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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11 #include <linux/kvm_types.h>
12 #include <linux/kvm_host.h>
13 #include <linux/kernel.h>
14 #include <linux/highmem.h>
15 #include <linux/psp.h>
16 #include <linux/psp-sev.h>
17 #include <linux/pagemap.h>
18 #include <linux/swap.h>
19 #include <linux/misc_cgroup.h>
20 #include <linux/processor.h>
21 #include <linux/trace_events.h>
22 #include <uapi/linux/sev-guest.h>
23
24 #include <asm/pkru.h>
25 #include <asm/trapnr.h>
26 #include <asm/fpu/xcr.h>
27 #include <asm/fpu/xstate.h>
28 #include <asm/debugreg.h>
29 #include <asm/msr.h>
30 #include <asm/sev.h>
31
32 #include "mmu.h"
33 #include "x86.h"
34 #include "svm.h"
35 #include "svm_ops.h"
36 #include "cpuid.h"
37 #include "trace.h"
38
39 #define GHCB_VERSION_MAX 2ULL
40 #define GHCB_VERSION_DEFAULT 2ULL
41 #define GHCB_VERSION_MIN 1ULL
42
43 #define GHCB_HV_FT_SUPPORTED (GHCB_HV_FT_SNP | GHCB_HV_FT_SNP_AP_CREATION)
44
45 /* enable/disable SEV support */
46 static bool sev_enabled = true;
47 module_param_named(sev, sev_enabled, bool, 0444);
48
49 /* enable/disable SEV-ES support */
50 static bool sev_es_enabled = true;
51 module_param_named(sev_es, sev_es_enabled, bool, 0444);
52
53 /* enable/disable SEV-SNP support */
54 static bool sev_snp_enabled = true;
55 module_param_named(sev_snp, sev_snp_enabled, bool, 0444);
56
57 /* enable/disable SEV-ES DebugSwap support */
58 static bool sev_es_debug_swap_enabled = true;
59 module_param_named(debug_swap, sev_es_debug_swap_enabled, bool, 0444);
60 static u64 sev_supported_vmsa_features;
61
62 #define AP_RESET_HOLD_NONE 0
63 #define AP_RESET_HOLD_NAE_EVENT 1
64 #define AP_RESET_HOLD_MSR_PROTO 2
65
66 /* As defined by SEV-SNP Firmware ABI, under "Guest Policy". */
67 #define SNP_POLICY_MASK_API_MINOR GENMASK_ULL(7, 0)
68 #define SNP_POLICY_MASK_API_MAJOR GENMASK_ULL(15, 8)
69 #define SNP_POLICY_MASK_SMT BIT_ULL(16)
70 #define SNP_POLICY_MASK_RSVD_MBO BIT_ULL(17)
71 #define SNP_POLICY_MASK_DEBUG BIT_ULL(19)
72 #define SNP_POLICY_MASK_SINGLE_SOCKET BIT_ULL(20)
73
74 #define SNP_POLICY_MASK_VALID (SNP_POLICY_MASK_API_MINOR | \
75 SNP_POLICY_MASK_API_MAJOR | \
76 SNP_POLICY_MASK_SMT | \
77 SNP_POLICY_MASK_RSVD_MBO | \
78 SNP_POLICY_MASK_DEBUG | \
79 SNP_POLICY_MASK_SINGLE_SOCKET)
80
81 #define INITIAL_VMSA_GPA 0xFFFFFFFFF000
82
83 static u8 sev_enc_bit;
84 static DECLARE_RWSEM(sev_deactivate_lock);
85 static DEFINE_MUTEX(sev_bitmap_lock);
86 unsigned int max_sev_asid;
87 static unsigned int min_sev_asid;
88 static unsigned long sev_me_mask;
89 static unsigned int nr_asids;
90 static unsigned long *sev_asid_bitmap;
91 static unsigned long *sev_reclaim_asid_bitmap;
92
93 static int snp_decommission_context(struct kvm *kvm);
94
95 struct enc_region {
96 struct list_head list;
97 unsigned long npages;
98 struct page **pages;
99 unsigned long uaddr;
100 unsigned long size;
101 };
102
103 /* Called with the sev_bitmap_lock held, or on shutdown */
sev_flush_asids(unsigned int min_asid,unsigned int max_asid)104 static int sev_flush_asids(unsigned int min_asid, unsigned int max_asid)
105 {
106 int ret, error = 0;
107 unsigned int asid;
108
109 /* Check if there are any ASIDs to reclaim before performing a flush */
110 asid = find_next_bit(sev_reclaim_asid_bitmap, nr_asids, min_asid);
111 if (asid > max_asid)
112 return -EBUSY;
113
114 /*
115 * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
116 * so it must be guarded.
117 */
118 down_write(&sev_deactivate_lock);
119
120 wbinvd_on_all_cpus();
121
122 if (sev_snp_enabled)
123 ret = sev_do_cmd(SEV_CMD_SNP_DF_FLUSH, NULL, &error);
124 else
125 ret = sev_guest_df_flush(&error);
126
127 up_write(&sev_deactivate_lock);
128
129 if (ret)
130 pr_err("SEV%s: DF_FLUSH failed, ret=%d, error=%#x\n",
131 sev_snp_enabled ? "-SNP" : "", ret, error);
132
133 return ret;
134 }
135
is_mirroring_enc_context(struct kvm * kvm)136 static inline bool is_mirroring_enc_context(struct kvm *kvm)
137 {
138 return !!to_kvm_sev_info(kvm)->enc_context_owner;
139 }
140
sev_vcpu_has_debug_swap(struct vcpu_svm * svm)141 static bool sev_vcpu_has_debug_swap(struct vcpu_svm *svm)
142 {
143 struct kvm_vcpu *vcpu = &svm->vcpu;
144 struct kvm_sev_info *sev = to_kvm_sev_info(vcpu->kvm);
145
146 return sev->vmsa_features & SVM_SEV_FEAT_DEBUG_SWAP;
147 }
148
149 /* Must be called with the sev_bitmap_lock held */
__sev_recycle_asids(unsigned int min_asid,unsigned int max_asid)150 static bool __sev_recycle_asids(unsigned int min_asid, unsigned int max_asid)
151 {
152 if (sev_flush_asids(min_asid, max_asid))
153 return false;
154
155 /* The flush process will flush all reclaimable SEV and SEV-ES ASIDs */
156 bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
157 nr_asids);
158 bitmap_zero(sev_reclaim_asid_bitmap, nr_asids);
159
160 return true;
161 }
162
sev_misc_cg_try_charge(struct kvm_sev_info * sev)163 static int sev_misc_cg_try_charge(struct kvm_sev_info *sev)
164 {
165 enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
166 return misc_cg_try_charge(type, sev->misc_cg, 1);
167 }
168
sev_misc_cg_uncharge(struct kvm_sev_info * sev)169 static void sev_misc_cg_uncharge(struct kvm_sev_info *sev)
170 {
171 enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
172 misc_cg_uncharge(type, sev->misc_cg, 1);
173 }
174
sev_asid_new(struct kvm_sev_info * sev)175 static int sev_asid_new(struct kvm_sev_info *sev)
176 {
177 /*
178 * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
179 * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.
180 * Note: min ASID can end up larger than the max if basic SEV support is
181 * effectively disabled by disallowing use of ASIDs for SEV guests.
182 */
183 unsigned int min_asid = sev->es_active ? 1 : min_sev_asid;
184 unsigned int max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
185 unsigned int asid;
186 bool retry = true;
187 int ret;
188
189 if (min_asid > max_asid)
190 return -ENOTTY;
191
192 WARN_ON(sev->misc_cg);
193 sev->misc_cg = get_current_misc_cg();
194 ret = sev_misc_cg_try_charge(sev);
195 if (ret) {
196 put_misc_cg(sev->misc_cg);
197 sev->misc_cg = NULL;
198 return ret;
199 }
200
201 mutex_lock(&sev_bitmap_lock);
202
203 again:
204 asid = find_next_zero_bit(sev_asid_bitmap, max_asid + 1, min_asid);
205 if (asid > max_asid) {
206 if (retry && __sev_recycle_asids(min_asid, max_asid)) {
207 retry = false;
208 goto again;
209 }
210 mutex_unlock(&sev_bitmap_lock);
211 ret = -EBUSY;
212 goto e_uncharge;
213 }
214
215 __set_bit(asid, sev_asid_bitmap);
216
217 mutex_unlock(&sev_bitmap_lock);
218
219 sev->asid = asid;
220 return 0;
221 e_uncharge:
222 sev_misc_cg_uncharge(sev);
223 put_misc_cg(sev->misc_cg);
224 sev->misc_cg = NULL;
225 return ret;
226 }
227
sev_get_asid(struct kvm * kvm)228 static unsigned int sev_get_asid(struct kvm *kvm)
229 {
230 return to_kvm_sev_info(kvm)->asid;
231 }
232
sev_asid_free(struct kvm_sev_info * sev)233 static void sev_asid_free(struct kvm_sev_info *sev)
234 {
235 struct svm_cpu_data *sd;
236 int cpu;
237
238 mutex_lock(&sev_bitmap_lock);
239
240 __set_bit(sev->asid, sev_reclaim_asid_bitmap);
241
242 for_each_possible_cpu(cpu) {
243 sd = per_cpu_ptr(&svm_data, cpu);
244 sd->sev_vmcbs[sev->asid] = NULL;
245 }
246
247 mutex_unlock(&sev_bitmap_lock);
248
249 sev_misc_cg_uncharge(sev);
250 put_misc_cg(sev->misc_cg);
251 sev->misc_cg = NULL;
252 }
253
sev_decommission(unsigned int handle)254 static void sev_decommission(unsigned int handle)
255 {
256 struct sev_data_decommission decommission;
257
258 if (!handle)
259 return;
260
261 decommission.handle = handle;
262 sev_guest_decommission(&decommission, NULL);
263 }
264
265 /*
266 * Transition a page to hypervisor-owned/shared state in the RMP table. This
267 * should not fail under normal conditions, but leak the page should that
268 * happen since it will no longer be usable by the host due to RMP protections.
269 */
kvm_rmp_make_shared(struct kvm * kvm,u64 pfn,enum pg_level level)270 static int kvm_rmp_make_shared(struct kvm *kvm, u64 pfn, enum pg_level level)
271 {
272 if (KVM_BUG_ON(rmp_make_shared(pfn, level), kvm)) {
273 snp_leak_pages(pfn, page_level_size(level) >> PAGE_SHIFT);
274 return -EIO;
275 }
276
277 return 0;
278 }
279
280 /*
281 * Certain page-states, such as Pre-Guest and Firmware pages (as documented
282 * in Chapter 5 of the SEV-SNP Firmware ABI under "Page States") cannot be
283 * directly transitioned back to normal/hypervisor-owned state via RMPUPDATE
284 * unless they are reclaimed first.
285 *
286 * Until they are reclaimed and subsequently transitioned via RMPUPDATE, they
287 * might not be usable by the host due to being set as immutable or still
288 * being associated with a guest ASID.
289 *
290 * Bug the VM and leak the page if reclaim fails, or if the RMP entry can't be
291 * converted back to shared, as the page is no longer usable due to RMP
292 * protections, and it's infeasible for the guest to continue on.
293 */
snp_page_reclaim(struct kvm * kvm,u64 pfn)294 static int snp_page_reclaim(struct kvm *kvm, u64 pfn)
295 {
296 struct sev_data_snp_page_reclaim data = {0};
297 int fw_err, rc;
298
299 data.paddr = __sme_set(pfn << PAGE_SHIFT);
300 rc = sev_do_cmd(SEV_CMD_SNP_PAGE_RECLAIM, &data, &fw_err);
301 if (KVM_BUG(rc, kvm, "Failed to reclaim PFN %llx, rc %d fw_err %d", pfn, rc, fw_err)) {
302 snp_leak_pages(pfn, 1);
303 return -EIO;
304 }
305
306 if (kvm_rmp_make_shared(kvm, pfn, PG_LEVEL_4K))
307 return -EIO;
308
309 return rc;
310 }
311
sev_unbind_asid(struct kvm * kvm,unsigned int handle)312 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
313 {
314 struct sev_data_deactivate deactivate;
315
316 if (!handle)
317 return;
318
319 deactivate.handle = handle;
320
321 /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
322 down_read(&sev_deactivate_lock);
323 sev_guest_deactivate(&deactivate, NULL);
324 up_read(&sev_deactivate_lock);
325
326 sev_decommission(handle);
327 }
328
329 /*
330 * This sets up bounce buffers/firmware pages to handle SNP Guest Request
331 * messages (e.g. attestation requests). See "SNP Guest Request" in the GHCB
332 * 2.0 specification for more details.
333 *
334 * Technically, when an SNP Guest Request is issued, the guest will provide its
335 * own request/response pages, which could in theory be passed along directly
336 * to firmware rather than using bounce pages. However, these pages would need
337 * special care:
338 *
339 * - Both pages are from shared guest memory, so they need to be protected
340 * from migration/etc. occurring while firmware reads/writes to them. At a
341 * minimum, this requires elevating the ref counts and potentially needing
342 * an explicit pinning of the memory. This places additional restrictions
343 * on what type of memory backends userspace can use for shared guest
344 * memory since there is some reliance on using refcounted pages.
345 *
346 * - The response page needs to be switched to Firmware-owned[1] state
347 * before the firmware can write to it, which can lead to potential
348 * host RMP #PFs if the guest is misbehaved and hands the host a
349 * guest page that KVM might write to for other reasons (e.g. virtio
350 * buffers/etc.).
351 *
352 * Both of these issues can be avoided completely by using separately-allocated
353 * bounce pages for both the request/response pages and passing those to
354 * firmware instead. So that's what is being set up here.
355 *
356 * Guest requests rely on message sequence numbers to ensure requests are
357 * issued to firmware in the order the guest issues them, so concurrent guest
358 * requests generally shouldn't happen. But a misbehaved guest could issue
359 * concurrent guest requests in theory, so a mutex is used to serialize
360 * access to the bounce buffers.
361 *
362 * [1] See the "Page States" section of the SEV-SNP Firmware ABI for more
363 * details on Firmware-owned pages, along with "RMP and VMPL Access Checks"
364 * in the APM for details on the related RMP restrictions.
365 */
snp_guest_req_init(struct kvm * kvm)366 static int snp_guest_req_init(struct kvm *kvm)
367 {
368 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
369 struct page *req_page;
370
371 req_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
372 if (!req_page)
373 return -ENOMEM;
374
375 sev->guest_resp_buf = snp_alloc_firmware_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
376 if (!sev->guest_resp_buf) {
377 __free_page(req_page);
378 return -EIO;
379 }
380
381 sev->guest_req_buf = page_address(req_page);
382 mutex_init(&sev->guest_req_mutex);
383
384 return 0;
385 }
386
snp_guest_req_cleanup(struct kvm * kvm)387 static void snp_guest_req_cleanup(struct kvm *kvm)
388 {
389 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
390
391 if (sev->guest_resp_buf)
392 snp_free_firmware_page(sev->guest_resp_buf);
393
394 if (sev->guest_req_buf)
395 __free_page(virt_to_page(sev->guest_req_buf));
396
397 sev->guest_req_buf = NULL;
398 sev->guest_resp_buf = NULL;
399 }
400
__sev_guest_init(struct kvm * kvm,struct kvm_sev_cmd * argp,struct kvm_sev_init * data,unsigned long vm_type)401 static int __sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp,
402 struct kvm_sev_init *data,
403 unsigned long vm_type)
404 {
405 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
406 struct sev_platform_init_args init_args = {0};
407 bool es_active = vm_type != KVM_X86_SEV_VM;
408 u64 valid_vmsa_features = es_active ? sev_supported_vmsa_features : 0;
409 int ret;
410
411 if (kvm->created_vcpus)
412 return -EINVAL;
413
414 if (data->flags)
415 return -EINVAL;
416
417 if (data->vmsa_features & ~valid_vmsa_features)
418 return -EINVAL;
419
420 if (data->ghcb_version > GHCB_VERSION_MAX || (!es_active && data->ghcb_version))
421 return -EINVAL;
422
423 if (unlikely(sev->active))
424 return -EINVAL;
425
426 sev->active = true;
427 sev->es_active = es_active;
428 sev->vmsa_features = data->vmsa_features;
429 sev->ghcb_version = data->ghcb_version;
430
431 /*
432 * Currently KVM supports the full range of mandatory features defined
433 * by version 2 of the GHCB protocol, so default to that for SEV-ES
434 * guests created via KVM_SEV_INIT2.
