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
  18 #include "x86.h"
  19 #include "svm.h"
  20
  21 static int sev_flush_asids(void);
  22 static DECLARE_RWSEM(sev_deactivate_lock);
  23 static DEFINE_MUTEX(sev_bitmap_lock);
  24 unsigned int max_sev_asid;
  25 static unsigned int min_sev_asid;
  26 static unsigned long *sev_asid_bitmap;
  27 static unsigned long *sev_reclaim_asid_bitmap;
  28 #define __sme_page_pa(x) __sme_set(page_to_pfn(x) << PAGE_SHIFT)
  29
  30 struct enc_region {
  31         struct list_head list;
  32         unsigned long npages;
  33         struct page **pages;
  34         unsigned long uaddr;
  35         unsigned long size;
  36 };
  37
  38 static int sev_flush_asids(void)
  39 {
  40         int ret, error = 0;
  41
  42         /*
  43          * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
  44          * so it must be guarded.
  45          */
  46         down_write(&sev_deactivate_lock);
  47
  48         wbinvd_on_all_cpus();
  49         ret = sev_guest_df_flush(&error);
  50
  51         up_write(&sev_deactivate_lock);
  52
  53         if (ret)
  54                 pr_err("SEV: DF_FLUSH failed, ret=%d, error=%#x\n", ret, error);
  55
  56         return ret;
  57 }
  58
  59 /* Must be called with the sev_bitmap_lock held */
  60 static bool __sev_recycle_asids(void)
  61 {
  62         int pos;
  63
  64         /* Check if there are any ASIDs to reclaim before performing a flush */
  65         pos = find_next_bit(sev_reclaim_asid_bitmap,
  66                             max_sev_asid, min_sev_asid - 1);
  67         if (pos >= max_sev_asid)
  68                 return false;
  69
  70         if (sev_flush_asids())
  71                 return false;
  72
  73         bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
  74                    max_sev_asid);
  75         bitmap_zero(sev_reclaim_asid_bitmap, max_sev_asid);
  76
  77         return true;
  78 }
  79
  80 static int sev_asid_new(void)
  81 {
  82         bool retry = true;
  83         int pos;
  84
  85         mutex_lock(&sev_bitmap_lock);
  86
  87         /*
  88          * SEV-enabled guest must use asid from min_sev_asid to max_sev_asid.
  89          */
  90 again:
  91         pos = find_next_zero_bit(sev_asid_bitmap, max_sev_asid, min_sev_asid - 1);
  92         if (pos >= max_sev_asid) {
  93                 if (retry && __sev_recycle_asids()) {
  94                         retry = false;
  95                         goto again;
  96                 }
  97                 mutex_unlock(&sev_bitmap_lock);
  98                 return -EBUSY;
  99         }
 100
 101         __set_bit(pos, sev_asid_bitmap);
 102
 103         mutex_unlock(&sev_bitmap_lock);
 104
 105         return pos + 1;
 106 }
 107
 108 static int sev_get_asid(struct kvm *kvm)
 109 {
 110         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 111
 112         return sev->asid;
 113 }
 114
 115 static void sev_asid_free(int asid)
 116 {
 117         struct svm_cpu_data *sd;
 118         int cpu, pos;
 119
 120         mutex_lock(&sev_bitmap_lock);
 121
 122         pos = asid - 1;
 123         __set_bit(pos, sev_reclaim_asid_bitmap);
 124
 125         for_each_possible_cpu(cpu) {
 126                 sd = per_cpu(svm_data, cpu);
 127                 sd->sev_vmcbs[pos] = NULL;
 128         }
 129
 130         mutex_unlock(&sev_bitmap_lock);
 131 }
 132
 133 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
 134 {
 135         struct sev_data_decommission *decommission;
 136         struct sev_data_deactivate *data;
 137
 138         if (!handle)
 139                 return;
 140
 141         data = kzalloc(sizeof(*data), GFP_KERNEL);
 142         if (!data)
 143                 return;
 144
 145         /* deactivate handle */
 146         data->handle = handle;
 147
 148         /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
 149         down_read(&sev_deactivate_lock);
 150         sev_guest_deactivate(data, NULL);
 151         up_read(&sev_deactivate_lock);
 152
 153         kfree(data);
 154
 155         decommission = kzalloc(sizeof(*decommission), GFP_KERNEL);
 156         if (!decommission)
 157                 return;
 158
 159         /* decommission handle */
 160         decommission->handle = handle;
 161         sev_guest_decommission(decommission, NULL);
 162
 163         kfree(decommission);
 164 }
 165
 166 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 167 {
 168         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 169         int asid, ret;
 170
 171         ret = -EBUSY;
 172         if (unlikely(sev->active))
 173                 return ret;
 174
