if (!sev_guest(kvm))
1101                 return;
1102
1103         mutex_lock(&kvm->lock);
1104
1105         /*
1106          * Ensure that all guest tagged cache entries are flushed before
1107          * releasing the pages back to the system for use. CLFLUSH will
1108          * not do this, so issue a WBINVD.
1109          */
1110         wbinvd_on_all_cpus();
1111
1112         /*
1113          * if userspace was terminated before unregistering the memory regions
1114          * then lets unpin all the registered memory.
1115          */
1116         if (!list_empty(head)) {
1117                 list_for_each_safe(pos, q, head) {
1118                         __unregister_enc_region_locked(kvm,
1119                                 list_entry(pos, struct enc_region, list));
1120                         cond_resched();
1121                 }
1122         }
1123
1124         mutex_unlock(&kvm->lock);
1125
1126         sev_unbind_asid(kvm, sev->handle);
1127         sev_asid_free(sev->asid);
1128 }
1129
1130 void __init sev_hardware_setup(void)
1131 {
1132         unsigned int eax, ebx, ecx, edx;
1133         bool sev_es_supported = false;
1134         bool sev_supported = false;
1135
1136         /* Does the CPU support SEV? */
1137         if (!boot_cpu_has(X86_FEATURE_SEV))
1138                 goto out;
1139
1140         /* Retrieve SEV CPUID information */
1141         cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
1142
1143         /* Maximum number of encrypted guests supported simultaneously */
1144         max_sev_asid = ecx;
1145
1146         if (!svm_sev_enabled())
1147                 goto out;
1148
1149         /* Minimum ASID value that should be used for SEV guest */
1150         min_sev_asid = edx;
1151
1152         /* Initialize SEV ASID bitmaps */
1153         sev_asid_bitmap = bitmap_zalloc(max_sev_asid, GFP_KERNEL);
1154         if (!sev_asid_bitmap)
1155                 goto out;
1156
1157         sev_reclaim_asid_bitmap = bitmap_zalloc(max_sev_asid, GFP_KERNEL);
1158         if (!sev_reclaim_asid_bitmap)
1159                 goto out;
1160
1161         pr_info("SEV supported: %u ASIDs\n", max_sev_asid - min_sev_asid + 1);
1162         sev_supported = true;
1163
1164         /* SEV-ES support requested? */
1165         if (!sev_es)
1166                 goto out;
1167
1168         /* Does the CPU support SEV-ES? */
1169         if (!boot_cpu_has(X86_FEATURE_SEV_ES))
1170                 goto out;
1171
1172         /* Has the system been allocated ASIDs for SEV-ES? */
1173         if (min_sev_asid == 1)
1174                 goto out;
1175
1176         pr_info("SEV-ES supported: %u ASIDs\n", min_sev_asid - 1);
1177         sev_es_supported = true;
1178
1179 out:
1180         sev = sev_supported;
1181         sev_es = sev_es_supported;
1182 }
1183
1184 void sev_hardware_teardown(void)
1185 {
1186         if (!svm_sev_enabled())
1187                 return;
1188
1189         bitmap_free(sev_asid_bitmap);
1190         bitmap_free(sev_reclaim_asid_bitmap);
1191
1192         sev_flush_asids();
1193 }
1194
1195 /*
1196  * Pages used by hardware to hold guest encrypted state must be flushed before
1197  * returning them to the system.
1198  */
1199 static void sev_flush_guest_memory(struct vcpu_svm *svm, void *va,
1200                                    unsigned long len)
1201 {
1202         /*
1203          * If hardware enforced cache coherency for encrypted mappings of the
1204          * same physical page is supported, nothing to do.
1205          */
1206         if (boot_cpu_has(X86_FEATURE_SME_COHERENT))
1207                 return;
1208
1209         /*
1210          * If the VM Page Flush MSR is supported, use it to flush the page
1211          * (using the page virtual address and the guest ASID).