435 */
436 if (sev->es_active && !sev->ghcb_version)
437 sev->ghcb_version = GHCB_VERSION_DEFAULT;
438
439 if (vm_type == KVM_X86_SNP_VM)
440 sev->vmsa_features |= SVM_SEV_FEAT_SNP_ACTIVE;
441
442 ret = sev_asid_new(sev);
443 if (ret)
444 goto e_no_asid;
445
446 init_args.probe = false;
447 ret = sev_platform_init(&init_args);
448 if (ret)
449 goto e_free;
450
451 /* This needs to happen after SEV/SNP firmware initialization. */
452 if (vm_type == KVM_X86_SNP_VM) {
453 ret = snp_guest_req_init(kvm);
454 if (ret)
455 goto e_free;
456 }
457
458 INIT_LIST_HEAD(&sev->regions_list);
459 INIT_LIST_HEAD(&sev->mirror_vms);
460 sev->need_init = false;
461
462 kvm_set_apicv_inhibit(kvm, APICV_INHIBIT_REASON_SEV);
463
464 return 0;
465
466 e_free:
467 argp->error = init_args.error;
468 sev_asid_free(sev);
469 sev->asid = 0;
470 e_no_asid:
471 sev->vmsa_features = 0;
472 sev->es_active = false;
473 sev->active = false;
474 return ret;
475 }
476
sev_guest_init(struct kvm * kvm,struct kvm_sev_cmd * argp)477 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
478 {
479 struct kvm_sev_init data = {
480 .vmsa_features = 0,
481 .ghcb_version = 0,
482 };
483 unsigned long vm_type;
484
485 if (kvm->arch.vm_type != KVM_X86_DEFAULT_VM)
486 return -EINVAL;
487
488 vm_type = (argp->id == KVM_SEV_INIT ? KVM_X86_SEV_VM : KVM_X86_SEV_ES_VM);
489
490 /*
491 * KVM_SEV_ES_INIT has been deprecated by KVM_SEV_INIT2, so it will
492 * continue to only ever support the minimal GHCB protocol version.
493 */
494 if (vm_type == KVM_X86_SEV_ES_VM)
495 data.ghcb_version = GHCB_VERSION_MIN;
496
497 return __sev_guest_init(kvm, argp, &data, vm_type);
498 }
499
sev_guest_init2(struct kvm * kvm,struct kvm_sev_cmd * argp)500 static int sev_guest_init2(struct kvm *kvm, struct kvm_sev_cmd *argp)
501 {
502 struct kvm_sev_init data;
503
504 if (!to_kvm_sev_info(kvm)->need_init)
505 return -EINVAL;
506
507 if (kvm->arch.vm_type != KVM_X86_SEV_VM &&
508 kvm->arch.vm_type != KVM_X86_SEV_ES_VM &&
509 kvm->arch.vm_type != KVM_X86_SNP_VM)
510 return -EINVAL;
511
512 if (copy_from_user(&data, u64_to_user_ptr(argp->data), sizeof(data)))
513 return -EFAULT;
514
515 return __sev_guest_init(kvm, argp, &data, kvm->arch.vm_type);
516 }
517
sev_bind_asid(struct kvm * kvm,unsigned int handle,int * error)518 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
519 {
520 unsigned int asid = sev_get_asid(kvm);
521 struct sev_data_activate activate;
522 int ret;
523
524 /* activate ASID on the given handle */
525 activate.handle = handle;
526 activate.asid = asid;
527 ret = sev_guest_activate(&activate, error);
528
529 return ret;
530 }
531
__sev_issue_cmd(int fd,int id,void * data,int * error)532 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
533 {
534 CLASS(fd, f)(fd);
535
536 if (fd_empty(f))
537 return -EBADF;
538
539 return sev_issue_cmd_external_user(fd_file(f), id, data, error);
540 }
541
sev_issue_cmd(struct kvm * kvm,int id,void * data,int * error)542 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
543 {
544 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
545
546 return __sev_issue_cmd(sev->fd, id, data, error);
547 }
548
sev_launch_start(struct kvm * kvm,struct kvm_sev_cmd * argp)549 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
550 {
551 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
552 struct sev_data_launch_start start;
553 struct kvm_sev_launch_start params;
554 void *dh_blob, *session_blob;
555 int *error = &argp->error;
556 int ret;
557
558 if (!sev_guest(kvm))
559 return -ENOTTY;
560
561 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data), sizeof(params)))
562 return -EFAULT;
563
564 sev->policy = params.policy;
565
566 memset(&start, 0, sizeof(start));
567
568 dh_blob = NULL;
569 if (params.dh_uaddr) {
570 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
571 if (IS_ERR(dh_blob))
572 return PTR_ERR(dh_blob);
573
574 start.dh_cert_address = __sme_set(__pa(dh_blob));
575 start.dh_cert_len = params.dh_len;
576 }
577
578 session_blob = NULL;
579 if (params.session_uaddr) {
580 session_blob = psp_copy_user_blob(params.session_uaddr, params.session_len);
581 if (IS_ERR(session_blob)) {
582 ret = PTR_ERR(session_blob);
583 goto e_free_dh;
584 }
585
586 start.session_address = __sme_set(__pa(session_blob));
587 start.session_len = params.session_len;
588 }
589
590 start.handle = params.handle;
591 start.policy = params.policy;
592
593 /* create memory encryption context */
594 ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_LAUNCH_START, &start, error);
595 if (ret)
596 goto e_free_session;
597
598 /* Bind ASID to this guest */
599 ret = sev_bind_asid(kvm, start.handle, error);
600 if (ret) {
601 sev_decommission(start.handle);
602 goto e_free_session;
603 }
604
605 /* return handle to userspace */
606 params.handle = start.handle;
607 if (copy_to_user(u64_to_user_ptr(argp->data), ¶ms, sizeof(params))) {
608 sev_unbind_asid(kvm, start.handle);
609 ret = -EFAULT;
610 goto e_free_session;
611 }
612
613 sev->handle = start.handle;
614 sev->fd = argp->sev_fd;
615
616 e_free_session:
617 kfree(session_blob);
618 e_free_dh:
619 kfree(dh_blob);
620 return ret;
621 }
622
sev_pin_memory(struct kvm * kvm,unsigned long uaddr,unsigned long ulen,unsigned long * n,unsigned int flags)623 static struct page **sev_pin_memory(struct kvm *kvm, unsigned long uaddr,
624 unsigned long ulen, unsigned long *n,
625 unsigned int flags)
626 {
627 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
628 unsigned long npages, size;
629 int npinned;
630 unsigned long locked, lock_limit;
631 struct page **pages;
632 unsigned long first, last;
633 int ret;
634
635 lockdep_assert_held(&kvm->lock);
636
637 if (ulen == 0 || uaddr + ulen < uaddr)
638 return ERR_PTR(-EINVAL);
639
640 /* Calculate number of pages. */
641 first = (uaddr & PAGE_MASK) >> PAGE_SHIFT;
642 last = ((uaddr + ulen - 1) & PAGE_MASK) >> PAGE_SHIFT;
643 npages = (last - first + 1);
644
645 locked = sev->pages_locked + npages;
646 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
647 if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
648 pr_err("SEV: %lu locked pages exceed the lock limit of %lu.\n", locked, lock_limit);
649 return ERR_PTR(-ENOMEM);
650 }
651
652 if (WARN_ON_ONCE(npages > INT_MAX))
653 return ERR_PTR(-EINVAL);
654
655 /* Avoid using vmalloc for smaller buffers. */
656 size = npages * sizeof(struct page *);
657 if (size > PAGE_SIZE)
658 pages = __vmalloc(size, GFP_KERNEL_ACCOUNT);
659 else
660 pages = kmalloc(size, GFP_KERNEL_ACCOUNT);
661
662 if (!pages)
663 return ERR_PTR(-ENOMEM);
664
665 /* Pin the user virtual address. */
666 npinned = pin_user_pages_fast(uaddr, npages, flags, pages);
667 if (npinned != npages) {
668 pr_err("SEV: Failure locking %lu pages.\n", npages);
669 ret = -ENOMEM;
670 goto err;
671 }
672
673 *n = npages;
674 sev->pages_locked = locked;
675
676 return pages;
677
678 err:
679 if (npinned > 0)
680 unpin_user_pages(pages, npinned);
681
682 kvfree(pages);
683 return ERR_PTR(ret);
684 }
685
sev_unpin_memory(struct kvm * kvm,struct page ** pages,unsigned long npages)686 static void sev_unpin_memory(struct kvm *kvm, struct page **pages,
687 unsigned long npages)
688 {
689 unpin_user_pages(pages, npages);
690 kvfree(pages);
691 to_kvm_sev_info(kvm)->pages_locked -= npages;
692 }
693
sev_clflush_pages(struct page * pages[],unsigned long npages)694 static void sev_clflush_pages(struct page *pages[], unsigned long npages)
695 {
696 uint8_t *page_virtual;
697 unsigned long i;
698
699 if (this_cpu_has(X86_FEATURE_SME_COHERENT) || npages == 0 ||
700 pages == NULL)
701 return;
702
703 for (i = 0; i < npages; i++) {
704 page_virtual = kmap_local_page(pages[i]);
705 clflush_cache_range(page_virtual, PAGE_SIZE);
706 kunmap_local(page_virtual);
707 cond_resched();
708 }
709 }
710
get_num_contig_pages(unsigned long idx,struct page ** inpages,unsigned long npages)711 static unsigned long get_num_contig_pages(unsigned long idx,
712 struct page **inpages, unsigned long npages)
713 {
714 unsigned long paddr, next_paddr;
715 unsigned long i = idx + 1, pages = 1;
716
717 /* find the number of contiguous pages starting from idx */
718 paddr = __sme_page_pa(inpages[idx]);
719 while (i < npages) {
720 next_paddr = __sme_page_pa(inpages[i++]);
721 if ((paddr + PAGE_SIZE) == next_paddr) {
722 pages++;
723 paddr = next_paddr;
724 continue;
725 }
726 break;
727 }
728
729 return pages;
730 }
731
sev_launch_update_data(struct kvm * kvm,struct kvm_sev_cmd * argp)732 static int sev_launch_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
733 {
734 unsigned long vaddr, vaddr_end, next_vaddr, npages, pages, size, i;
735 struct kvm_sev_launch_update_data params;
736 struct sev_data_launch_update_data data;
737 struct page **inpages;
738 int ret;
739
740 if (!sev_guest(kvm))
741 return -ENOTTY;
742
743 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data), sizeof(params)))
744 return -EFAULT;
745
746 vaddr = params.uaddr;
747 size = params.len;
748 vaddr_end = vaddr + size;
749
750 /* Lock the user memory. */
751 inpages = sev_pin_memory(kvm, vaddr, size, &npages, FOLL_WRITE);
752 if (IS_ERR(inpages))
753 return PTR_ERR(inpages);
754
755 /*
756 * Flush (on non-coherent CPUs) before LAUNCH_UPDATE encrypts pages in
757 * place; the cache may contain the data that was written unencrypted.
758 */
759 sev_clflush_pages(inpages, npages);
760
761 data.reserved = 0;
762 data.handle = to_kvm_sev_info(kvm)->handle;
763
764 for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
765 int offset, len;
766
767 /*
768 * If the user buffer is not page-aligned, calculate the offset
769 * within the page.
770 */
771 offset = vaddr & (PAGE_SIZE - 1);
772
773 /* Calculate the number of pages that can be encrypted in one go. */
774 pages = get_num_contig_pages(i, inpages, npages);
775
776 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
777
778 data.len = len;
779 data.address = __sme_page_pa(inpages[i]) + offset;
780 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, &data, &argp->error);
781 if (ret)
782 goto e_unpin;
783
784 size -= len;
785 next_vaddr = vaddr + len;
786 }
787
788 e_unpin:
789 /* content of memory is updated, mark pages dirty */
790 for (i = 0; i < npages; i++) {
791 set_page_dirty_lock(inpages[i]);
792 mark_page_accessed(inpages[i]);
793 }
794 /* unlock the user pages */
795 sev_unpin_memory(kvm, inpages, npages);
796 return ret;
797 }
798
sev_es_sync_vmsa(struct vcpu_svm * svm)799 static int sev_es_sync_vmsa(struct vcpu_svm *svm)
800 {
801 struct kvm_vcpu *vcpu = &svm->vcpu;
802 struct kvm_sev_info *sev = to_kvm_sev_info(vcpu->kvm);
803 struct sev_es_save_area *save = svm->sev_es.vmsa;
804 struct xregs_state *xsave;
805 const u8 *s;
806 u8 *d;
807 int i;
808
809 /* Check some debug related fields before encrypting the VMSA */
810 if (svm->vcpu.guest_debug || (svm->vmcb->save.dr7 & ~DR7_FIXED_1))
811 return -EINVAL;
812
813 /*
814 * SEV-ES will use a VMSA that is pointed to by the VMCB, not
815 * the traditional VMSA that is part of the VMCB. Copy the
816 * traditional VMSA as it has been built so far (in prep
817 * for LAUNCH_UPDATE_VMSA) to be the initial SEV-ES state.
818 */
819 memcpy(save, &svm->vmcb->save, sizeof(svm->vmcb->save));
820
821 /* Sync registgers */
822 save->rax = svm->vcpu.arch.regs[VCPU_REGS_RAX];
823 save->rbx = svm->vcpu.arch.regs[VCPU_REGS_RBX];
824 save->rcx = svm->vcpu.arch.regs[VCPU_REGS_RCX];
825 save->rdx = svm->vcpu.arch.regs[VCPU_REGS_RDX];
826 save->rsp = svm->vcpu.arch.regs[VCPU_REGS_RSP];
827 save->rbp = svm->vcpu.arch.regs[VCPU_REGS_RBP];
828 save->rsi = svm->vcpu.arch.regs[VCPU_REGS_RSI];
829 save->rdi = svm->vcpu.arch.regs[VCPU_REGS_RDI];
830 #ifdef CONFIG_X86_64
831 save->r8 = svm->vcpu.arch.regs[VCPU_REGS_R8];
832 save->r9 = svm->vcpu.arch.regs[VCPU_REGS_R9];
833 save->r10 = svm->vcpu.arch.regs[VCPU_REGS_R10];
834 save->r11 = svm->vcpu.arch.regs[VCPU_REGS_R11];
835 save->r12 = svm->vcpu.arch.regs[VCPU_REGS_R12];
836 save->r13 = svm->vcpu.arch.regs[VCPU_REGS_R13];
837 save->r14 = svm->vcpu.arch.regs[VCPU_REGS_R14];
838 save->r15 = svm->vcpu.arch.regs[VCPU_REGS_R15];
839 #endif
840 save->rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
841
842 /* Sync some non-GPR registers before encrypting */
843 save->xcr0 = svm->vcpu.arch.xcr0;
844 save->pkru = svm->vcpu.arch.pkru;
845 save->xss = svm->vcpu.arch.ia32_xss;
846 save->dr6 = svm->vcpu.arch.dr6;
847
848 save->sev_features = sev->vmsa_features;
849
850 /*
851 * Skip FPU and AVX setup with KVM_SEV_ES_INIT to avoid
852 * breaking older measurements.