 175         asid = sev_asid_new();
 176         if (asid < 0)
 177                 return ret;
 178
 179         ret = sev_platform_init(&argp->error);
 180         if (ret)
 181                 goto e_free;
 182
 183         sev->active = true;
 184         sev->asid = asid;
 185         INIT_LIST_HEAD(&sev->regions_list);
 186
 187         return 0;
 188
 189 e_free:
 190         sev_asid_free(asid);
 191         return ret;
 192 }
 193
 194 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
 195 {
 196         struct sev_data_activate *data;
 197         int asid = sev_get_asid(kvm);
 198         int ret;
 199
 200         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 201         if (!data)
 202                 return -ENOMEM;
 203
 204         /* activate ASID on the given handle */
 205         data->handle = handle;
 206         data->asid   = asid;
 207         ret = sev_guest_activate(data, error);
 208         kfree(data);
 209
 210         return ret;
 211 }
 212
 213 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
 214 {
 215         struct fd f;
 216         int ret;
 217
 218         f = fdget(fd);
 219         if (!f.file)
 220                 return -EBADF;
 221
 222         ret = sev_issue_cmd_external_user(f.file, id, data, error);
 223
 224         fdput(f);
 225         return ret;
 226 }
 227
 228 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
 229 {
 230         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 231
 232         return __sev_issue_cmd(sev->fd, id, data, error);
 233 }
 234
 235 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
 236 {
 237         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 238         struct sev_data_launch_start *start;
 239         struct kvm_sev_launch_start params;
 240         void *dh_blob, *session_blob;
 241         int *error = &argp->error;
 242         int ret;
 243
 244         if (!sev_guest(kvm))
 245                 return -ENOTTY;
 246
 247         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 248                 return -EFAULT;
 249
 250         start = kzalloc(sizeof(*start), GFP_KERNEL_ACCOUNT);
 251         if (!start)
 252                 return -ENOMEM;
 253
 254         dh_blob = NULL;
 255         if (params.dh_uaddr) {
 256                 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
 257                 if (IS_ERR(dh_blob)) {
 258                         ret = PTR_ERR(dh_blob);
 259                         goto e_free;
 260                 }
 261
 262                 start->dh_cert_address = __sme_set(__pa(dh_blob));
 263                 start->dh_cert_len = params.dh_len;
 264         }
 265
 266         session_blob = NULL;
 267         if (params.session_uaddr) {
 268                 session_blob = psp_copy_user_blob(params.session_uaddr, params.session_len);
 269                 if (IS_ERR(session_blob)) {
 270                         ret = PTR_ERR(session_blob);
 271                         goto e_free_dh;
 272                 }
 273
 274                 start->session_address = __sme_set(__pa(session_blob));
 275                 start->session_len = params.session_len;
 276         }
 277
 278         start->handle = params.handle;
 279         start->policy = params.policy;
 280
 281         /* create memory encryption context */
 282         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_LAUNCH_START, start, error);
 283         if (ret)
 284                 goto e_free_session;
 285
 286         /* Bind ASID to this guest */
 287         ret = sev_bind_asid(kvm, start->handle, error);
 288         if (ret)
 289                 goto e_free_session;
 290
 291         /* return handle to userspace */
 292         params.handle = start->handle;
 293         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params))) {
 294                 sev_unbind_asid(kvm, start->handle);
 295                 ret = -EFAULT;
 296                 goto e_free_session;
 297         }
 298
 299         sev->handle = start->handle;
 300         sev->fd = argp->sev_fd;
 301
 302 e_free_session:
 303         kfree(session_blob);
 304 e_free_dh:
 305         kfree(dh_blob);
 306 e_free:
 307         kfree(start);
 308         return ret;
 309 }
 310
 311 static struct page **sev_pin_memory(struct kvm *kvm, unsigned long uaddr,
 312                                     unsigned long ulen, unsigned long *n,
 313                                     int write)
 314 {
 315         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 316         unsigned long npages, size;
 317         int npinned;
 318         unsigned long locked, lock_limit;
 319         struct page **pages;