1212          */
1213         if (boot_cpu_has(X86_FEATURE_VM_PAGE_FLUSH)) {
1214                 struct kvm_sev_info *sev;
1215                 unsigned long va_start;
1216                 u64 start, stop;
1217
1218                 /* Align start and stop to page boundaries. */
1219                 va_start = (unsigned long)va;
1220                 start = (u64)va_start & PAGE_MASK;
1221                 stop = PAGE_ALIGN((u64)va_start + len);
1222
1223                 if (start < stop) {
1224                         sev = &to_kvm_svm(svm->vcpu.kvm)->sev_info;
1225
1226                         while (start < stop) {
1227                                 wrmsrl(MSR_AMD64_VM_PAGE_FLUSH,
1228                                        start | sev->asid);
1229
1230                                 start += PAGE_SIZE;
1231                         }
1232
1233                         return;
1234                 }
1235
1236                 WARN(1, "Address overflow, using WBINVD\n");
1237         }
1238
1239         /*
1240          * Hardware should always have one of the above features,
1241          * but if not, use WBINVD and issue a warning.
1242          */
1243         WARN_ONCE(1, "Using WBINVD to flush guest memory\n");
1244         wbinvd_on_all_cpus();
1245 }
1246
1247 void sev_free_vcpu(struct kvm_vcpu *vcpu)
1248 {
1249         struct vcpu_svm *svm;
1250
1251         if (!sev_es_guest(vcpu->kvm))
1252                 return;
1253
1254         svm = to_svm(vcpu);
1255
1256         if (vcpu->arch.guest_state_protected)
1257                 sev_flush_guest_memory(svm, svm->vmsa, PAGE_SIZE);
1258         __free_page(virt_to_page(svm->vmsa));
1259 }
1260
1261 static void dump_ghcb(struct vcpu_svm *svm)
1262 {
1263         struct ghcb *ghcb = svm->ghcb;
1264         unsigned int nbits;
1265
1266         /* Re-use the dump_invalid_vmcb module parameter */
1267         if (!dump_invalid_vmcb) {
1268                 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
1269                 return;
1270         }
1271
1272         nbits = sizeof(ghcb->save.valid_bitmap) * 8;
1273
1274         pr_err("GHCB (GPA=%016llx):\n", svm->vmcb->control.ghcb_gpa);
1275         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
1276                ghcb->save.sw_exit_code, ghcb_sw_exit_code_is_valid(ghcb));
1277         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
1278                ghcb->save.sw_exit_info_1, ghcb_sw_exit_info_1_is_valid(ghcb));
1279         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
1280                ghcb->save.sw_exit_info_2, ghcb_sw_exit_info_2_is_valid(ghcb));
1281         pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
1282                ghcb->save.sw_scratch, ghcb_sw_scratch_is_valid(ghcb));
1283         pr_err("%-20s%*pb\n", "valid_bitmap", nbits, ghcb->save.valid_bitmap);
1284 }
1285
1286 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
1287 {
1288         struct kvm_vcpu *vcpu = &svm->vcpu;
1289         struct ghcb *ghcb = svm->ghcb;
1290
1291         /*
1292          * The GHCB protocol so far allows for the following data
1293          * to be returned:
1294          *   GPRs RAX, RBX, RCX, RDX
1295          *
1296          * Copy their values to the GHCB if they are dirty.
1297          */
1298         if (kvm_register_is_dirty(vcpu, VCPU_REGS_RAX))
1299                 ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
1300         if (kvm_register_is_dirty(vcpu, VCPU_REGS_RBX))
1301                 ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
1302         if (kvm_register_is_dirty(vcpu, VCPU_REGS_RCX))
1303                 ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
1304         if (kvm_register_is_dirty(vcpu, VCPU_REGS_RDX))
1305                 ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
1306 }
1307
1308 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
1309 {
1310         struct vmcb_control_area *control = &svm->vmcb->control;
1311         struct kvm_vcpu *vcpu = &svm->vcpu;
1312         struct ghcb *ghcb = svm->ghcb;
1313         u64 exit_code;
1314
1315         /*
1316          * The GHCB protocol so far allows for the following data
1317          * to be supplied:
1318          *   GPRs RAX, RBX, RCX, RDX
1319          *   XCR0
1320          *   CPL
1321          *
1322          * VMMCALL allows the guest to provide extra registers. KVM also
1323          * expects RSI for hypercalls, so include that, too.
1324          *
1325          * Copy their values to the appropriate location if supplied.