853 */
854 if (vcpu->kvm->arch.vm_type != KVM_X86_DEFAULT_VM) {
855 xsave = &vcpu->arch.guest_fpu.fpstate->regs.xsave;
856 save->x87_dp = xsave->i387.rdp;
857 save->mxcsr = xsave->i387.mxcsr;
858 save->x87_ftw = xsave->i387.twd;
859 save->x87_fsw = xsave->i387.swd;
860 save->x87_fcw = xsave->i387.cwd;
861 save->x87_fop = xsave->i387.fop;
862 save->x87_ds = 0;
863 save->x87_cs = 0;
864 save->x87_rip = xsave->i387.rip;
865
866 for (i = 0; i < 8; i++) {
867 /*
868 * The format of the x87 save area is undocumented and
869 * definitely not what you would expect. It consists of
870 * an 8*8 bytes area with bytes 0-7, and an 8*2 bytes
871 * area with bytes 8-9 of each register.
872 */
873 d = save->fpreg_x87 + i * 8;
874 s = ((u8 *)xsave->i387.st_space) + i * 16;
875 memcpy(d, s, 8);
876 save->fpreg_x87[64 + i * 2] = s[8];
877 save->fpreg_x87[64 + i * 2 + 1] = s[9];
878 }
879 memcpy(save->fpreg_xmm, xsave->i387.xmm_space, 256);
880
881 s = get_xsave_addr(xsave, XFEATURE_YMM);
882 if (s)
883 memcpy(save->fpreg_ymm, s, 256);
884 else
885 memset(save->fpreg_ymm, 0, 256);
886 }
887
888 pr_debug("Virtual Machine Save Area (VMSA):\n");
889 print_hex_dump_debug("", DUMP_PREFIX_NONE, 16, 1, save, sizeof(*save), false);
890
891 return 0;
892 }
893
__sev_launch_update_vmsa(struct kvm * kvm,struct kvm_vcpu * vcpu,int * error)894 static int __sev_launch_update_vmsa(struct kvm *kvm, struct kvm_vcpu *vcpu,
895 int *error)
896 {
897 struct sev_data_launch_update_vmsa vmsa;
898 struct vcpu_svm *svm = to_svm(vcpu);
899 int ret;
900
901 if (vcpu->guest_debug) {
902 pr_warn_once("KVM_SET_GUEST_DEBUG for SEV-ES guest is not supported");
903 return -EINVAL;
904 }
905
906 /* Perform some pre-encryption checks against the VMSA */
907 ret = sev_es_sync_vmsa(svm);
908 if (ret)
909 return ret;
910
911 /*
912 * The LAUNCH_UPDATE_VMSA command will perform in-place encryption of
913 * the VMSA memory content (i.e it will write the same memory region
914 * with the guest's key), so invalidate it first.
915 */
916 clflush_cache_range(svm->sev_es.vmsa, PAGE_SIZE);
917
918 vmsa.reserved = 0;
919 vmsa.handle = to_kvm_sev_info(kvm)->handle;
920 vmsa.address = __sme_pa(svm->sev_es.vmsa);
921 vmsa.len = PAGE_SIZE;
922 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_VMSA, &vmsa, error);
923 if (ret)
924 return ret;
925
926 /*
927 * SEV-ES guests maintain an encrypted version of their FPU
928 * state which is restored and saved on VMRUN and VMEXIT.
929 * Mark vcpu->arch.guest_fpu->fpstate as scratch so it won't
930 * do xsave/xrstor on it.
931 */
932 fpstate_set_confidential(&vcpu->arch.guest_fpu);
933 vcpu->arch.guest_state_protected = true;
934
935 /*
936 * SEV-ES guest mandates LBR Virtualization to be _always_ ON. Enable it
937 * only after setting guest_state_protected because KVM_SET_MSRS allows
938 * dynamic toggling of LBRV (for performance reason) on write access to
939 * MSR_IA32_DEBUGCTLMSR when guest_state_protected is not set.
940 */
941 svm_enable_lbrv(vcpu);
942 return 0;
943 }
944
sev_launch_update_vmsa(struct kvm * kvm,struct kvm_sev_cmd * argp)945 static int sev_launch_update_vmsa(struct kvm *kvm, struct kvm_sev_cmd *argp)
946 {
947 struct kvm_vcpu *vcpu;
948 unsigned long i;
949 int ret;
950
951 if (!sev_es_guest(kvm))
952 return -ENOTTY;
953
954 kvm_for_each_vcpu(i, vcpu, kvm) {
955 ret = mutex_lock_killable(&vcpu->mutex);
956 if (ret)
957 return ret;
958
959 ret = __sev_launch_update_vmsa(kvm, vcpu, &argp->error);
960
961 mutex_unlock(&vcpu->mutex);
962 if (ret)
963 return ret;
964 }
965
966 return 0;
967 }
968
sev_launch_measure(struct kvm * kvm,struct kvm_sev_cmd * argp)969 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
970 {
971 void __user *measure = u64_to_user_ptr(argp->data);
972 struct sev_data_launch_measure data;
973 struct kvm_sev_launch_measure params;
974 void __user *p = NULL;
975 void *blob = NULL;
976 int ret;
977
978 if (!sev_guest(kvm))
979 return -ENOTTY;
980
981 if (copy_from_user(¶ms, measure, sizeof(params)))
982 return -EFAULT;
983
984 memset(&data, 0, sizeof(data));
985
986 /* User wants to query the blob length */
987 if (!params.len)
988 goto cmd;
989
990 p = u64_to_user_ptr(params.uaddr);
991 if (p) {
992 if (params.len > SEV_FW_BLOB_MAX_SIZE)
993 return -EINVAL;
994
995 blob = kzalloc(params.len, GFP_KERNEL_ACCOUNT);
996 if (!blob)
997 return -ENOMEM;
998
999 data.address = __psp_pa(blob);
1000 data.len = params.len;
1001 }
1002
1003 cmd:
1004 data.handle = to_kvm_sev_info(kvm)->handle;
1005 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, &data, &argp->error);
1006
1007 /*
1008 * If we query the session length, FW responded with expected data.
1009 */
1010 if (!params.len)
1011 goto done;
1012
1013 if (ret)
1014 goto e_free_blob;
1015
1016 if (blob) {
1017 if (copy_to_user(p, blob, params.len))
1018 ret = -EFAULT;
1019 }
1020
1021 done:
1022 params.len = data.len;
1023 if (copy_to_user(measure, ¶ms, sizeof(params)))
1024 ret = -EFAULT;
1025 e_free_blob:
1026 kfree(blob);
1027 return ret;
1028 }
1029
sev_launch_finish(struct kvm * kvm,struct kvm_sev_cmd * argp)1030 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1031 {
1032 struct sev_data_launch_finish data;
1033
1034 if (!sev_guest(kvm))
1035 return -ENOTTY;
1036
1037 data.handle = to_kvm_sev_info(kvm)->handle;
1038 return sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, &data, &argp->error);
1039 }
1040
sev_guest_status(struct kvm * kvm,struct kvm_sev_cmd * argp)1041 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
1042 {
1043 struct kvm_sev_guest_status params;
1044 struct sev_data_guest_status data;
1045 int ret;
1046
1047 if (!sev_guest(kvm))
1048 return -ENOTTY;
1049
1050 memset(&data, 0, sizeof(data));
1051
1052 data.handle = to_kvm_sev_info(kvm)->handle;
1053 ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, &data, &argp->error);
1054 if (ret)
1055 return ret;
1056
1057 params.policy = data.policy;
1058 params.state = data.state;
1059 params.handle = data.handle;
1060
1061 if (copy_to_user(u64_to_user_ptr(argp->data), ¶ms, sizeof(params)))
1062 ret = -EFAULT;
1063
1064 return ret;
1065 }
1066
__sev_issue_dbg_cmd(struct kvm * kvm,unsigned long src,unsigned long dst,int size,int * error,bool enc)1067 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
1068 unsigned long dst, int size,
1069 int *error, bool enc)
1070 {
1071 struct sev_data_dbg data;
1072
1073 data.reserved = 0;
1074 data.handle = to_kvm_sev_info(kvm)->handle;
1075 data.dst_addr = dst;
1076 data.src_addr = src;
1077 data.len = size;
1078
1079 return sev_issue_cmd(kvm,
1080 enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
1081 &data, error);
1082 }
1083
__sev_dbg_decrypt(struct kvm * kvm,unsigned long src_paddr,unsigned long dst_paddr,int sz,int * err)1084 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
1085 unsigned long dst_paddr, int sz, int *err)
1086 {
1087 int offset;
1088
1089 /*
1090 * Its safe to read more than we are asked, caller should ensure that
1091 * destination has enough space.
1092 */
1093 offset = src_paddr & 15;
1094 src_paddr = round_down(src_paddr, 16);
1095 sz = round_up(sz + offset, 16);
1096
1097 return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
1098 }
1099
__sev_dbg_decrypt_user(struct kvm * kvm,unsigned long paddr,void __user * dst_uaddr,unsigned long dst_paddr,int size,int * err)1100 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
1101 void __user *dst_uaddr,
1102 unsigned long dst_paddr,
1103 int size, int *err)
1104 {
1105 struct page *tpage = NULL;
1106 int ret, offset;
1107
1108 /* if inputs are not 16-byte then use intermediate buffer */
1109 if (!IS_ALIGNED(dst_paddr, 16) ||
1110 !IS_ALIGNED(paddr, 16) ||
1111 !IS_ALIGNED(size, 16)) {
1112 tpage = (void *)alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
1113 if (!tpage)
1114 return -ENOMEM;
1115
1116 dst_paddr = __sme_page_pa(tpage);
1117 }
1118
1119 ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
1120 if (ret)
1121 goto e_free;
1122
1123 if (tpage) {
1124 offset = paddr & 15;
1125 if (copy_to_user(dst_uaddr, page_address(tpage) + offset, size))
1126 ret = -EFAULT;
1127 }
1128
1129 e_free:
1130 if (tpage)
1131 __free_page(tpage);
1132
1133 return ret;
1134 }
1135
__sev_dbg_encrypt_user(struct kvm * kvm,unsigned long paddr,void __user * vaddr,unsigned long dst_paddr,void __user * dst_vaddr,int size,int * error)1136 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
1137 void __user *vaddr,
1138 unsigned long dst_paddr,
1139 void __user *dst_vaddr,
1140 int size, int *error)
1141 {
1142 struct page *src_tpage = NULL;
1143 struct page *dst_tpage = NULL;
1144 int ret, len = size;
1145
1146 /* If source buffer is not aligned then use an intermediate buffer */
1147 if (!IS_ALIGNED((unsigned long)vaddr, 16)) {
1148 src_tpage = alloc_page(GFP_KERNEL_ACCOUNT);
1149 if (!src_tpage)
1150 return -ENOMEM;
1151
1152 if (copy_from_user(page_address(src_tpage), vaddr, size)) {
1153 __free_page(src_tpage);
1154 return -EFAULT;
1155 }
1156
1157 paddr = __sme_page_pa(src_tpage);
1158 }
1159
1160 /*
1161 * If destination buffer or length is not aligned then do read-modify-write:
1162 * - decrypt destination in an intermediate buffer
1163 * - copy the source buffer in an intermediate buffer
1164 * - use the intermediate buffer as source buffer
1165 */
1166 if (!IS_ALIGNED((unsigned long)dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
1167 int dst_offset;
1168
1169 dst_tpage = alloc_page(GFP_KERNEL_ACCOUNT);
1170 if (!dst_tpage) {
1171 ret = -ENOMEM;
1172 goto e_free;
1173 }
1174
1175 ret = __sev_dbg_decrypt(kvm, dst_paddr,
1176 __sme_page_pa(dst_tpage), size, error);
1177 if (ret)
1178 goto e_free;
1179
1180 /*
1181 * If source is kernel buffer then use memcpy() otherwise
1182 * copy_from_user().
1183 */
1184 dst_offset = dst_paddr & 15;
1185
1186 if (src_tpage)
1187 memcpy(page_address(dst_tpage) + dst_offset,
1188 page_address(src_tpage), size);
1189 else {
1190 if (copy_from_user(page_address(dst_tpage) + dst_offset,
1191 vaddr, size)) {
1192 ret = -EFAULT;
1193 goto e_free;
1194 }
1195 }
1196
1197 paddr = __sme_page_pa(dst_tpage);
1198 dst_paddr = round_down(dst_paddr, 16);
1199 len = round_up(size, 16);
1200 }
1201
1202 ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
1203
1204 e_free:
1205 if (src_tpage)
1206 __free_page(src_tpage);
1207 if (dst_tpage)
1208 __free_page(dst_tpage);
1209 return ret;
1210 }
1211
sev_dbg_crypt(struct kvm * kvm,struct kvm_sev_cmd * argp,bool dec)1212 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
1213 {
1214 unsigned long vaddr, vaddr_end, next_vaddr;
1215 unsigned long dst_vaddr;
1216 struct page **src_p, **dst_p;
1217 struct kvm_sev_dbg debug;
1218 unsigned long n;
1219 unsigned int size;
1220 int ret;
1221
1222 if (!sev_guest(kvm))
1223 return -ENOTTY;
1224
1225 if (copy_from_user(&debug, u64_to_user_ptr(argp->data), sizeof(debug)))
1226 return -EFAULT;
1227
1228 if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
1229 return -EINVAL;
1230 if (!debug.dst_uaddr)
1231 return -EINVAL;
1232
1233 vaddr = debug.src_uaddr;
1234 size = debug.len;
1235 vaddr_end = vaddr + size;
1236 dst_vaddr = debug.dst_uaddr;
1237
1238 for (; vaddr < vaddr_end; vaddr = next_vaddr) {
1239 int len, s_off, d_off;
1240
1241 /* lock userspace source and destination page */
1242 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
1243 if (IS_ERR(src_p))
1244 return PTR_ERR(src_p);
1245
1246 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, FOLL_WRITE);
1247 if (IS_ERR(dst_p)) {
1248 sev_unpin_memory(kvm, src_p, n);
1249 return PTR_ERR(dst_p);
1250 }
1251
1252 /*
1253 * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
1254 * the pages; flush the destination too so that future accesses do not
1255 * see stale data.
1256 */
1257 sev_clflush_pages(src_p, 1);
1258 sev_clflush_pages(dst_p, 1);
1259
1260 /*
1261 * Since user buffer may not be page aligned, calculate the
1262 * offset within the page.
1263 */
1264 s_off = vaddr & ~PAGE_MASK;
1265 d_off = dst_vaddr & ~PAGE_MASK;
1266 len = min_t(size_t, (PAGE_SIZE - s_off), size);
1267
1268 if (dec)
1269 ret = __sev_dbg_decrypt_user(kvm,
1270 __sme_page_pa(src_p[0]) + s_off,
1271 (void __user *)dst_vaddr,
1272 __sme_page_pa(dst_p[0]) + d_off,
1273 len, &argp->error);
1274 else
1275 ret = __sev_dbg_encrypt_user(kvm,
1276 __sme_page_pa(src_p[0]) + s_off,
1277 (void __user *)vaddr,
1278 __sme_page_pa(dst_p[0]) + d_off,
1279 (void __user *)dst_vaddr,
1280 len, &argp->error);
1281
1282 sev_unpin_memory(kvm, src_p, n);
1283 sev_unpin_memory(kvm, dst_p, n);
1284
1285 if (ret)
1286 goto err;
1287
1288 next_vaddr = vaddr + len;
1289 dst_vaddr = dst_vaddr + len;
1290 size -= len;
1291 }
1292 err:
1293 return ret;
1294 }
1295
sev_launch_secret(struct kvm * kvm,struct kvm_sev_cmd * argp)1296 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
1297 {
1298 struct sev_data_launch_secret data;
1299 struct kvm_sev_launch_secret params;
1300 struct page **pages;
1301 void *blob, *hdr;
1302 unsigned long n, i;
1303 int ret, offset;
1304
1305 if (!sev_guest(kvm))
1306 return -ENOTTY;
1307
1308 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data), sizeof(params)))
1309 return -EFAULT;
1310
1311 pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, FOLL_WRITE);
1312 if (IS_ERR(pages))
1313 return PTR_ERR(pages);
1314
1315 /*
1316 * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
1317 * place; the cache may contain the data that was written unencrypted.
1318 */
1319 sev_clflush_pages(pages, n);
1320
1321 /*
1322 * The secret must be copied into contiguous memory region, lets verify
1323 * that userspace memory pages are contiguous before we issue command.