 320         unsigned long first, last;
 321
 322         if (ulen == 0 || uaddr + ulen < uaddr)
 323                 return ERR_PTR(-EINVAL);
 324
 325         /* Calculate number of pages. */
 326         first = (uaddr & PAGE_MASK) >> PAGE_SHIFT;
 327         last = ((uaddr + ulen - 1) & PAGE_MASK) >> PAGE_SHIFT;
 328         npages = (last - first + 1);
 329
 330         locked = sev->pages_locked + npages;
 331         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 332         if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
 333                 pr_err("SEV: %lu locked pages exceed the lock limit of %lu.\n", locked, lock_limit);
 334                 return ERR_PTR(-ENOMEM);
 335         }
 336
 337         if (WARN_ON_ONCE(npages > INT_MAX))
 338                 return ERR_PTR(-EINVAL);
 339
 340         /* Avoid using vmalloc for smaller buffers. */
 341         size = npages * sizeof(struct page *);
 342         if (size > PAGE_SIZE)
 343                 pages = __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 344         else
 345                 pages = kmalloc(size, GFP_KERNEL_ACCOUNT);
 346
 347         if (!pages)
 348                 return ERR_PTR(-ENOMEM);
 349
 350         /* Pin the user virtual address. */
 351         npinned = pin_user_pages_fast(uaddr, npages, write ? FOLL_WRITE : 0, pages);
 352         if (npinned != npages) {
 353                 pr_err("SEV: Failure locking %lu pages.\n", npages);
 354                 goto err;
 355         }
 356
 357         *n = npages;
 358         sev->pages_locked = locked;
 359
 360         return pages;
 361
 362 err:
 363         if (npinned > 0) {
 364                 unpin_user_pages(pages, npinned);
 365                 npinned = -ENOMEM;
 366         }
 367
 368         kvfree(pages);
 369         return ERR_PTR(npinned);
 370 }
 371
 372 static void sev_unpin_memory(struct kvm *kvm, struct page **pages,
 373                              unsigned long npages)
 374 {
 375         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 376
 377         unpin_user_pages(pages, npages);
 378         kvfree(pages);
 379         sev->pages_locked -= npages;
 380 }
 381
 382 static void sev_clflush_pages(struct page *pages[], unsigned long npages)
 383 {
 384         uint8_t *page_virtual;
 385         unsigned long i;
 386
 387         if (npages == 0 || pages == NULL)
 388                 return;
 389
 390         for (i = 0; i < npages; i++) {
 391                 page_virtual = kmap_atomic(pages[i]);
 392                 clflush_cache_range(page_virtual, PAGE_SIZE);
 393                 kunmap_atomic(page_virtual);
 394         }
 395 }
 396
 397 static unsigned long get_num_contig_pages(unsigned long idx,
 398                                 struct page **inpages, unsigned long npages)
 399 {
 400         unsigned long paddr, next_paddr;
 401         unsigned long i = idx + 1, pages = 1;
 402
 403         /* find the number of contiguous pages starting from idx */
 404         paddr = __sme_page_pa(inpages[idx]);
 405         while (i < npages) {
 406                 next_paddr = __sme_page_pa(inpages[i++]);
 407                 if ((paddr + PAGE_SIZE) == next_paddr) {
 408                         pages++;
 409                         paddr = next_paddr;
 410                         continue;
 411                 }
 412                 break;
 413         }
 414
 415         return pages;
 416 }
 417
 418 static int sev_launch_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
 419 {
 420         unsigned long vaddr, vaddr_end, next_vaddr, npages, pages, size, i;
 421         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 422         struct kvm_sev_launch_update_data params;
 423         struct sev_data_launch_update_data *data;
 424         struct page **inpages;
 425         int ret;
 426
 427         if (!sev_guest(kvm))
 428                 return -ENOTTY;
 429
 430         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 431                 return -EFAULT;
 432
 433         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 434         if (!data)
 435                 return -ENOMEM;
 436
 437         vaddr = params.uaddr;
 438         size = params.len;
 439         vaddr_end = vaddr + size;
 440
 441         /* Lock the user memory. */
 442         inpages = sev_pin_memory(kvm, vaddr, size, &npages, 1);
 443         if (!inpages) {
 444                 ret = -ENOMEM;
 445                 goto e_free;
 446         }
 447
 448         /*
 449          * The LAUNCH_UPDATE command will perform in-place encryption of the
 450          * memory content (i.e it will write the same memory region with C=1).