1326          */
1327         memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
1328
1329         vcpu->arch.regs[VCPU_REGS_RAX] = ghcb_get_rax_if_valid(ghcb);
1330         vcpu->arch.regs[VCPU_REGS_RBX] = ghcb_get_rbx_if_valid(ghcb);
1331         vcpu->arch.regs[VCPU_REGS_RCX] = ghcb_get_rcx_if_valid(ghcb);
1332         vcpu->arch.regs[VCPU_REGS_RDX] = ghcb_get_rdx_if_valid(ghcb);
1333         vcpu->arch.regs[VCPU_REGS_RSI] = ghcb_get_rsi_if_valid(ghcb);
1334
1335         svm->vmcb->save.cpl = ghcb_get_cpl_if_valid(ghcb);
1336
1337         if (ghcb_xcr0_is_valid(ghcb)) {
1338                 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
1339                 kvm_update_cpuid_runtime(vcpu);
1340         }
1341
1342         /* Copy the GHCB exit information into the VMCB fields */
1343         exit_code = ghcb_get_sw_exit_code(ghcb);
1344         control->exit_code = lower_32_bits(exit_code);
1345         control->exit_code_hi = upper_32_bits(exit_code);
1346         control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
1347         control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
1348
1349         /* Clear the valid entries fields */
1350         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
1351 }
1352
1353 static int sev_es_validate_vmgexit(struct vcpu_svm *svm)
1354 {
1355         struct kvm_vcpu *vcpu;
1356         struct ghcb *ghcb;
1357         u64 exit_code = 0;
1358
1359         ghcb = svm->ghcb;
1360
1361         /* Only GHCB Usage code 0 is supported */
1362         if (ghcb->ghcb_usage)
1363                 goto vmgexit_err;
1364
1365         /*
1366          * Retrieve the exit code now even though is may not be marked valid
1367          * as it could help with debugging.
1368          */
1369         exit_code = ghcb_get_sw_exit_code(ghcb);
1370
1371         if (!ghcb_sw_exit_code_is_valid(ghcb) ||
1372             !ghcb_sw_exit_info_1_is_valid(ghcb) ||
1373             !ghcb_sw_exit_info_2_is_valid(ghcb))
1374                 goto vmgexit_err;
1375
1376         switch (ghcb_get_sw_exit_code(ghcb)) {
1377         case SVM_EXIT_READ_DR7:
1378                 break;
1379         case SVM_EXIT_WRITE_DR7:
1380                 if (!ghcb_rax_is_valid(ghcb))
1381                         goto vmgexit_err;
1382                 break;
1383         case SVM_EXIT_RDTSC:
1384                 break;
1385         case SVM_EXIT_RDPMC:
1386                 if (!ghcb_rcx_is_valid(ghcb))
1387                         goto vmgexit_err;
1388                 break;
1389         case SVM_EXIT_CPUID:
1390                 if (!ghcb_rax_is_valid(ghcb) ||
1391                     !ghcb_rcx_is_valid(ghcb))
1392                         goto vmgexit_err;
1393                 if (ghcb_get_rax(ghcb) == 0xd)
1394                         if (!ghcb_xcr0_is_valid(ghcb))
1395                                 goto vmgexit_err;
1396                 break;
1397         case SVM_EXIT_INVD:
1398                 break;
1399         case SVM_EXIT_IOIO:
1400                 if (!(ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_TYPE_MASK))
1401                         if (!ghcb_rax_is_valid(ghcb))
1402                                 goto vmgexit_err;
1403                 break;
1404         case SVM_EXIT_MSR:
1405                 if (!ghcb_rcx_is_valid(ghcb))
1406                         goto vmgexit_err;
1407                 if (ghcb_get_sw_exit_info_1(ghcb)) {
1408                         if (!ghcb_rax_is_valid(ghcb) ||
1409                             !ghcb_rdx_is_valid(ghcb))
1410                                 goto vmgexit_err;
1411                 }
1412                 break;
1413         case SVM_EXIT_VMMCALL:
1414                 if (!ghcb_rax_is_valid(ghcb) ||
1415                     !ghcb_cpl_is_valid(ghcb))
1416                         goto vmgexit_err;
1417                 break;
1418         case SVM_EXIT_RDTSCP:
1419                 break;
1420         case SVM_EXIT_WBINVD:
1421                 break;
1422         case SVM_EXIT_MONITOR:
1423                 if (!ghcb_rax_is_valid(ghcb) ||
1424                     !ghcb_rcx_is_valid(ghcb) ||
1425                     !ghcb_rdx_is_valid(ghcb))
1426                         goto vmgexit_err;
1427                 break;
1428         case SVM_EXIT_MWAIT:
1429                 if (!ghcb_rax_is_valid(ghcb) ||
1430                     !ghcb_rcx_is_valid(ghcb))
1431                         goto vmgexit_err;
1432                 break;
1433         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
1434                 break;
1435         default:
1436                 goto vmgexit_err;
1437         }
1438
1439         return 0;
1440
1441 vmgexit_err:
1442         vcpu = &svm->vcpu;
1443
1444         if (ghcb->ghcb_usage) {
1445                 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
1446                             ghcb->ghcb_usage);
1447         } else {
1448                 vcpu_unimpl(vcpu, "vmgexit: exit reason %#llx is not valid\n",
1449                             exit_code);
1450                 dump_ghcb(svm);
1451         }
1452
1453         vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
1454         vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
1455         vcpu->run->internal.ndata = 2;
1456         vcpu->run->internal.data[0] = exit_code;
1457         vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
1458
1459         return -EINVAL;
1460 }
1461
1462 static void pre_sev_es_run(struct vcpu_svm *svm)
1463 {
1464         if (!svm->ghcb)
1465                 return;
1466
1467         sev_es_sync_to_ghcb(svm);
1468
1469         kvm_vcpu_unmap(&svm->vcpu, &svm->ghcb_map, true);
1470         svm->ghcb = NULL;
1471 }
1472
1473 void pre_sev_run(struct vcpu_svm *svm, int cpu)
1474 {
1475         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1476         int asid = sev_get_asid(svm->vcpu.kvm);
1477
1478         /* Perform any SEV-ES pre-run actions */
1479         pre_sev_es_run(svm);
1480
1481         /* Assign the asid allocated with this SEV guest */
1482         svm->asid = asid;
1483
1484         /*
1485          * Flush guest TLB:
1486          *
1487          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
1488          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
1489          */
1490         if (sd->sev_vmcbs[asid] == svm->vmcb &&
1491             svm->vcpu.arch.last_vmentry_cpu == cpu)
1492                 return;
1493
1494         sd->sev_vmcbs[asid] = svm->vmcb;
1495         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
1496         vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
1497 }
1498
1499 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
1500 {
1501         return -EINVAL;
1502 }
1503
1504 int sev_handle_vmgexit(struct vcpu_svm *svm)
1505 {
1506         struct vmcb_control_area *control = &svm->vmcb->control;
1507         u64 ghcb_gpa, exit_code;
1508         struct ghcb *ghcb;
1509         int ret;
1510
1511         /* Validate the GHCB */
1512         ghcb_gpa = control->ghcb_gpa;
1513         if (ghcb_gpa & GHCB_MSR_INFO_MASK)
1514                 return sev_handle_vmgexit_msr_protocol(svm);
1515
1516         if (!ghcb_gpa) {
1517                 vcpu_unimpl(&svm->vcpu, "vmgexit: GHCB gpa is not set\n");
1518                 return -EINVAL;
1519         }
1520
1521         if (kvm_vcpu_map(&svm->vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->ghcb_map)) {
1522                 /* Unable to map GHCB from guest */
1523                 vcpu_unimpl(&svm->vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
1524                             ghcb_gpa);
1525                 return -EINVAL;
1526         }
1527
1528         svm->ghcb = svm->ghcb_map.hva;
1529         ghcb = svm->ghcb_map.hva;
1530
1531         exit_code = ghcb_get_sw_exit_code(ghcb);
1532
1533         ret = sev_es_validate_vmgexit(svm);
1534         if (ret)
1535                 return ret;
1536
1537         sev_es_sync_from_ghcb(svm);
1538         ghcb_set_sw_exit_info_1(ghcb, 0);
1539         ghcb_set_sw_exit_info_2(ghcb, 0);
1540
1541         ret = -EINVAL;
1542         switch (exit_code) {
1543         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
1544                 vcpu_unimpl(&svm->vcpu,
1545                             "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
1546                             control->exit_info_1, control->exit_info_2);
1547                 break;
1548         default:
1549                 ret = svm_invoke_exit_handler(svm, exit_code);
1550         }
1551
1552         return ret;
1553 }