1324 */
1325 if (get_num_contig_pages(0, pages, n) != n) {
1326 ret = -EINVAL;
1327 goto e_unpin_memory;
1328 }
1329
1330 memset(&data, 0, sizeof(data));
1331
1332 offset = params.guest_uaddr & (PAGE_SIZE - 1);
1333 data.guest_address = __sme_page_pa(pages[0]) + offset;
1334 data.guest_len = params.guest_len;
1335
1336 blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1337 if (IS_ERR(blob)) {
1338 ret = PTR_ERR(blob);
1339 goto e_unpin_memory;
1340 }
1341
1342 data.trans_address = __psp_pa(blob);
1343 data.trans_len = params.trans_len;
1344
1345 hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1346 if (IS_ERR(hdr)) {
1347 ret = PTR_ERR(hdr);
1348 goto e_free_blob;
1349 }
1350 data.hdr_address = __psp_pa(hdr);
1351 data.hdr_len = params.hdr_len;
1352
1353 data.handle = to_kvm_sev_info(kvm)->handle;
1354 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, &data, &argp->error);
1355
1356 kfree(hdr);
1357
1358 e_free_blob:
1359 kfree(blob);
1360 e_unpin_memory:
1361 /* content of memory is updated, mark pages dirty */
1362 for (i = 0; i < n; i++) {
1363 set_page_dirty_lock(pages[i]);
1364 mark_page_accessed(pages[i]);
1365 }
1366 sev_unpin_memory(kvm, pages, n);
1367 return ret;
1368 }
1369
sev_get_attestation_report(struct kvm * kvm,struct kvm_sev_cmd * argp)1370 static int sev_get_attestation_report(struct kvm *kvm, struct kvm_sev_cmd *argp)
1371 {
1372 void __user *report = u64_to_user_ptr(argp->data);
1373 struct sev_data_attestation_report data;
1374 struct kvm_sev_attestation_report params;
1375 void __user *p;
1376 void *blob = NULL;
1377 int ret;
1378
1379 if (!sev_guest(kvm))
1380 return -ENOTTY;
1381
1382 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data), sizeof(params)))
1383 return -EFAULT;
1384
1385 memset(&data, 0, sizeof(data));
1386
1387 /* User wants to query the blob length */
1388 if (!params.len)
1389 goto cmd;
1390
1391 p = u64_to_user_ptr(params.uaddr);
1392 if (p) {
1393 if (params.len > SEV_FW_BLOB_MAX_SIZE)
1394 return -EINVAL;
1395
1396 blob = kzalloc(params.len, GFP_KERNEL_ACCOUNT);
1397 if (!blob)
1398 return -ENOMEM;
1399
1400 data.address = __psp_pa(blob);
1401 data.len = params.len;
1402 memcpy(data.mnonce, params.mnonce, sizeof(params.mnonce));
1403 }
1404 cmd:
1405 data.handle = to_kvm_sev_info(kvm)->handle;
1406 ret = sev_issue_cmd(kvm, SEV_CMD_ATTESTATION_REPORT, &data, &argp->error);
1407 /*
1408 * If we query the session length, FW responded with expected data.
1409 */
1410 if (!params.len)
1411 goto done;
1412
1413 if (ret)
1414 goto e_free_blob;
1415
1416 if (blob) {
1417 if (copy_to_user(p, blob, params.len))
1418 ret = -EFAULT;
1419 }
1420
1421 done:
1422 params.len = data.len;
1423 if (copy_to_user(report, ¶ms, sizeof(params)))
1424 ret = -EFAULT;
1425 e_free_blob:
1426 kfree(blob);
1427 return ret;
1428 }
1429
1430 /* Userspace wants to query session length. */
1431 static int
__sev_send_start_query_session_length(struct kvm * kvm,struct kvm_sev_cmd * argp,struct kvm_sev_send_start * params)1432 __sev_send_start_query_session_length(struct kvm *kvm, struct kvm_sev_cmd *argp,
1433 struct kvm_sev_send_start *params)
1434 {
1435 struct sev_data_send_start data;
1436 int ret;
1437
1438 memset(&data, 0, sizeof(data));
1439 data.handle = to_kvm_sev_info(kvm)->handle;
1440 ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1441
1442 params->session_len = data.session_len;
1443 if (copy_to_user(u64_to_user_ptr(argp->data), params,
1444 sizeof(struct kvm_sev_send_start)))
1445 ret = -EFAULT;
1446
1447 return ret;
1448 }
1449
sev_send_start(struct kvm * kvm,struct kvm_sev_cmd * argp)1450 static int sev_send_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1451 {
1452 struct sev_data_send_start data;
1453 struct kvm_sev_send_start params;
1454 void *amd_certs, *session_data;
1455 void *pdh_cert, *plat_certs;
1456 int ret;
1457
1458 if (!sev_guest(kvm))
1459 return -ENOTTY;
1460
1461 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data),
1462 sizeof(struct kvm_sev_send_start)))
1463 return -EFAULT;
1464
1465 /* if session_len is zero, userspace wants to query the session length */
1466 if (!params.session_len)
1467 return __sev_send_start_query_session_length(kvm, argp,
1468 ¶ms);
1469
1470 /* some sanity checks */
1471 if (!params.pdh_cert_uaddr || !params.pdh_cert_len ||
1472 !params.session_uaddr || params.session_len > SEV_FW_BLOB_MAX_SIZE)
1473 return -EINVAL;
1474
1475 /* allocate the memory to hold the session data blob */
1476 session_data = kzalloc(params.session_len, GFP_KERNEL_ACCOUNT);
1477 if (!session_data)
1478 return -ENOMEM;
1479
1480 /* copy the certificate blobs from userspace */
1481 pdh_cert = psp_copy_user_blob(params.pdh_cert_uaddr,
1482 params.pdh_cert_len);
1483 if (IS_ERR(pdh_cert)) {
1484 ret = PTR_ERR(pdh_cert);
1485 goto e_free_session;
1486 }
1487
1488 plat_certs = psp_copy_user_blob(params.plat_certs_uaddr,
1489 params.plat_certs_len);
1490 if (IS_ERR(plat_certs)) {
1491 ret = PTR_ERR(plat_certs);
1492 goto e_free_pdh;
1493 }
1494
1495 amd_certs = psp_copy_user_blob(params.amd_certs_uaddr,
1496 params.amd_certs_len);
1497 if (IS_ERR(amd_certs)) {
1498 ret = PTR_ERR(amd_certs);
1499 goto e_free_plat_cert;
1500 }
1501
1502 /* populate the FW SEND_START field with system physical address */
1503 memset(&data, 0, sizeof(data));
1504 data.pdh_cert_address = __psp_pa(pdh_cert);
1505 data.pdh_cert_len = params.pdh_cert_len;
1506 data.plat_certs_address = __psp_pa(plat_certs);
1507 data.plat_certs_len = params.plat_certs_len;
1508 data.amd_certs_address = __psp_pa(amd_certs);
1509 data.amd_certs_len = params.amd_certs_len;
1510 data.session_address = __psp_pa(session_data);
1511 data.session_len = params.session_len;
1512 data.handle = to_kvm_sev_info(kvm)->handle;
1513
1514 ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1515
1516 if (!ret && copy_to_user(u64_to_user_ptr(params.session_uaddr),
1517 session_data, params.session_len)) {
1518 ret = -EFAULT;
1519 goto e_free_amd_cert;
1520 }
1521
1522 params.policy = data.policy;
1523 params.session_len = data.session_len;
1524 if (copy_to_user(u64_to_user_ptr(argp->data), ¶ms,
1525 sizeof(struct kvm_sev_send_start)))
1526 ret = -EFAULT;
1527
1528 e_free_amd_cert:
1529 kfree(amd_certs);
1530 e_free_plat_cert:
1531 kfree(plat_certs);
1532 e_free_pdh:
1533 kfree(pdh_cert);
1534 e_free_session:
1535 kfree(session_data);
1536 return ret;
1537 }
1538
1539 /* Userspace wants to query either header or trans length. */
1540 static int
__sev_send_update_data_query_lengths(struct kvm * kvm,struct kvm_sev_cmd * argp,struct kvm_sev_send_update_data * params)1541 __sev_send_update_data_query_lengths(struct kvm *kvm, struct kvm_sev_cmd *argp,
1542 struct kvm_sev_send_update_data *params)
1543 {
1544 struct sev_data_send_update_data data;
1545 int ret;
1546
1547 memset(&data, 0, sizeof(data));
1548 data.handle = to_kvm_sev_info(kvm)->handle;
1549 ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1550
1551 params->hdr_len = data.hdr_len;
1552 params->trans_len = data.trans_len;
1553
1554 if (copy_to_user(u64_to_user_ptr(argp->data), params,
1555 sizeof(struct kvm_sev_send_update_data)))
1556 ret = -EFAULT;
1557
1558 return ret;
1559 }
1560
sev_send_update_data(struct kvm * kvm,struct kvm_sev_cmd * argp)1561 static int sev_send_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1562 {
1563 struct sev_data_send_update_data data;
1564 struct kvm_sev_send_update_data params;
1565 void *hdr, *trans_data;
1566 struct page **guest_page;
1567 unsigned long n;
1568 int ret, offset;
1569
1570 if (!sev_guest(kvm))
1571 return -ENOTTY;
1572
1573 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data),
1574 sizeof(struct kvm_sev_send_update_data)))
1575 return -EFAULT;
1576
1577 /* userspace wants to query either header or trans length */
1578 if (!params.trans_len || !params.hdr_len)
1579 return __sev_send_update_data_query_lengths(kvm, argp, ¶ms);
1580
1581 if (!params.trans_uaddr || !params.guest_uaddr ||
1582 !params.guest_len || !params.hdr_uaddr)
1583 return -EINVAL;
1584
1585 /* Check if we are crossing the page boundary */
1586 offset = params.guest_uaddr & (PAGE_SIZE - 1);
1587 if (params.guest_len > PAGE_SIZE || (params.guest_len + offset) > PAGE_SIZE)
1588 return -EINVAL;
1589
1590 /* Pin guest memory */
1591 guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1592 PAGE_SIZE, &n, 0);
1593 if (IS_ERR(guest_page))
1594 return PTR_ERR(guest_page);
1595
1596 /* allocate memory for header and transport buffer */
1597 ret = -ENOMEM;
1598 hdr = kzalloc(params.hdr_len, GFP_KERNEL);
1599 if (!hdr)
1600 goto e_unpin;
1601
1602 trans_data = kzalloc(params.trans_len, GFP_KERNEL);
1603 if (!trans_data)
1604 goto e_free_hdr;
1605
1606 memset(&data, 0, sizeof(data));
1607 data.hdr_address = __psp_pa(hdr);
1608 data.hdr_len = params.hdr_len;
1609 data.trans_address = __psp_pa(trans_data);
1610 data.trans_len = params.trans_len;
1611
1612 /* The SEND_UPDATE_DATA command requires C-bit to be always set. */
1613 data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1614 data.guest_address |= sev_me_mask;
1615 data.guest_len = params.guest_len;
1616 data.handle = to_kvm_sev_info(kvm)->handle;
1617
1618 ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1619
1620 if (ret)
1621 goto e_free_trans_data;
1622
1623 /* copy transport buffer to user space */
1624 if (copy_to_user(u64_to_user_ptr(params.trans_uaddr),
1625 trans_data, params.trans_len)) {
1626 ret = -EFAULT;
1627 goto e_free_trans_data;
1628 }
1629
1630 /* Copy packet header to userspace. */
1631 if (copy_to_user(u64_to_user_ptr(params.hdr_uaddr), hdr,
1632 params.hdr_len))
1633 ret = -EFAULT;
1634
1635 e_free_trans_data:
1636 kfree(trans_data);
1637 e_free_hdr:
1638 kfree(hdr);
1639 e_unpin:
1640 sev_unpin_memory(kvm, guest_page, n);
1641
1642 return ret;
1643 }
1644
sev_send_finish(struct kvm * kvm,struct kvm_sev_cmd * argp)1645 static int sev_send_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1646 {
1647 struct sev_data_send_finish data;
1648
1649 if (!sev_guest(kvm))
1650 return -ENOTTY;
1651
1652 data.handle = to_kvm_sev_info(kvm)->handle;
1653 return sev_issue_cmd(kvm, SEV_CMD_SEND_FINISH, &data, &argp->error);
1654 }
1655
sev_send_cancel(struct kvm * kvm,struct kvm_sev_cmd * argp)1656 static int sev_send_cancel(struct kvm *kvm, struct kvm_sev_cmd *argp)
1657 {
1658 struct sev_data_send_cancel data;
1659
1660 if (!sev_guest(kvm))
1661 return -ENOTTY;
1662
1663 data.handle = to_kvm_sev_info(kvm)->handle;
1664 return sev_issue_cmd(kvm, SEV_CMD_SEND_CANCEL, &data, &argp->error);
1665 }
1666
sev_receive_start(struct kvm * kvm,struct kvm_sev_cmd * argp)1667 static int sev_receive_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1668 {
1669 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
1670 struct sev_data_receive_start start;
1671 struct kvm_sev_receive_start params;
1672 int *error = &argp->error;
1673 void *session_data;
1674 void *pdh_data;
1675 int ret;
1676
1677 if (!sev_guest(kvm))
1678 return -ENOTTY;
1679
1680 /* Get parameter from the userspace */
1681 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data),
1682 sizeof(struct kvm_sev_receive_start)))
1683 return -EFAULT;
1684
1685 /* some sanity checks */
1686 if (!params.pdh_uaddr || !params.pdh_len ||
1687 !params.session_uaddr || !params.session_len)
1688 return -EINVAL;
1689
1690 pdh_data = psp_copy_user_blob(params.pdh_uaddr, params.pdh_len);
1691 if (IS_ERR(pdh_data))
1692 return PTR_ERR(pdh_data);
1693
1694 session_data = psp_copy_user_blob(params.session_uaddr,
1695 params.session_len);
1696 if (IS_ERR(session_data)) {
1697 ret = PTR_ERR(session_data);
1698 goto e_free_pdh;
1699 }
1700
1701 memset(&start, 0, sizeof(start));
1702 start.handle = params.handle;
1703 start.policy = params.policy;
1704 start.pdh_cert_address = __psp_pa(pdh_data);
1705 start.pdh_cert_len = params.pdh_len;
1706 start.session_address = __psp_pa(session_data);
1707 start.session_len = params.session_len;
1708
1709 /* create memory encryption context */
1710 ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_RECEIVE_START, &start,
1711 error);
1712 if (ret)
1713 goto e_free_session;
1714
1715 /* Bind ASID to this guest */
1716 ret = sev_bind_asid(kvm, start.handle, error);
1717 if (ret) {
1718 sev_decommission(start.handle);
1719 goto e_free_session;
1720 }
1721
1722 params.handle = start.handle;
1723 if (copy_to_user(u64_to_user_ptr(argp->data),
1724 ¶ms, sizeof(struct kvm_sev_receive_start))) {
1725 ret = -EFAULT;
1726 sev_unbind_asid(kvm, start.handle);
1727 goto e_free_session;
1728 }
1729
1730 sev->handle = start.handle;
1731 sev->fd = argp->sev_fd;
1732
1733 e_free_session:
1734 kfree(session_data);
1735 e_free_pdh:
1736 kfree(pdh_data);
1737
1738 return ret;
1739 }
1740
sev_receive_update_data(struct kvm * kvm,struct kvm_sev_cmd * argp)1741 static int sev_receive_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1742 {
1743 struct kvm_sev_receive_update_data params;
1744 struct sev_data_receive_update_data data;
1745 void *hdr = NULL, *trans = NULL;
1746 struct page **guest_page;
1747 unsigned long n;
1748 int ret, offset;
1749
1750 if (!sev_guest(kvm))
1751 return -EINVAL;
1752
1753 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data),
1754 sizeof(struct kvm_sev_receive_update_data)))
1755 return -EFAULT;
1756
1757 if (!params.hdr_uaddr || !params.hdr_len ||
1758 !params.guest_uaddr || !params.guest_len ||
1759 !params.trans_uaddr || !params.trans_len)
1760 return -EINVAL;
1761
1762 /* Check if we are crossing the page boundary */
1763 offset = params.guest_uaddr & (PAGE_SIZE - 1);
1764 if (params.guest_len > PAGE_SIZE || (params.guest_len + offset) > PAGE_SIZE)
1765 return -EINVAL;
1766
1767 hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1768 if (IS_ERR(hdr))
1769 return PTR_ERR(hdr);
1770
1771 trans = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1772 if (IS_ERR(trans)) {
1773 ret = PTR_ERR(trans);
1774 goto e_free_hdr;
1775 }
1776
1777 memset(&data, 0, sizeof(data));
1778 data.hdr_address = __psp_pa(hdr);
1779 data.hdr_len = params.hdr_len;
1780 data.trans_address = __psp_pa(trans);
1781 data.trans_len = params.trans_len;
1782
1783 /* Pin guest memory */
1784 guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1785 PAGE_SIZE, &n, FOLL_WRITE);
1786 if (IS_ERR(guest_page)) {
1787 ret = PTR_ERR(guest_page);
1788 goto e_free_trans;
1789 }
1790
1791 /*
1792 * Flush (on non-coherent CPUs) before RECEIVE_UPDATE_DATA, the PSP
1793 * encrypts the written data with the guest's key, and the cache may
1794 * contain dirty, unencrypted data.