 451          * It's possible that the cache may contain the data with C=0, i.e.,
 452          * unencrypted so invalidate it first.
 453          */
 454         sev_clflush_pages(inpages, npages);
 455
 456         for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
 457                 int offset, len;
 458
 459                 /*
 460                  * If the user buffer is not page-aligned, calculate the offset
 461                  * within the page.
 462                  */
 463                 offset = vaddr & (PAGE_SIZE - 1);
 464
 465                 /* Calculate the number of pages that can be encrypted in one go. */
 466                 pages = get_num_contig_pages(i, inpages, npages);
 467
 468                 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
 469
 470                 data->handle = sev->handle;
 471                 data->len = len;
 472                 data->address = __sme_page_pa(inpages[i]) + offset;
 473                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, data, &argp->error);
 474                 if (ret)
 475                         goto e_unpin;
 476
 477                 size -= len;
 478                 next_vaddr = vaddr + len;
 479         }
 480
 481 e_unpin:
 482         /* content of memory is updated, mark pages dirty */
 483         for (i = 0; i < npages; i++) {
 484                 set_page_dirty_lock(inpages[i]);
 485                 mark_page_accessed(inpages[i]);
 486         }
 487         /* unlock the user pages */
 488         sev_unpin_memory(kvm, inpages, npages);
 489 e_free:
 490         kfree(data);
 491         return ret;
 492 }
 493
 494 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 495 {
 496         void __user *measure = (void __user *)(uintptr_t)argp->data;
 497         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 498         struct sev_data_launch_measure *data;
 499         struct kvm_sev_launch_measure params;
 500         void __user *p = NULL;
 501         void *blob = NULL;
 502         int ret;
 503
 504         if (!sev_guest(kvm))
 505                 return -ENOTTY;
 506
 507         if (copy_from_user(&params, measure, sizeof(params)))
 508                 return -EFAULT;
 509
 510         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 511         if (!data)
 512                 return -ENOMEM;
 513
 514         /* User wants to query the blob length */
 515         if (!params.len)
 516                 goto cmd;
 517
 518         p = (void __user *)(uintptr_t)params.uaddr;
 519         if (p) {
 520                 if (params.len > SEV_FW_BLOB_MAX_SIZE) {
 521                         ret = -EINVAL;
 522                         goto e_free;
 523                 }
 524
 525                 ret = -ENOMEM;
 526                 blob = kmalloc(params.len, GFP_KERNEL);
 527                 if (!blob)
 528                         goto e_free;
 529
 530                 data->address = __psp_pa(blob);
 531                 data->len = params.len;
 532         }
 533
 534 cmd:
 535         data->handle = sev->handle;
 536         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, data, &argp->error);
 537
 538         /*
 539          * If we query the session length, FW responded with expected data.
 540          */
 541         if (!params.len)
 542                 goto done;
 543
 544         if (ret)
 545                 goto e_free_blob;
 546
 547         if (blob) {
 548                 if (copy_to_user(p, blob, params.len))
 549                         ret = -EFAULT;
 550         }
 551
 552 done:
 553         params.len = data->len;
 554         if (copy_to_user(measure, &params, sizeof(params)))
 555                 ret = -EFAULT;
 556 e_free_blob:
 557         kfree(blob);
 558 e_free:
 559         kfree(data);
 560         return ret;
 561 }
 562
 563 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 564 {
 565         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 566         struct sev_data_launch_finish *data;
 567         int ret;
 568
 569         if (!sev_guest(kvm))
 570                 return -ENOTTY;
 571
 572         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 573         if (!data)
 574                 return -ENOMEM;
 575
 576         data->handle = sev->handle;
 577         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, data, &argp->error);
 578
 579         kfree(data);
 580         return ret;
 581 }
 582
 583 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 584 {
 585         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 586         struct kvm_sev_guest_status params;
 587         struct sev_data_guest_status *data;
 588         int ret;
 589
 590         if (!sev_guest(kvm))
 591                 return -ENOTTY;
 592
 593         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 594         if (!data)
 595                 return -ENOMEM;
 596
 597         data->handle = sev->handle;
 598         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, data, &argp->error);
 599         if (ret)
 600                 goto e_free;
 601
 602         params.policy = data->policy;
 603         params.state = data->state;
 604         params.handle = data->handle;
 605
 606         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 607                 ret = -EFAULT;
 608 e_free:
 609         kfree(data);
 610         return ret;
 611 }
 612
 613 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 614                                unsigned long dst, int size,
 615                                int *error, bool enc)
 616 {
 617         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 618         struct sev_data_dbg *data;
 619         int ret;
 620
 621         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 622         if (!data)
 623                 return -ENOMEM;
 624
 625         data->handle = sev->handle;
 626         data->dst_addr = dst;
 627         data->src_addr = src;
 628         data->len = size;
 629
 630         ret = sev_issue_cmd(kvm,
 631                             enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 632                             data, error);
 633         kfree(data);
 634         return ret;
 635 }
 636
 637 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 638                              unsigned long dst_paddr, int sz, int *err)
 639 {
 640         int offset;
 641
 642         /*
 643          * Its safe to read more than we are asked, caller should ensure that
 644          * destination has enough space.
 645          */
 646         src_paddr = round_down(src_paddr, 16);
 647         offset = src_paddr & 15;
 648         sz = round_up(sz + offset, 16);
 649
 650         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 651 }
 652
 653 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 654                                   unsigned long __user dst_uaddr,
 655                                   unsigned long dst_paddr,
 656                                   int size, int *err)
 657 {
 658         struct page *tpage = NULL;
 659         int ret, offset;
 660
 661         /* if inputs are not 16-byte then use intermediate buffer */
 662         if (!IS_ALIGNED(dst_paddr, 16) ||
 663             !IS_ALIGNED(paddr,     16) ||
 664             !IS_ALIGNED(size,      16)) {
 665                 tpage = (void *)alloc_page(GFP_KERNEL);
 666                 if (!tpage)
 667                         return -ENOMEM;
 668
 669                 dst_paddr = __sme_page_pa(tpage);
 670         }
 671
 672         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 673         if (ret)
 674                 goto e_free;
 675
 676         if (tpage) {
 677                 offset = paddr & 15;
 678                 if (copy_to_user((void __user *)(uintptr_t)dst_uaddr,
 679                                  page_address(tpage) + offset, size))
 680                         ret = -EFAULT;
 681         }
 682
 683 e_free:
 684         if (tpage)
 685                 __free_page(tpage);
 686
 687         return ret;
 688 }
 689
 690 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 691                                   unsigned long __user vaddr,
 692                                   unsigned long dst_paddr,
 693                                   unsigned long __user dst_vaddr,
 694                                   int size, int *error)
 695 {
 696         struct page *src_tpage = NULL;
 697         struct page *dst_tpage = NULL;
 698         int ret, len = size;
 699
 700         /* If source buffer is not aligned then use an intermediate buffer */
 701         if (!IS_ALIGNED(vaddr, 16)) {
 702                 src_tpage = alloc_page(GFP_KERNEL);
 703                 if (!src_tpage)
 704                         return -ENOMEM;
 705
 706                 if (copy_from_user(page_address(src_tpage),
 707                                 (void __user *)(uintptr_t)vaddr, size)) {
 708                         __free_page(src_tpage);
 709                         return -EFAULT;
 710                 }
 711
 712                 paddr = __sme_page_pa(src_tpage);
 713         }
 714
 715         /*
 716          *  If destination buffer or length is not aligned then do read-modify-write:
 717          *   - decrypt destination in an intermediate buffer
 718          *   - copy the source buffer in an intermediate buffer
 719          *   - use the intermediate buffer as source buffer
 720          */
 721         if (!IS_ALIGNED(dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 722                 int dst_offset;
 723
 724                 dst_tpage = alloc_page(GFP_KERNEL);
 725                 if (!dst_tpage) {
 726                         ret = -ENOMEM;
 727                         goto e_free;
 728                 }
 729
 730                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 731                                         __sme_page_pa(dst_tpage), size, error);
 732                 if (ret)
 733                         goto e_free;
 734
 735                 /*
 736                  *  If source is kernel buffer then use memcpy() otherwise
 737                  *  copy_from_user().