1795 */
1796 sev_clflush_pages(guest_page, n);
1797
1798 /* The RECEIVE_UPDATE_DATA command requires C-bit to be always set. */
1799 data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1800 data.guest_address |= sev_me_mask;
1801 data.guest_len = params.guest_len;
1802 data.handle = to_kvm_sev_info(kvm)->handle;
1803
1804 ret = sev_issue_cmd(kvm, SEV_CMD_RECEIVE_UPDATE_DATA, &data,
1805 &argp->error);
1806
1807 sev_unpin_memory(kvm, guest_page, n);
1808
1809 e_free_trans:
1810 kfree(trans);
1811 e_free_hdr:
1812 kfree(hdr);
1813
1814 return ret;
1815 }
1816
sev_receive_finish(struct kvm * kvm,struct kvm_sev_cmd * argp)1817 static int sev_receive_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1818 {
1819 struct sev_data_receive_finish data;
1820
1821 if (!sev_guest(kvm))
1822 return -ENOTTY;
1823
1824 data.handle = to_kvm_sev_info(kvm)->handle;
1825 return sev_issue_cmd(kvm, SEV_CMD_RECEIVE_FINISH, &data, &argp->error);
1826 }
1827
is_cmd_allowed_from_mirror(u32 cmd_id)1828 static bool is_cmd_allowed_from_mirror(u32 cmd_id)
1829 {
1830 /*
1831 * Allow mirrors VM to call KVM_SEV_LAUNCH_UPDATE_VMSA to enable SEV-ES
1832 * active mirror VMs. Also allow the debugging and status commands.
1833 */
1834 if (cmd_id == KVM_SEV_LAUNCH_UPDATE_VMSA ||
1835 cmd_id == KVM_SEV_GUEST_STATUS || cmd_id == KVM_SEV_DBG_DECRYPT ||
1836 cmd_id == KVM_SEV_DBG_ENCRYPT)
1837 return true;
1838
1839 return false;
1840 }
1841
sev_lock_two_vms(struct kvm * dst_kvm,struct kvm * src_kvm)1842 static int sev_lock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1843 {
1844 struct kvm_sev_info *dst_sev = to_kvm_sev_info(dst_kvm);
1845 struct kvm_sev_info *src_sev = to_kvm_sev_info(src_kvm);
1846 int r = -EBUSY;
1847
1848 if (dst_kvm == src_kvm)
1849 return -EINVAL;
1850
1851 /*
1852 * Bail if these VMs are already involved in a migration to avoid
1853 * deadlock between two VMs trying to migrate to/from each other.
1854 */
1855 if (atomic_cmpxchg_acquire(&dst_sev->migration_in_progress, 0, 1))
1856 return -EBUSY;
1857
1858 if (atomic_cmpxchg_acquire(&src_sev->migration_in_progress, 0, 1))
1859 goto release_dst;
1860
1861 r = -EINTR;
1862 if (mutex_lock_killable(&dst_kvm->lock))
1863 goto release_src;
1864 if (mutex_lock_killable_nested(&src_kvm->lock, SINGLE_DEPTH_NESTING))
1865 goto unlock_dst;
1866 return 0;
1867
1868 unlock_dst:
1869 mutex_unlock(&dst_kvm->lock);
1870 release_src:
1871 atomic_set_release(&src_sev->migration_in_progress, 0);
1872 release_dst:
1873 atomic_set_release(&dst_sev->migration_in_progress, 0);
1874 return r;
1875 }
1876
sev_unlock_two_vms(struct kvm * dst_kvm,struct kvm * src_kvm)1877 static void sev_unlock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1878 {
1879 struct kvm_sev_info *dst_sev = to_kvm_sev_info(dst_kvm);
1880 struct kvm_sev_info *src_sev = to_kvm_sev_info(src_kvm);
1881
1882 mutex_unlock(&dst_kvm->lock);
1883 mutex_unlock(&src_kvm->lock);
1884 atomic_set_release(&dst_sev->migration_in_progress, 0);
1885 atomic_set_release(&src_sev->migration_in_progress, 0);
1886 }
1887
sev_migrate_from(struct kvm * dst_kvm,struct kvm * src_kvm)1888 static void sev_migrate_from(struct kvm *dst_kvm, struct kvm *src_kvm)
1889 {
1890 struct kvm_sev_info *dst = to_kvm_sev_info(dst_kvm);
1891 struct kvm_sev_info *src = to_kvm_sev_info(src_kvm);
1892 struct kvm_vcpu *dst_vcpu, *src_vcpu;
1893 struct vcpu_svm *dst_svm, *src_svm;
1894 struct kvm_sev_info *mirror;
1895 unsigned long i;
1896
1897 dst->active = true;
1898 dst->asid = src->asid;
1899 dst->handle = src->handle;
1900 dst->pages_locked = src->pages_locked;
1901 dst->enc_context_owner = src->enc_context_owner;
1902 dst->es_active = src->es_active;
1903 dst->vmsa_features = src->vmsa_features;
1904
1905 src->asid = 0;
1906 src->active = false;
1907 src->handle = 0;
1908 src->pages_locked = 0;
1909 src->enc_context_owner = NULL;
1910 src->es_active = false;
1911
1912 list_cut_before(&dst->regions_list, &src->regions_list, &src->regions_list);
1913
1914 /*
1915 * If this VM has mirrors, "transfer" each mirror's refcount of the
1916 * source to the destination (this KVM). The caller holds a reference
1917 * to the source, so there's no danger of use-after-free.
1918 */
1919 list_cut_before(&dst->mirror_vms, &src->mirror_vms, &src->mirror_vms);
1920 list_for_each_entry(mirror, &dst->mirror_vms, mirror_entry) {
1921 kvm_get_kvm(dst_kvm);
1922 kvm_put_kvm(src_kvm);
1923 mirror->enc_context_owner = dst_kvm;
1924 }
1925
1926 /*
1927 * If this VM is a mirror, remove the old mirror from the owners list
1928 * and add the new mirror to the list.
1929 */
1930 if (is_mirroring_enc_context(dst_kvm)) {
1931 struct kvm_sev_info *owner_sev_info = to_kvm_sev_info(dst->enc_context_owner);
1932
1933 list_del(&src->mirror_entry);
1934 list_add_tail(&dst->mirror_entry, &owner_sev_info->mirror_vms);
1935 }
1936
1937 kvm_for_each_vcpu(i, dst_vcpu, dst_kvm) {
1938 dst_svm = to_svm(dst_vcpu);
1939
1940 sev_init_vmcb(dst_svm);
1941
1942 if (!dst->es_active)
1943 continue;
1944
1945 /*
1946 * Note, the source is not required to have the same number of
1947 * vCPUs as the destination when migrating a vanilla SEV VM.
1948 */
1949 src_vcpu = kvm_get_vcpu(src_kvm, i);
1950 src_svm = to_svm(src_vcpu);
1951
1952 /*
1953 * Transfer VMSA and GHCB state to the destination. Nullify and
1954 * clear source fields as appropriate, the state now belongs to
1955 * the destination.
1956 */
1957 memcpy(&dst_svm->sev_es, &src_svm->sev_es, sizeof(src_svm->sev_es));
1958 dst_svm->vmcb->control.ghcb_gpa = src_svm->vmcb->control.ghcb_gpa;
1959 dst_svm->vmcb->control.vmsa_pa = src_svm->vmcb->control.vmsa_pa;
1960 dst_vcpu->arch.guest_state_protected = true;
1961
1962 memset(&src_svm->sev_es, 0, sizeof(src_svm->sev_es));
1963 src_svm->vmcb->control.ghcb_gpa = INVALID_PAGE;
1964 src_svm->vmcb->control.vmsa_pa = INVALID_PAGE;
1965 src_vcpu->arch.guest_state_protected = false;
1966 }
1967 }
1968
sev_check_source_vcpus(struct kvm * dst,struct kvm * src)1969 static int sev_check_source_vcpus(struct kvm *dst, struct kvm *src)
1970 {
1971 struct kvm_vcpu *src_vcpu;
1972 unsigned long i;
1973
1974 if (src->created_vcpus != atomic_read(&src->online_vcpus) ||
1975 dst->created_vcpus != atomic_read(&dst->online_vcpus))
1976 return -EBUSY;
1977
1978 if (!sev_es_guest(src))
1979 return 0;
1980
1981 if (atomic_read(&src->online_vcpus) != atomic_read(&dst->online_vcpus))
1982 return -EINVAL;
1983
1984 kvm_for_each_vcpu(i, src_vcpu, src) {
1985 if (!src_vcpu->arch.guest_state_protected)
1986 return -EINVAL;
1987 }
1988
1989 return 0;
1990 }
1991
sev_vm_move_enc_context_from(struct kvm * kvm,unsigned int source_fd)1992 int sev_vm_move_enc_context_from(struct kvm *kvm, unsigned int source_fd)
1993 {
1994 struct kvm_sev_info *dst_sev = to_kvm_sev_info(kvm);
1995 struct kvm_sev_info *src_sev, *cg_cleanup_sev;
1996 CLASS(fd, f)(source_fd);
1997 struct kvm *source_kvm;
1998 bool charged = false;
1999 int ret;
2000
2001 if (fd_empty(f))
2002 return -EBADF;
2003
2004 if (!file_is_kvm(fd_file(f)))
2005 return -EBADF;
2006
2007 source_kvm = fd_file(f)->private_data;
2008 ret = sev_lock_two_vms(kvm, source_kvm);
2009 if (ret)
2010 return ret;
2011
2012 if (kvm->arch.vm_type != source_kvm->arch.vm_type ||
2013 sev_guest(kvm) || !sev_guest(source_kvm)) {
2014 ret = -EINVAL;
2015 goto out_unlock;
2016 }
2017
2018 src_sev = to_kvm_sev_info(source_kvm);
2019
2020 dst_sev->misc_cg = get_current_misc_cg();
2021 cg_cleanup_sev = dst_sev;
2022 if (dst_sev->misc_cg != src_sev->misc_cg) {
2023 ret = sev_misc_cg_try_charge(dst_sev);
2024 if (ret)
2025 goto out_dst_cgroup;
2026 charged = true;
2027 }
2028
2029 ret = kvm_lock_all_vcpus(kvm);
2030 if (ret)
2031 goto out_dst_cgroup;
2032 ret = kvm_lock_all_vcpus(source_kvm);
2033 if (ret)
2034 goto out_dst_vcpu;
2035
2036 ret = sev_check_source_vcpus(kvm, source_kvm);
2037 if (ret)
2038 goto out_source_vcpu;
2039
2040 sev_migrate_from(kvm, source_kvm);
2041 kvm_vm_dead(source_kvm);
2042 cg_cleanup_sev = src_sev;
2043 ret = 0;
2044
2045 out_source_vcpu:
2046 kvm_unlock_all_vcpus(source_kvm);
2047 out_dst_vcpu:
2048 kvm_unlock_all_vcpus(kvm);
2049 out_dst_cgroup:
2050 /* Operates on the source on success, on the destination on failure. */
2051 if (charged)
2052 sev_misc_cg_uncharge(cg_cleanup_sev);
2053 put_misc_cg(cg_cleanup_sev->misc_cg);
2054 cg_cleanup_sev->misc_cg = NULL;
2055 out_unlock:
2056 sev_unlock_two_vms(kvm, source_kvm);
2057 return ret;
2058 }
2059
sev_dev_get_attr(u32 group,u64 attr,u64 * val)2060 int sev_dev_get_attr(u32 group, u64 attr, u64 *val)
2061 {
2062 if (group != KVM_X86_GRP_SEV)
2063 return -ENXIO;
2064
2065 switch (attr) {
2066 case KVM_X86_SEV_VMSA_FEATURES:
2067 *val = sev_supported_vmsa_features;
2068 return 0;
2069
2070 default:
2071 return -ENXIO;
2072 }
2073 }
2074
2075 /*
2076 * The guest context contains all the information, keys and metadata
2077 * associated with the guest that the firmware tracks to implement SEV
2078 * and SNP features. The firmware stores the guest context in hypervisor
2079 * provide page via the SNP_GCTX_CREATE command.