 738                  */
 739                 dst_offset = dst_paddr & 15;
 740
 741                 if (src_tpage)
 742                         memcpy(page_address(dst_tpage) + dst_offset,
 743                                page_address(src_tpage), size);
 744                 else {
 745                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 746                                            (void __user *)(uintptr_t)vaddr, size)) {
 747                                 ret = -EFAULT;
 748                                 goto e_free;
 749                         }
 750                 }
 751
 752                 paddr = __sme_page_pa(dst_tpage);
 753                 dst_paddr = round_down(dst_paddr, 16);
 754                 len = round_up(size, 16);
 755         }
 756
 757         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 758
 759 e_free:
 760         if (src_tpage)
 761                 __free_page(src_tpage);
 762         if (dst_tpage)
 763                 __free_page(dst_tpage);
 764         return ret;
 765 }
 766
 767 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 768 {
 769         unsigned long vaddr, vaddr_end, next_vaddr;
 770         unsigned long dst_vaddr;
 771         struct page **src_p, **dst_p;
 772         struct kvm_sev_dbg debug;
 773         unsigned long n;
 774         unsigned int size;
 775         int ret;
 776
 777         if (!sev_guest(kvm))
 778                 return -ENOTTY;
 779
 780         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 781                 return -EFAULT;
 782
 783         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 784                 return -EINVAL;
 785         if (!debug.dst_uaddr)
 786                 return -EINVAL;
 787
 788         vaddr = debug.src_uaddr;
 789         size = debug.len;
 790         vaddr_end = vaddr + size;
 791         dst_vaddr = debug.dst_uaddr;
 792
 793         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 794                 int len, s_off, d_off;
 795
 796                 /* lock userspace source and destination page */
 797                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 798                 if (!src_p)
 799                         return -EFAULT;
 800
 801                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 802                 if (!dst_p) {
 803                         sev_unpin_memory(kvm, src_p, n);
 804                         return -EFAULT;
 805                 }
 806
 807                 /*
 808                  * The DBG_{DE,EN}CRYPT commands will perform {dec,en}cryption of the
 809                  * memory content (i.e it will write the same memory region with C=1).
 810                  * It's possible that the cache may contain the data with C=0, i.e.,
 811                  * unencrypted so invalidate it first.
 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;
 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          * The secret must be copied into contiguous memory region, lets verify
 874          * that userspace memory pages are contiguous before we issue command.
 875          */
 876         if (get_num_contig_pages(0, pages, n) != n) {
 877                 ret = -EINVAL;
 878                 goto e_unpin_memory;
 879         }
 880
 881         ret = -ENOMEM;
 882         data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
 883         if (!data)
 884                 goto e_unpin_memory;
 885
 886         offset = params.guest_uaddr & (PAGE_SIZE - 1);
 887         data->guest_address = __sme_page_pa(pages[0]) + offset;
 888         data->guest_len = params.guest_len;
 889
 890         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
 891         if (IS_ERR(blob)) {
 892                 ret = PTR_ERR(blob);
 893                 goto e_free;
 894         }
 895
 896         data->trans_address = __psp_pa(blob);
 897         data->trans_len = params.trans_len;
 898
 899         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
 900         if (IS_ERR(hdr)) {
 901                 ret = PTR_ERR(hdr);
 902                 goto e_free_blob;
 903         }
 904         data->hdr_address = __psp_pa(hdr);
 905         data->hdr_len = params.hdr_len;
 906
 907         data->handle = sev->handle;
 908         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, data, &argp->error);
 909
 910         kfree(hdr);
 911
 912 e_free_blob:
 913         kfree(blob);
 914 e_free:
 915         kfree(data);
 916 e_unpin_memory:
 917         sev_unpin_memory(kvm, pages, n);
 918         return ret;
 919 }
 920
 921 int svm_mem_enc_op(struct kvm *kvm, void __user *argp)
 922 {
 923         struct kvm_sev_cmd sev_cmd;
 924         int r;
 925
 926         if (!svm_sev_enabled())
 927                 return -ENOTTY;