2080 */
snp_context_create(struct kvm * kvm,struct kvm_sev_cmd * argp)2081 static void *snp_context_create(struct kvm *kvm, struct kvm_sev_cmd *argp)
2082 {
2083 struct sev_data_snp_addr data = {};
2084 void *context;
2085 int rc;
2086
2087 /* Allocate memory for context page */
2088 context = snp_alloc_firmware_page(GFP_KERNEL_ACCOUNT);
2089 if (!context)
2090 return NULL;
2091
2092 data.address = __psp_pa(context);
2093 rc = __sev_issue_cmd(argp->sev_fd, SEV_CMD_SNP_GCTX_CREATE, &data, &argp->error);
2094 if (rc) {
2095 pr_warn("Failed to create SEV-SNP context, rc %d fw_error %d",
2096 rc, argp->error);
2097 snp_free_firmware_page(context);
2098 return NULL;
2099 }
2100
2101 return context;
2102 }
2103
snp_bind_asid(struct kvm * kvm,int * error)2104 static int snp_bind_asid(struct kvm *kvm, int *error)
2105 {
2106 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2107 struct sev_data_snp_activate data = {0};
2108
2109 data.gctx_paddr = __psp_pa(sev->snp_context);
2110 data.asid = sev_get_asid(kvm);
2111 return sev_issue_cmd(kvm, SEV_CMD_SNP_ACTIVATE, &data, error);
2112 }
2113
snp_launch_start(struct kvm * kvm,struct kvm_sev_cmd * argp)2114 static int snp_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
2115 {
2116 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2117 struct sev_data_snp_launch_start start = {0};
2118 struct kvm_sev_snp_launch_start params;
2119 int rc;
2120
2121 if (!sev_snp_guest(kvm))
2122 return -ENOTTY;
2123
2124 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data), sizeof(params)))
2125 return -EFAULT;
2126
2127 /* Don't allow userspace to allocate memory for more than 1 SNP context. */
2128 if (sev->snp_context)
2129 return -EINVAL;
2130
2131 if (params.flags)
2132 return -EINVAL;
2133
2134 if (params.policy & ~SNP_POLICY_MASK_VALID)
2135 return -EINVAL;
2136
2137 /* Check for policy bits that must be set */
2138 if (!(params.policy & SNP_POLICY_MASK_RSVD_MBO) ||
2139 !(params.policy & SNP_POLICY_MASK_SMT))
2140 return -EINVAL;
2141
2142 if (params.policy & SNP_POLICY_MASK_SINGLE_SOCKET)
2143 return -EINVAL;
2144
2145 sev->policy = params.policy;
2146
2147 sev->snp_context = snp_context_create(kvm, argp);
2148 if (!sev->snp_context)
2149 return -ENOTTY;
2150
2151 start.gctx_paddr = __psp_pa(sev->snp_context);
2152 start.policy = params.policy;
2153 memcpy(start.gosvw, params.gosvw, sizeof(params.gosvw));
2154 rc = __sev_issue_cmd(argp->sev_fd, SEV_CMD_SNP_LAUNCH_START, &start, &argp->error);
2155 if (rc) {
2156 pr_debug("%s: SEV_CMD_SNP_LAUNCH_START firmware command failed, rc %d\n",
2157 __func__, rc);
2158 goto e_free_context;
2159 }
2160
2161 sev->fd = argp->sev_fd;
2162 rc = snp_bind_asid(kvm, &argp->error);
2163 if (rc) {
2164 pr_debug("%s: Failed to bind ASID to SEV-SNP context, rc %d\n",
2165 __func__, rc);
2166 goto e_free_context;
2167 }
2168
2169 return 0;
2170
2171 e_free_context:
2172 snp_decommission_context(kvm);
2173
2174 return rc;
2175 }
2176
2177 struct sev_gmem_populate_args {
2178 __u8 type;
2179 int sev_fd;
2180 int fw_error;
2181 };
2182
sev_gmem_post_populate(struct kvm * kvm,gfn_t gfn_start,kvm_pfn_t pfn,void __user * src,int order,void * opaque)2183 static int sev_gmem_post_populate(struct kvm *kvm, gfn_t gfn_start, kvm_pfn_t pfn,
2184 void __user *src, int order, void *opaque)
2185 {
2186 struct sev_gmem_populate_args *sev_populate_args = opaque;
2187 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2188 int n_private = 0, ret, i;
2189 int npages = (1 << order);
2190 gfn_t gfn;
2191
2192 if (WARN_ON_ONCE(sev_populate_args->type != KVM_SEV_SNP_PAGE_TYPE_ZERO && !src))
2193 return -EINVAL;
2194
2195 for (gfn = gfn_start, i = 0; gfn < gfn_start + npages; gfn++, i++) {
2196 struct sev_data_snp_launch_update fw_args = {0};
2197 bool assigned = false;
2198 int level;
2199
2200 ret = snp_lookup_rmpentry((u64)pfn + i, &assigned, &level);
2201 if (ret || assigned) {
2202 pr_debug("%s: Failed to ensure GFN 0x%llx RMP entry is initial shared state, ret: %d assigned: %d\n",
2203 __func__, gfn, ret, assigned);
2204 ret = ret ? -EINVAL : -EEXIST;
2205 goto err;
2206 }
2207
2208 if (src) {
2209 void *vaddr = kmap_local_pfn(pfn + i);
2210
2211 if (copy_from_user(vaddr, src + i * PAGE_SIZE, PAGE_SIZE)) {
2212 ret = -EFAULT;
2213 goto err;
2214 }
2215 kunmap_local(vaddr);
2216 }
2217
2218 ret = rmp_make_private(pfn + i, gfn << PAGE_SHIFT, PG_LEVEL_4K,
2219 sev_get_asid(kvm), true);
2220 if (ret)
2221 goto err;
2222
2223 n_private++;
2224
2225 fw_args.gctx_paddr = __psp_pa(sev->snp_context);
2226 fw_args.address = __sme_set(pfn_to_hpa(pfn + i));
2227 fw_args.page_size = PG_LEVEL_TO_RMP(PG_LEVEL_4K);
2228 fw_args.page_type = sev_populate_args->type;
2229
2230 ret = __sev_issue_cmd(sev_populate_args->sev_fd, SEV_CMD_SNP_LAUNCH_UPDATE,
2231 &fw_args, &sev_populate_args->fw_error);
2232 if (ret)
2233 goto fw_err;
2234 }
2235
2236 return 0;
2237
2238 fw_err:
2239 /*
2240 * If the firmware command failed handle the reclaim and cleanup of that
2241 * PFN specially vs. prior pages which can be cleaned up below without
2242 * needing to reclaim in advance.
2243 *
2244 * Additionally, when invalid CPUID function entries are detected,
2245 * firmware writes the expected values into the page and leaves it
2246 * unencrypted so it can be used for debugging and error-reporting.
2247 *
2248 * Copy this page back into the source buffer so userspace can use this
2249 * information to provide information on which CPUID leaves/fields
2250 * failed CPUID validation.
2251 */
2252 if (!snp_page_reclaim(kvm, pfn + i) &&
2253 sev_populate_args->type == KVM_SEV_SNP_PAGE_TYPE_CPUID &&
2254 sev_populate_args->fw_error == SEV_RET_INVALID_PARAM) {
2255 void *vaddr = kmap_local_pfn(pfn + i);
2256
2257 if (copy_to_user(src + i * PAGE_SIZE, vaddr, PAGE_SIZE))
2258 pr_debug("Failed to write CPUID page back to userspace\n");
2259
2260 kunmap_local(vaddr);
2261 }
2262
2263 /* pfn + i is hypervisor-owned now, so skip below cleanup for it. */
2264 n_private--;
2265
2266 err:
2267 pr_debug("%s: exiting with error ret %d (fw_error %d), restoring %d gmem PFNs to shared.\n",
2268 __func__, ret, sev_populate_args->fw_error, n_private);
2269 for (i = 0; i < n_private; i++)
2270 kvm_rmp_make_shared(kvm, pfn + i, PG_LEVEL_4K);
2271
2272 return ret;
2273 }
2274
snp_launch_update(struct kvm * kvm,struct kvm_sev_cmd * argp)2275 static int snp_launch_update(struct kvm *kvm, struct kvm_sev_cmd *argp)
2276 {
2277 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2278 struct sev_gmem_populate_args sev_populate_args = {0};
2279 struct kvm_sev_snp_launch_update params;
2280 struct kvm_memory_slot *memslot;
2281 long npages, count;
2282 void __user *src;
2283 int ret = 0;
2284
2285 if (!sev_snp_guest(kvm) || !sev->snp_context)
2286 return -EINVAL;
2287
2288 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data), sizeof(params)))
2289 return -EFAULT;
2290
2291 pr_debug("%s: GFN start 0x%llx length 0x%llx type %d flags %d\n", __func__,
2292 params.gfn_start, params.len, params.type, params.flags);
2293
2294 if (!PAGE_ALIGNED(params.len) || params.flags ||
2295 (params.type != KVM_SEV_SNP_PAGE_TYPE_NORMAL &&
2296 params.type != KVM_SEV_SNP_PAGE_TYPE_ZERO &&
2297 params.type != KVM_SEV_SNP_PAGE_TYPE_UNMEASURED &&
2298 params.type != KVM_SEV_SNP_PAGE_TYPE_SECRETS &&
2299 params.type != KVM_SEV_SNP_PAGE_TYPE_CPUID))
2300 return -EINVAL;
2301
2302 npages = params.len / PAGE_SIZE;
2303
2304 /*
2305 * For each GFN that's being prepared as part of the initial guest
2306 * state, the following pre-conditions are verified:
2307 *
2308 * 1) The backing memslot is a valid private memslot.
2309 * 2) The GFN has been set to private via KVM_SET_MEMORY_ATTRIBUTES
2310 * beforehand.
2311 * 3) The PFN of the guest_memfd has not already been set to private
2312 * in the RMP table.
2313 *
2314 * The KVM MMU relies on kvm->mmu_invalidate_seq to retry nested page
2315 * faults if there's a race between a fault and an attribute update via
2316 * KVM_SET_MEMORY_ATTRIBUTES, and a similar approach could be utilized
2317 * here. However, kvm->slots_lock guards against both this as well as
2318 * concurrent memslot updates occurring while these checks are being
2319 * performed, so use that here to make it easier to reason about the
2320 * initial expected state and better guard against unexpected
2321 * situations.
2322 */
2323 mutex_lock(&kvm->slots_lock);
2324
2325 memslot = gfn_to_memslot(kvm, params.gfn_start);
2326 if (!kvm_slot_can_be_private(memslot)) {
2327 ret = -EINVAL;
2328 goto out;
2329 }
2330
2331 sev_populate_args.sev_fd = argp->sev_fd;
2332 sev_populate_args.type = params.type;
2333 src = params.type == KVM_SEV_SNP_PAGE_TYPE_ZERO ? NULL : u64_to_user_ptr(params.uaddr);
2334
2335 count = kvm_gmem_populate(kvm, params.gfn_start, src, npages,
2336 sev_gmem_post_populate, &sev_populate_args);
2337 if (count < 0) {
2338 argp->error = sev_populate_args.fw_error;
2339 pr_debug("%s: kvm_gmem_populate failed, ret %ld (fw_error %d)\n",
2340 __func__, count, argp->error);
2341 ret = -EIO;
2342 } else {
2343 params.gfn_start += count;
2344 params.len -= count * PAGE_SIZE;
2345 if (params.type != KVM_SEV_SNP_PAGE_TYPE_ZERO)
2346 params.uaddr += count * PAGE_SIZE;
2347
2348 ret = 0;
2349 if (copy_to_user(u64_to_user_ptr(argp->data), ¶ms, sizeof(params)))
2350 ret = -EFAULT;
2351 }
2352
2353 out:
2354 mutex_unlock(&kvm->slots_lock);
2355
2356 return ret;
2357 }
2358
snp_launch_update_vmsa(struct kvm * kvm,struct kvm_sev_cmd * argp)2359 static int snp_launch_update_vmsa(struct kvm *kvm, struct kvm_sev_cmd *argp)
2360 {
2361 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2362 struct sev_data_snp_launch_update data = {};
2363 struct kvm_vcpu *vcpu;
2364 unsigned long i;
2365 int ret;
2366
2367 data.gctx_paddr = __psp_pa(sev->snp_context);
2368 data.page_type = SNP_PAGE_TYPE_VMSA;
2369
2370 kvm_for_each_vcpu(i, vcpu, kvm) {
2371 struct vcpu_svm *svm = to_svm(vcpu);
2372 u64 pfn = __pa(svm->sev_es.vmsa) >> PAGE_SHIFT;
2373
2374 ret = sev_es_sync_vmsa(svm);
2375 if (ret)
2376 return ret;
2377
2378 /* Transition the VMSA page to a firmware state. */
2379 ret = rmp_make_private(pfn, INITIAL_VMSA_GPA, PG_LEVEL_4K, sev->asid, true);
2380 if (ret)
2381 return ret;
2382
2383 /* Issue the SNP command to encrypt the VMSA */
2384 data.address = __sme_pa(svm->sev_es.vmsa);
2385 ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_SNP_LAUNCH_UPDATE,
2386 &data, &argp->error);
2387 if (ret) {
2388 snp_page_reclaim(kvm, pfn);
2389
2390 return ret;
2391 }
2392
2393 svm->vcpu.arch.guest_state_protected = true;
2394 /*
2395 * SEV-ES (and thus SNP) guest mandates LBR Virtualization to
2396 * be _always_ ON. Enable it only after setting
2397 * guest_state_protected because KVM_SET_MSRS allows dynamic
2398 * toggling of LBRV (for performance reason) on write access to
2399 * MSR_IA32_DEBUGCTLMSR when guest_state_protected is not set.
2400 */
2401 svm_enable_lbrv(vcpu);
2402 }
2403
2404 return 0;
2405 }
2406
snp_launch_finish(struct kvm * kvm,struct kvm_sev_cmd * argp)2407 static int snp_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
2408 {
2409 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2410 struct kvm_sev_snp_launch_finish params;
2411 struct sev_data_snp_launch_finish *data;
2412 void *id_block = NULL, *id_auth = NULL;
2413 int ret;
2414
2415 if (!sev_snp_guest(kvm))
2416 return -ENOTTY;
2417
2418 if (!sev->snp_context)
2419 return -EINVAL;
2420
2421 if (copy_from_user(¶ms, u64_to_user_ptr(argp->data), sizeof(params)))
2422 return -EFAULT;
2423
2424 if (params.flags)
2425 return -EINVAL;
2426
2427 /* Measure all vCPUs using LAUNCH_UPDATE before finalizing the launch flow. */
2428 ret = snp_launch_update_vmsa(kvm, argp);
2429 if (ret)
2430 return ret;
2431
2432 data = kzalloc(sizeof(*data), GFP_KERNEL_ACCOUNT);
2433 if (!data)
2434 return -ENOMEM;
2435
2436 if (params.id_block_en) {
2437 id_block = psp_copy_user_blob(params.id_block_uaddr, KVM_SEV_SNP_ID_BLOCK_SIZE);
2438 if (IS_ERR(id_block)) {
2439 ret = PTR_ERR(id_block);
2440 goto e_free;
2441 }
2442
2443 data->id_block_en = 1;
2444 data->id_block_paddr = __sme_pa(id_block);
2445
2446 id_auth = psp_copy_user_blob(params.id_auth_uaddr, KVM_SEV_SNP_ID_AUTH_SIZE);
2447 if (IS_ERR(id_auth)) {
2448 ret = PTR_ERR(id_auth);
2449 goto e_free_id_block;
2450 }
2451
2452 data->id_auth_paddr = __sme_pa(id_auth);
2453
2454 if (params.auth_key_en)
2455 data->auth_key_en = 1;
2456 }
2457
2458 data->vcek_disabled = params.vcek_disabled;
2459
2460 memcpy(data->host_data, params.host_data, KVM_SEV_SNP_FINISH_DATA_SIZE);
2461 data->gctx_paddr = __psp_pa(sev->snp_context);
2462 ret = sev_issue_cmd(kvm, SEV_CMD_SNP_LAUNCH_FINISH, data, &argp->error);
2463
2464 /*
2465 * Now that there will be no more SNP_LAUNCH_UPDATE ioctls, private pages
2466 * can be given to the guest simply by marking the RMP entry as private.
2467 * This can happen on first access and also with KVM_PRE_FAULT_MEMORY.
2468 */
2469 if (!ret)
2470 kvm->arch.pre_fault_allowed = true;
2471
2472 kfree(id_auth);
2473
2474 e_free_id_block:
2475 kfree(id_block);
2476
2477 e_free:
2478 kfree(data);
2479
2480 return ret;
2481 }
2482
sev_mem_enc_ioctl(struct kvm * kvm,void __user * argp)2483 int sev_mem_enc_ioctl(struct kvm *kvm, void __user *argp)
2484 {
2485 struct kvm_sev_cmd sev_cmd;
2486 int r;
2487
2488 if (!sev_enabled)
2489 return -ENOTTY;
2490
2491 if (!argp)
2492 return 0;
2493
2494 if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
2495 return -EFAULT;
2496
2497 mutex_lock(&kvm->lock);
2498
2499 /* Only the enc_context_owner handles some memory enc operations. */
2500 if (is_mirroring_enc_context(kvm) &&
2501 !is_cmd_allowed_from_mirror(sev_cmd.id)) {
2502 r = -EINVAL;
2503 goto out;
2504 }
2505
2506 /*
2507 * Once KVM_SEV_INIT2 initializes a KVM instance as an SNP guest, only
2508 * allow the use of SNP-specific commands.
2509 */
2510 if (sev_snp_guest(kvm) && sev_cmd.id < KVM_SEV_SNP_LAUNCH_START) {
2511 r = -EPERM;
2512 goto out;
2513 }
2514
2515 switch (sev_cmd.id) {
2516 case KVM_SEV_ES_INIT:
2517 if (!sev_es_enabled) {
2518 r = -ENOTTY;
2519 goto out;
2520 }
2521 fallthrough;
2522 case KVM_SEV_INIT:
2523 r = sev_guest_init(kvm, &sev_cmd);
2524 break;
2525 case KVM_SEV_INIT2:
2526 r = sev_guest_init2(kvm, &sev_cmd);
2527 break;
2528 case KVM_SEV_LAUNCH_START:
2529 r = sev_launch_start(kvm, &sev_cmd);
2530 break;
2531 case KVM_SEV_LAUNCH_UPDATE_DATA:
2532 r = sev_launch_update_data(kvm, &sev_cmd);
2533 break;
2534 case KVM_SEV_LAUNCH_UPDATE_VMSA:
2535 r = sev_launch_update_vmsa(kvm, &sev_cmd);
2536 break;
2537 case KVM_SEV_LAUNCH_MEASURE:
2538 r = sev_launch_measure(kvm, &sev_cmd);
2539 break;
2540 case KVM_SEV_LAUNCH_FINISH:
2541 r = sev_launch_finish(kvm, &sev_cmd);
2542 break;
2543 case KVM_SEV_GUEST_STATUS:
2544 r = sev_guest_status(kvm, &sev_cmd);
2545 break;
2546 case KVM_SEV_DBG_DECRYPT:
2547 r = sev_dbg_crypt(kvm, &sev_cmd, true);
2548 break;
2549 case KVM_SEV_DBG_ENCRYPT:
2550 r = sev_dbg_crypt(kvm, &sev_cmd, false);
2551 break;
2552 case KVM_SEV_LAUNCH_SECRET:
2553 r = sev_launch_secret(kvm, &sev_cmd);
2554 break;
2555 case KVM_SEV_GET_ATTESTATION_REPORT:
2556 r = sev_get_attestation_report(kvm, &sev_cmd);
2557 break;
2558 case KVM_SEV_SEND_START:
2559 r = sev_send_start(kvm, &sev_cmd);
2560 break;
2561 case KVM_SEV_SEND_UPDATE_DATA:
2562 r = sev_send_update_data(kvm, &sev_cmd);
2563 break;
2564 case KVM_SEV_SEND_FINISH:
2565 r = sev_send_finish(kvm, &sev_cmd);
2566 break;
2567 case KVM_SEV_SEND_CANCEL:
2568 r = sev_send_cancel(kvm, &sev_cmd);
2569 break;
2570 case KVM_SEV_RECEIVE_START:
2571 r = sev_receive_start(kvm, &sev_cmd);
2572 break;
2573 case KVM_SEV_RECEIVE_UPDATE_DATA:
2574 r = sev_receive_update_data(kvm, &sev_cmd);
2575 break;
2576 case KVM_SEV_RECEIVE_FINISH:
2577 r = sev_receive_finish(kvm, &sev_cmd);
2578 break;
2579 case KVM_SEV_SNP_LAUNCH_START:
2580 r = snp_launch_start(kvm, &sev_cmd);
2581 break;
2582 case KVM_SEV_SNP_LAUNCH_UPDATE:
2583 r = snp_launch_update(kvm, &sev_cmd);
2584 break;
2585 case KVM_SEV_SNP_LAUNCH_FINISH:
2586 r = snp_launch_finish(kvm, &sev_cmd);
2587 break;
2588 default:
2589 r = -EINVAL;
2590 goto out;
2591 }
2592
2593 if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
2594 r = -EFAULT;
2595
2596 out:
2597 mutex_unlock(&kvm->lock);
2598 return r;
2599 }
2600
sev_mem_enc_register_region(struct kvm * kvm,struct kvm_enc_region * range)2601 int sev_mem_enc_register_region(struct kvm *kvm,
2602 struct kvm_enc_region *range)
2603 {
2604 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2605 struct enc_region *region;
2606 int ret = 0;
2607
2608 if (!sev_guest(kvm))
2609 return -ENOTTY;
2610
2611 /* If kvm is mirroring encryption context it isn't responsible for it */
2612 if (is_mirroring_enc_context(kvm))
2613 return -EINVAL;
2614
2615 if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
2616 return -EINVAL;
2617
2618 region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
2619 if (!region)
2620 return -ENOMEM;
2621
2622 mutex_lock(&kvm->lock);
2623 region->pages = sev_pin_memory(kvm, range->addr, range->size, ®ion->npages,
2624 FOLL_WRITE | FOLL_LONGTERM);
2625 if (IS_ERR(region->pages)) {
2626 ret = PTR_ERR(region->pages);
2627 mutex_unlock(&kvm->lock);
2628 goto e_free;
2629 }
2630
2631 /*
2632 * The guest may change the memory encryption attribute from C=0 -> C=1
2633 * or vice versa for this memory range. Lets make sure caches are
2634 * flushed to ensure that guest data gets written into memory with
2635 * correct C-bit. Note, this must be done before dropping kvm->lock,
2636 * as region and its array of pages can be freed by a different task
2637 * once kvm->lock is released.
2638 */
2639 sev_clflush_pages(region->pages, region->npages);
2640
2641 region->uaddr = range->addr;
2642 region->size = range->size;
2643
2644 list_add_tail(®ion->list, &sev->regions_list);
2645 mutex_unlock(&kvm->lock);
2646
2647 return ret;
2648
2649 e_free:
2650 kfree(region);
2651 return ret;
2652 }
2653
2654 static struct enc_region *
find_enc_region(struct kvm * kvm,struct kvm_enc_region * range)2655 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
2656 {
2657 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2658 struct list_head *head = &sev->regions_list;
2659 struct enc_region *i;
2660
2661 list_for_each_entry(i, head, list) {
2662 if (i->uaddr == range->addr &&
2663 i->size == range->size)
2664 return i;
2665 }
2666
2667 return NULL;
2668 }
2669
__unregister_enc_region_locked(struct kvm * kvm,struct enc_region * region)2670 static void __unregister_enc_region_locked(struct kvm *kvm,
2671 struct enc_region *region)
2672 {
2673 sev_unpin_memory(kvm, region->pages, region->npages);
2674 list_del(®ion->list);
2675 kfree(region);
2676 }
2677
sev_mem_enc_unregister_region(struct kvm * kvm,struct kvm_enc_region * range)2678 int sev_mem_enc_unregister_region(struct kvm *kvm,
2679 struct kvm_enc_region *range)
2680 {
2681 struct enc_region *region;
2682 int ret;
2683
2684 /* If kvm is mirroring encryption context it isn't responsible for it */
2685 if (is_mirroring_enc_context(kvm))
2686 return -EINVAL;
2687
2688 mutex_lock(&kvm->lock);
2689
2690 if (!sev_guest(kvm)) {
2691 ret = -ENOTTY;
2692 goto failed;
2693 }
2694
2695 region = find_enc_region(kvm, range);
2696 if (!region) {
2697 ret = -EINVAL;
2698 goto failed;
2699 }
2700
2701 /*
2702 * Ensure that all guest tagged cache entries are flushed before
2703 * releasing the pages back to the system for use. CLFLUSH will
2704 * not do this, so issue a WBINVD.
2705 */
2706 wbinvd_on_all_cpus();
2707
2708 __unregister_enc_region_locked(kvm, region);
2709
2710 mutex_unlock(&kvm->lock);
2711 return 0;
2712
2713 failed:
2714 mutex_unlock(&kvm->lock);
2715 return ret;
2716 }
2717
sev_vm_copy_enc_context_from(struct kvm * kvm,unsigned int source_fd)2718 int sev_vm_copy_enc_context_from(struct kvm *kvm, unsigned int source_fd)
2719 {
2720 CLASS(fd, f)(source_fd);
2721 struct kvm *source_kvm;
2722 struct kvm_sev_info *source_sev, *mirror_sev;
2723 int ret;
2724
2725 if (fd_empty(f))
2726 return -EBADF;
2727
2728 if (!file_is_kvm(fd_file(f)))
2729 return -EBADF;
2730
2731 source_kvm = fd_file(f)->private_data;
2732 ret = sev_lock_two_vms(kvm, source_kvm);
2733 if (ret)
2734 return ret;
2735
2736 /*
2737 * Mirrors of mirrors should work, but let's not get silly. Also
2738 * disallow out-of-band SEV/SEV-ES init if the target is already an
2739 * SEV guest, or if vCPUs have been created. KVM relies on vCPUs being
2740 * created after SEV/SEV-ES initialization, e.g. to init intercepts.
2741 */
2742 if (sev_guest(kvm) || !sev_guest(source_kvm) ||
2743 is_mirroring_enc_context(source_kvm) || kvm->created_vcpus) {
2744 ret = -EINVAL;
2745 goto e_unlock;
2746 }
2747
2748 /*
2749 * The mirror kvm holds an enc_context_owner ref so its asid can't
2750 * disappear until we're done with it
2751 */
2752 source_sev = to_kvm_sev_info(source_kvm);
2753 kvm_get_kvm(source_kvm);
2754 mirror_sev = to_kvm_sev_info(kvm);
2755 list_add_tail(&mirror_sev->mirror_entry, &source_sev->mirror_vms);
2756
2757 /* Set enc_context_owner and copy its encryption context over */
2758 mirror_sev->enc_context_owner = source_kvm;
2759 mirror_sev->active = true;
2760 mirror_sev->asid = source_sev->asid;
2761 mirror_sev->fd = source_sev->fd;
2762 mirror_sev->es_active = source_sev->es_active;
2763 mirror_sev->need_init = false;
2764 mirror_sev->handle = source_sev->handle;
2765 INIT_LIST_HEAD(&mirror_sev->regions_list);
2766 INIT_LIST_HEAD(&mirror_sev->mirror_vms);
2767 ret = 0;
2768
2769 /*
2770 * Do not copy ap_jump_table. Since the mirror does not share the same
2771 * KVM contexts as the original, and they may have different
2772 * memory-views.
2773 */
2774
2775 e_unlock:
2776 sev_unlock_two_vms(kvm, source_kvm);
2777 return ret;
2778 }
2779
snp_decommission_context(struct kvm * kvm)2780 static int snp_decommission_context(struct kvm *kvm)
2781 {
2782 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2783 struct sev_data_snp_addr data = {};
2784 int ret;
2785
2786 /* If context is not created then do nothing */
2787 if (!sev->snp_context)
2788 return 0;
2789
2790 /* Do the decommision, which will unbind the ASID from the SNP context */
2791 data.address = __sme_pa(sev->snp_context);
2792 down_write(&sev_deactivate_lock);
2793 ret = sev_do_cmd(SEV_CMD_SNP_DECOMMISSION, &data, NULL);
2794 up_write(&sev_deactivate_lock);
2795
2796 if (WARN_ONCE(ret, "Failed to release guest context, ret %d", ret))
2797 return ret;
2798
2799 snp_free_firmware_page(sev->snp_context);
2800 sev->snp_context = NULL;
2801
2802 return 0;
2803 }
2804
sev_vm_destroy(struct kvm * kvm)2805 void sev_vm_destroy(struct kvm *kvm)
2806 {
2807 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
2808 struct list_head *head = &sev->regions_list;
2809 struct list_head *pos, *q;
2810
2811 if (!sev_guest(kvm))
2812 return;
2813
2814 WARN_ON(!list_empty(&sev->mirror_vms));
2815
2816 /* If this is a mirror_kvm release the enc_context_owner and skip sev cleanup */
2817 if (is_mirroring_enc_context(kvm)) {
2818 struct kvm *owner_kvm = sev->enc_context_owner;
2819
2820 mutex_lock(&owner_kvm->lock);
2821 list_del(&sev->mirror_entry);
2822 mutex_unlock(&owner_kvm->lock);
2823 kvm_put_kvm(owner_kvm);
2824 return;
2825 }
2826
2827 /*
2828 * Ensure that all guest tagged cache entries are flushed before
2829 * releasing the pages back to the system for use. CLFLUSH will
2830 * not do this, so issue a WBINVD.
2831 */
2832 wbinvd_on_all_cpus();
2833
2834 /*
2835 * if userspace was terminated before unregistering the memory regions
2836 * then lets unpin all the registered memory.
2837 */
2838 if (!list_empty(head)) {
2839 list_for_each_safe(pos, q, head) {
2840 __unregister_enc_region_locked(kvm,
2841 list_entry(pos, struct enc_region, list));
2842 cond_resched();
2843 }
2844 }
2845
2846 if (sev_snp_guest(kvm)) {
2847 snp_guest_req_cleanup(kvm);
2848
2849 /*
2850 * Decomission handles unbinding of the ASID. If it fails for
2851 * some unexpected reason, just leak the ASID.
2852 */
2853 if (snp_decommission_context(kvm))
2854 return;
2855 } else {
2856 sev_unbind_asid(kvm, sev->handle);
2857 }
2858
2859 sev_asid_free(sev);
2860 }
2861
sev_set_cpu_caps(void)2862 void __init sev_set_cpu_caps(void)
2863 {
2864 if (sev_enabled) {
2865 kvm_cpu_cap_set(X86_FEATURE_SEV);
2866 kvm_caps.supported_vm_types |= BIT(KVM_X86_SEV_VM);
2867 }
2868 if (sev_es_enabled) {
2869 kvm_cpu_cap_set(X86_FEATURE_SEV_ES);
2870 kvm_caps.supported_vm_types |= BIT(KVM_X86_SEV_ES_VM);
2871 }
2872 if (sev_snp_enabled) {
2873 kvm_cpu_cap_set(X86_FEATURE_SEV_SNP);
2874 kvm_caps.supported_vm_types |= BIT(KVM_X86_SNP_VM);
2875 }
2876 }
2877
is_sev_snp_initialized(void)2878 static bool is_sev_snp_initialized(void)
2879 {
2880 struct sev_user_data_snp_status *status;
2881 struct sev_data_snp_addr buf;
2882 bool initialized = false;
2883 int ret, error = 0;
2884
2885 status = snp_alloc_firmware_page(GFP_KERNEL | __GFP_ZERO);
2886 if (!status)
2887 return false;
2888
2889 buf.address = __psp_pa(status);
2890 ret = sev_do_cmd(SEV_CMD_SNP_PLATFORM_STATUS, &buf, &error);
2891 if (ret) {
2892 pr_err("SEV: SNP_PLATFORM_STATUS failed ret=%d, fw_error=%d (%#x)\n",
2893 ret, error, error);
2894 goto out;
2895 }
2896
2897 initialized = !!status->state;
2898
2899 out:
2900 snp_free_firmware_page(status);
2901
2902 return initialized;
2903 }
2904
sev_hardware_setup(void)2905 void __init sev_hardware_setup(void)
2906 {
2907 unsigned int eax, ebx, ecx, edx, sev_asid_count, sev_es_asid_count;
2908 struct sev_platform_init_args init_args = {0};
2909 bool sev_snp_supported = false;
2910 bool sev_es_supported = false;
2911 bool sev_supported = false;
2912
2913 if (!sev_enabled || !npt_enabled || !nrips)
2914 goto out;
2915
2916 /*
2917 * SEV must obviously be supported in hardware. Sanity check that the
2918 * CPU supports decode assists, which is mandatory for SEV guests to
2919 * support instruction emulation. Ditto for flushing by ASID, as SEV
2920 * guests are bound to a single ASID, i.e. KVM can't rotate to a new
2921 * ASID to effect a TLB flush.
2922 */
2923 if (!boot_cpu_has(X86_FEATURE_SEV) ||
2924 WARN_ON_ONCE(!boot_cpu_has(X86_FEATURE_DECODEASSISTS)) ||
2925 WARN_ON_ONCE(!boot_cpu_has(X86_FEATURE_FLUSHBYASID)))
2926 goto out;
2927
2928 /*
2929 * The kernel's initcall infrastructure lacks the ability to express
2930 * dependencies between initcalls, whereas the modules infrastructure
2931 * automatically handles dependencies via symbol loading. Ensure the
2932 * PSP SEV driver is initialized before proceeding if KVM is built-in,
2933 * as the dependency isn't handled by the initcall infrastructure.
2934 */
2935 if (IS_BUILTIN(CONFIG_KVM_AMD) && sev_module_init())
2936 goto out;
2937
2938 /* Retrieve SEV CPUID information */
2939 cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
2940
2941 /* Set encryption bit location for SEV-ES guests */
2942 sev_enc_bit = ebx & 0x3f;
2943
2944 /* Maximum number of encrypted guests supported simultaneously */
2945 max_sev_asid = ecx;
2946 if (!max_sev_asid)
2947 goto out;
2948
2949 /* Minimum ASID value that should be used for SEV guest */
2950 min_sev_asid = edx;
2951 sev_me_mask = 1UL << (ebx & 0x3f);
2952
2953 /*
2954 * Initialize SEV ASID bitmaps. Allocate space for ASID 0 in the bitmap,
2955 * even though it's never used, so that the bitmap is indexed by the
2956 * actual ASID.