 928
 929         if (!argp)
 930                 return 0;
 931
 932         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
 933                 return -EFAULT;
 934
 935         mutex_lock(&kvm->lock);
 936
 937         switch (sev_cmd.id) {
 938         case KVM_SEV_INIT:
 939                 r = sev_guest_init(kvm, &sev_cmd);
 940                 break;
 941         case KVM_SEV_LAUNCH_START:
 942                 r = sev_launch_start(kvm, &sev_cmd);
 943                 break;
 944         case KVM_SEV_LAUNCH_UPDATE_DATA:
 945                 r = sev_launch_update_data(kvm, &sev_cmd);
 946                 break;
 947         case KVM_SEV_LAUNCH_MEASURE:
 948                 r = sev_launch_measure(kvm, &sev_cmd);
 949                 break;
 950         case KVM_SEV_LAUNCH_FINISH:
 951                 r = sev_launch_finish(kvm, &sev_cmd);
 952                 break;
 953         case KVM_SEV_GUEST_STATUS:
 954                 r = sev_guest_status(kvm, &sev_cmd);
 955                 break;
 956         case KVM_SEV_DBG_DECRYPT:
 957                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
 958                 break;
 959         case KVM_SEV_DBG_ENCRYPT:
 960                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
 961                 break;
 962         case KVM_SEV_LAUNCH_SECRET:
 963                 r = sev_launch_secret(kvm, &sev_cmd);
 964                 break;
 965         default:
 966                 r = -EINVAL;
 967                 goto out;
 968         }
 969
 970         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
 971                 r = -EFAULT;
 972
 973 out:
 974         mutex_unlock(&kvm->lock);
 975         return r;
 976 }
 977
 978 int svm_register_enc_region(struct kvm *kvm,
 979                             struct kvm_enc_region *range)
 980 {
 981         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 982         struct enc_region *region;
 983         int ret = 0;
 984
 985         if (!sev_guest(kvm))
 986                 return -ENOTTY;
 987
 988         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
 989                 return -EINVAL;
 990
 991         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
 992         if (!region)
 993                 return -ENOMEM;
 994
 995         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
 996         if (IS_ERR(region->pages)) {
 997                 ret = PTR_ERR(region->pages);
 998                 goto e_free;
 999         }
1000
1001         /*
1002          * The guest may change the memory encryption attribute from C=0 -> C=1
1003          * or vice versa for this memory range. Lets make sure caches are
1004          * flushed to ensure that guest data gets written into memory with
1005          * correct C-bit.
1006          */
1007         sev_clflush_pages(region->pages, region->npages);
1008
1009         region->uaddr = range->addr;
1010         region->size = range->size;
1011
1012         mutex_lock(&kvm->lock);
1013         list_add_tail(&region->list, &sev->regions_list);
1014         mutex_unlock(&kvm->lock);
1015
1016         return ret;
1017
1018 e_free:
1019         kfree(region);
1020         return ret;
1021 }
1022
1023 static struct enc_region *
1024 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
1025 {
1026         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1027         struct list_head *head = &sev->regions_list;
1028         struct enc_region *i;
1029
1030         list_for_each_entry(i, head, list) {
1031                 if (i->uaddr == range->addr &&
1032                     i->size == range->size)
1033                         return i;
1034         }
1035
1036         return NULL;
1037 }
1038
1039 static void __unregister_enc_region_locked(struct kvm *kvm,
1040                                            struct enc_region *region)
1041 {
1042         sev_unpin_memory(kvm, region->pages, region->npages);
1043         list_del(&region->list);
1044         kfree(region);
1045 }
1046
1047 int svm_unregister_enc_region(struct kvm *kvm,
1048                               struct kvm_enc_region *range)
1049 {
1050         struct enc_region *region;
1051         int ret;
1052
1053         mutex_lock(&kvm->lock);
1054
1055         if (!sev_guest(kvm)) {
1056                 ret = -ENOTTY;
1057                 goto failed;
1058         }
1059
1060         region = find_enc_region(kvm, range);
1061         if (!region) {
1062                 ret = -EINVAL;
1063                 goto failed;
1064         }
1065
1066         /*
1067          * Ensure that all guest tagged cache entries are flushed before
1068          * releasing the pages back to the system for use. CLFLUSH will
1069          * not do this, so issue a WBINVD.
1070          */
1071         wbinvd_on_all_cpus();
1072
1073         __unregister_enc_region_locked(kvm, region);
1074
1075         mutex_unlock(&kvm->lock);
1076         return 0;
1077
1078 failed:
1079         mutex_unlock(&kvm->lock);
1080         return ret;
1081 }
1082
1083 void sev_vm_destroy(struct kvm *kvm)
1084 {
1085         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1086         struct list_head *head = &sev->regions_list;
1087         struct list_head *pos, *q;
1088
1089         if (!sev_guest(kvm))
1090                 return;
1091
1092         mutex_lock(&kvm->lock);
1093
1094         /*
1095          * Ensure that all guest tagged cache entries are flushed before
1096          * releasing the pages back to the system for use. CLFLUSH will
1097          * not do this, so issue a WBINVD.
1098          */
1099         wbinvd_on_all_cpus();
1100
1101         /*
1102          * if userspace was terminated before unregistering the memory regions
1103          * then lets unpin all the registered memory.
1104          */
1105         if (!list_empty(head)) {
1106                 list_for_each_safe(pos, q, head) {
1107                         __unregister_enc_region_locked(kvm,
1108                                 list_entry(pos, struct enc_region, list));
1109                 }
1110         }
1111
1112         mutex_unlock(&kvm->lock);
1113
1114         sev_unbind_asid(kvm, sev->handle);
1115         sev_asid_free(sev->asid);
1116 }
1117
1118 int __init sev_hardware_setup(void)
1119 {
1120         struct sev_user_data_status *status;
1121         int rc;
1122
1123         /* Maximum number of encrypted guests supported simultaneously */
1124         max_sev_asid = cpuid_ecx(0x8000001F);
1125
1126         if (!svm_sev_enabled())
1127                 return 1;
1128
1129         /* Minimum ASID value that should be used for SEV guest */
1130         min_sev_asid = cpuid_edx(0x8000001F);
1131
1132         /* Initialize SEV ASID bitmaps */
1133         sev_asid_bitmap = bitmap_zalloc(max_sev_asid, GFP_KERNEL);
1134         if (!sev_asid_bitmap)
1135                 return 1;
1136
1137         sev_reclaim_asid_bitmap = bitmap_zalloc(max_sev_asid, GFP_KERNEL);
1138         if (!sev_reclaim_asid_bitmap)
1139                 return 1;
1140
1141         status = kmalloc(sizeof(*status), GFP_KERNEL);
1142         if (!status)
1143                 return 1;
1144
1145         /*
1146          * Check SEV platform status.
1147          *
1148          * PLATFORM_STATUS can be called in any state, if we failed to query
1149          * the PLATFORM status then either PSP firmware does not support SEV
1150          * feature or SEV firmware is dead.
1151          */
1152         rc = sev_platform_status(status, NULL);
1153         if (rc)
1154                 goto err;
1155
1156         pr_info("SEV supported\n");
1157
1158 err:
1159         kfree(status);
1160         return rc;
1161 }
1162
1163 void sev_hardware_teardown(void)
1164 {
1165         if (!svm_sev_enabled())
1166                 return;
1167
1168         bitmap_free(sev_asid_bitmap);
1169         bitmap_free(sev_reclaim_asid_bitmap);
1170
1171         sev_flush_asids();
1172 }
1173
1174 void pre_sev_run(struct vcpu_svm *svm, int cpu)
1175 {
1176         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
1177         int asid = sev_get_asid(svm->vcpu.kvm);
1178
1179         /* Assign the asid allocated with this SEV guest */
1180         svm->vmcb->control.asid = asid;
1181
1182         /*
1183          * Flush guest TLB:
1184          *
1185          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
1186          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
1187          */
1188         if (sd->sev_vmcbs[asid] == svm->vmcb &&
1189             svm->vcpu.arch.last_vmentry_cpu == cpu)
1190                 return;
1191
1192         sd->sev_vmcbs[asid] = svm->vmcb;
1193         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
1194         mark_dirty(svm->vmcb, VMCB_ASID);
1195 }