2957 */
2958 nr_asids = max_sev_asid + 1;
2959 sev_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2960 if (!sev_asid_bitmap)
2961 goto out;
2962
2963 sev_reclaim_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2964 if (!sev_reclaim_asid_bitmap) {
2965 bitmap_free(sev_asid_bitmap);
2966 sev_asid_bitmap = NULL;
2967 goto out;
2968 }
2969
2970 if (min_sev_asid <= max_sev_asid) {
2971 sev_asid_count = max_sev_asid - min_sev_asid + 1;
2972 WARN_ON_ONCE(misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count));
2973 }
2974 sev_supported = true;
2975
2976 /* SEV-ES support requested? */
2977 if (!sev_es_enabled)
2978 goto out;
2979
2980 /*
2981 * SEV-ES requires MMIO caching as KVM doesn't have access to the guest
2982 * instruction stream, i.e. can't emulate in response to a #NPF and
2983 * instead relies on #NPF(RSVD) being reflected into the guest as #VC
2984 * (the guest can then do a #VMGEXIT to request MMIO emulation).
2985 */
2986 if (!enable_mmio_caching)
2987 goto out;
2988
2989 /* Does the CPU support SEV-ES? */
2990 if (!boot_cpu_has(X86_FEATURE_SEV_ES))
2991 goto out;
2992
2993 if (!lbrv) {
2994 WARN_ONCE(!boot_cpu_has(X86_FEATURE_LBRV),
2995 "LBRV must be present for SEV-ES support");
2996 goto out;
2997 }
2998
2999 /* Has the system been allocated ASIDs for SEV-ES? */
3000 if (min_sev_asid == 1)
3001 goto out;
3002
3003 sev_es_asid_count = min_sev_asid - 1;
3004 WARN_ON_ONCE(misc_cg_set_capacity(MISC_CG_RES_SEV_ES, sev_es_asid_count));
3005 sev_es_supported = true;
3006 sev_snp_supported = sev_snp_enabled && cc_platform_has(CC_ATTR_HOST_SEV_SNP);
3007
3008 out:
3009 if (sev_enabled) {
3010 init_args.probe = true;
3011 if (sev_platform_init(&init_args))
3012 sev_supported = sev_es_supported = sev_snp_supported = false;
3013 else if (sev_snp_supported)
3014 sev_snp_supported = is_sev_snp_initialized();
3015 }
3016
3017 if (boot_cpu_has(X86_FEATURE_SEV))
3018 pr_info("SEV %s (ASIDs %u - %u)\n",
3019 sev_supported ? min_sev_asid <= max_sev_asid ? "enabled" :
3020 "unusable" :
3021 "disabled",
3022 min_sev_asid, max_sev_asid);
3023 if (boot_cpu_has(X86_FEATURE_SEV_ES))
3024 pr_info("SEV-ES %s (ASIDs %u - %u)\n",
3025 str_enabled_disabled(sev_es_supported),
3026 min_sev_asid > 1 ? 1 : 0, min_sev_asid - 1);
3027 if (boot_cpu_has(X86_FEATURE_SEV_SNP))
3028 pr_info("SEV-SNP %s (ASIDs %u - %u)\n",
3029 str_enabled_disabled(sev_snp_supported),
3030 min_sev_asid > 1 ? 1 : 0, min_sev_asid - 1);
3031
3032 sev_enabled = sev_supported;
3033 sev_es_enabled = sev_es_supported;
3034 sev_snp_enabled = sev_snp_supported;
3035
3036 if (!sev_es_enabled || !cpu_feature_enabled(X86_FEATURE_DEBUG_SWAP) ||
3037 !cpu_feature_enabled(X86_FEATURE_NO_NESTED_DATA_BP))
3038 sev_es_debug_swap_enabled = false;
3039
3040 sev_supported_vmsa_features = 0;
3041 if (sev_es_debug_swap_enabled)
3042 sev_supported_vmsa_features |= SVM_SEV_FEAT_DEBUG_SWAP;
3043 }
3044
sev_hardware_unsetup(void)3045 void sev_hardware_unsetup(void)
3046 {
3047 if (!sev_enabled)
3048 return;
3049
3050 /* No need to take sev_bitmap_lock, all VMs have been destroyed. */
3051 sev_flush_asids(1, max_sev_asid);
3052
3053 bitmap_free(sev_asid_bitmap);
3054 bitmap_free(sev_reclaim_asid_bitmap);
3055
3056 misc_cg_set_capacity(MISC_CG_RES_SEV, 0);
3057 misc_cg_set_capacity(MISC_CG_RES_SEV_ES, 0);
3058
3059 sev_platform_shutdown();
3060 }
3061
sev_cpu_init(struct svm_cpu_data * sd)3062 int sev_cpu_init(struct svm_cpu_data *sd)
3063 {
3064 if (!sev_enabled)
3065 return 0;
3066
3067 sd->sev_vmcbs = kcalloc(nr_asids, sizeof(void *), GFP_KERNEL);
3068 if (!sd->sev_vmcbs)
3069 return -ENOMEM;
3070
3071 return 0;
3072 }
3073
3074 /*
3075 * Pages used by hardware to hold guest encrypted state must be flushed before
3076 * returning them to the system.
3077 */
sev_flush_encrypted_page(struct kvm_vcpu * vcpu,void * va)3078 static void sev_flush_encrypted_page(struct kvm_vcpu *vcpu, void *va)
3079 {
3080 unsigned int asid = sev_get_asid(vcpu->kvm);
3081
3082 /*
3083 * Note! The address must be a kernel address, as regular page walk
3084 * checks are performed by VM_PAGE_FLUSH, i.e. operating on a user
3085 * address is non-deterministic and unsafe. This function deliberately
3086 * takes a pointer to deter passing in a user address.
3087 */
3088 unsigned long addr = (unsigned long)va;
3089
3090 /*
3091 * If CPU enforced cache coherency for encrypted mappings of the
3092 * same physical page is supported, use CLFLUSHOPT instead. NOTE: cache
3093 * flush is still needed in order to work properly with DMA devices.
3094 */
3095 if (boot_cpu_has(X86_FEATURE_SME_COHERENT)) {
3096 clflush_cache_range(va, PAGE_SIZE);
3097 return;
3098 }
3099
3100 /*
3101 * VM Page Flush takes a host virtual address and a guest ASID. Fall
3102 * back to WBINVD if this faults so as not to make any problems worse
3103 * by leaving stale encrypted data in the cache.
3104 */
3105 if (WARN_ON_ONCE(wrmsrq_safe(MSR_AMD64_VM_PAGE_FLUSH, addr | asid)))
3106 goto do_wbinvd;
3107
3108 return;
3109
3110 do_wbinvd:
3111 wbinvd_on_all_cpus();
3112 }
3113
sev_guest_memory_reclaimed(struct kvm * kvm)3114 void sev_guest_memory_reclaimed(struct kvm *kvm)
3115 {
3116 /*
3117 * With SNP+gmem, private/encrypted memory is unreachable via the
3118 * hva-based mmu notifiers, so these events are only actually
3119 * pertaining to shared pages where there is no need to perform
3120 * the WBINVD to flush associated caches.
3121 */
3122 if (!sev_guest(kvm) || sev_snp_guest(kvm))
3123 return;
3124
3125 wbinvd_on_all_cpus();
3126 }
3127
sev_free_vcpu(struct kvm_vcpu * vcpu)3128 void sev_free_vcpu(struct kvm_vcpu *vcpu)
3129 {
3130 struct vcpu_svm *svm;
3131
3132 if (!sev_es_guest(vcpu->kvm))
3133 return;
3134
3135 svm = to_svm(vcpu);
3136
3137 /*
3138 * If it's an SNP guest, then the VMSA was marked in the RMP table as
3139 * a guest-owned page. Transition the page to hypervisor state before
3140 * releasing it back to the system.
3141 */
3142 if (sev_snp_guest(vcpu->kvm)) {
3143 u64 pfn = __pa(svm->sev_es.vmsa) >> PAGE_SHIFT;
3144
3145 if (kvm_rmp_make_shared(vcpu->kvm, pfn, PG_LEVEL_4K))
3146 goto skip_vmsa_free;
3147 }
3148
3149 if (vcpu->arch.guest_state_protected)
3150 sev_flush_encrypted_page(vcpu, svm->sev_es.vmsa);
3151
3152 __free_page(virt_to_page(svm->sev_es.vmsa));
3153
3154 skip_vmsa_free:
3155 if (svm->sev_es.ghcb_sa_free)
3156 kvfree(svm->sev_es.ghcb_sa);
3157 }
3158
kvm_ghcb_get_sw_exit_code(struct vmcb_control_area * control)3159 static u64 kvm_ghcb_get_sw_exit_code(struct vmcb_control_area *control)
3160 {
3161 return (((u64)control->exit_code_hi) << 32) | control->exit_code;
3162 }
3163
dump_ghcb(struct vcpu_svm * svm)3164 static void dump_ghcb(struct vcpu_svm *svm)
3165 {
3166 struct vmcb_control_area *control = &svm->vmcb->control;
3167 unsigned int nbits;
3168
3169 /* Re-use the dump_invalid_vmcb module parameter */
3170 if (!dump_invalid_vmcb) {
3171 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
3172 return;
3173 }
3174
3175 nbits = sizeof(svm->sev_es.valid_bitmap) * 8;
3176
3177 /*
3178 * Print KVM's snapshot of the GHCB values that were (unsuccessfully)
3179 * used to handle the exit. If the guest has since modified the GHCB
3180 * itself, dumping the raw GHCB won't help debug why KVM was unable to
3181 * handle the VMGEXIT that KVM observed.
3182 */
3183 pr_err("GHCB (GPA=%016llx) snapshot:\n", svm->vmcb->control.ghcb_gpa);
3184 pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
3185 kvm_ghcb_get_sw_exit_code(control), kvm_ghcb_sw_exit_code_is_valid(svm));
3186 pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
3187 control->exit_info_1, kvm_ghcb_sw_exit_info_1_is_valid(svm));
3188 pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
3189 control->exit_info_2, kvm_ghcb_sw_exit_info_2_is_valid(svm));
3190 pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
3191 svm->sev_es.sw_scratch, kvm_ghcb_sw_scratch_is_valid(svm));
3192 pr_err("%-20s%*pb\n", "valid_bitmap", nbits, svm->sev_es.valid_bitmap);
3193 }
3194
sev_es_sync_to_ghcb(struct vcpu_svm * svm)3195 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
3196 {
3197 struct kvm_vcpu *vcpu = &svm->vcpu;
3198 struct ghcb *ghcb = svm->sev_es.ghcb;
3199
3200 /*
3201 * The GHCB protocol so far allows for the following data
3202 * to be returned:
3203 * GPRs RAX, RBX, RCX, RDX
3204 *
3205 * Copy their values, even if they may not have been written during the
3206 * VM-Exit. It's the guest's responsibility to not consume random data.
3207 */
3208 ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
3209 ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
3210 ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
3211 ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
3212 }
3213
sev_es_sync_from_ghcb(struct vcpu_svm * svm)3214 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
3215 {
3216 struct vmcb_control_area *control = &svm->vmcb->control;
3217 struct kvm_vcpu *vcpu = &svm->vcpu;
3218 struct ghcb *ghcb = svm->sev_es.ghcb;
3219 u64 exit_code;
3220
3221 /*
3222 * The GHCB protocol so far allows for the following data
3223 * to be supplied:
3224 * GPRs RAX, RBX, RCX, RDX
3225 * XCR0
3226 * CPL
3227 *
3228 * VMMCALL allows the guest to provide extra registers. KVM also
3229 * expects RSI for hypercalls, so include that, too.
3230 *
3231 * Copy their values to the appropriate location if supplied.
3232 */
3233 memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
3234
3235 BUILD_BUG_ON(sizeof(svm->sev_es.valid_bitmap) != sizeof(ghcb->save.valid_bitmap));
3236 memcpy(&svm->sev_es.valid_bitmap, &ghcb->save.valid_bitmap, sizeof(ghcb->save.valid_bitmap));
3237
3238 vcpu->arch.regs[VCPU_REGS_RAX] = kvm_ghcb_get_rax_if_valid(svm, ghcb);
3239 vcpu->arch.regs[VCPU_REGS_RBX] = kvm_ghcb_get_rbx_if_valid(svm, ghcb);
3240 vcpu->arch.regs[VCPU_REGS_RCX] = kvm_ghcb_get_rcx_if_valid(svm, ghcb);
3241 vcpu->arch.regs[VCPU_REGS_RDX] = kvm_ghcb_get_rdx_if_valid(svm, ghcb);
3242 vcpu->arch.regs[VCPU_REGS_RSI] = kvm_ghcb_get_rsi_if_valid(svm, ghcb);
3243
3244 svm->vmcb->save.cpl = kvm_ghcb_get_cpl_if_valid(svm, ghcb);
3245
3246 if (kvm_ghcb_xcr0_is_valid(svm)) {
3247 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
3248 vcpu->arch.cpuid_dynamic_bits_dirty = true;
3249 }
3250
3251 /* Copy the GHCB exit information into the VMCB fields */
3252 exit_code = ghcb_get_sw_exit_code(ghcb);
3253 control->exit_code = lower_32_bits(exit_code);
3254 control->exit_code_hi = upper_32_bits(exit_code);
3255 control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
3256 control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
3257 svm->sev_es.sw_scratch = kvm_ghcb_get_sw_scratch_if_valid(svm, ghcb);
3258
3259 /* Clear the valid entries fields */
3260 memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
3261 }
3262
sev_es_validate_vmgexit(struct vcpu_svm * svm)3263 static int sev_es_validate_vmgexit(struct vcpu_svm *svm)
3264 {
3265 struct vmcb_control_area *control = &svm->vmcb->control;
3266 struct kvm_vcpu *vcpu = &svm->vcpu;
3267 u64 exit_code;
3268 u64 reason;
3269
3270 /*
3271 * Retrieve the exit code now even though it may not be marked valid
3272 * as it could help with debugging.
3273 */
3274 exit_code = kvm_ghcb_get_sw_exit_code(control);
3275
3276 /* Only GHCB Usage code 0 is supported */
3277 if (svm->sev_es.ghcb->ghcb_usage) {
3278 reason = GHCB_ERR_INVALID_USAGE;
3279 goto vmgexit_err;
3280 }
3281
3282 reason = GHCB_ERR_MISSING_INPUT;
3283
3284 if (!kvm_ghcb_sw_exit_code_is_valid(svm) ||
3285 !kvm_ghcb_sw_exit_info_1_is_valid(svm) ||
3286 !kvm_ghcb_sw_exit_info_2_is_valid(svm))
3287 goto vmgexit_err;
3288
3289 switch (exit_code) {
3290 case SVM_EXIT_READ_DR7:
3291 break;
3292 case SVM_EXIT_WRITE_DR7:
3293 if (!kvm_ghcb_rax_is_valid(svm))
3294 goto vmgexit_err;
3295 break;
3296 case SVM_EXIT_RDTSC:
3297 break;
3298 case SVM_EXIT_RDPMC:
3299 if (!kvm_ghcb_rcx_is_valid(svm))
3300 goto vmgexit_err;
3301 break;
3302 case SVM_EXIT_CPUID:
3303 if (!kvm_ghcb_rax_is_valid(svm) ||
3304 !kvm_ghcb_rcx_is_valid(svm))
3305 goto vmgexit_err;
3306 if (vcpu->arch.regs[VCPU_REGS_RAX] == 0xd)
3307 if (!kvm_ghcb_xcr0_is_valid(svm))
3308 goto vmgexit_err;
3309 break;
3310 case SVM_EXIT_INVD:
3311 break;
3312 case SVM_EXIT_IOIO:
3313 if (control->exit_info_1 & SVM_IOIO_STR_MASK) {
3314 if (!kvm_ghcb_sw_scratch_is_valid(svm))
3315 goto vmgexit_err;
3316 } else {
3317 if (!(control->exit_info_1 & SVM_IOIO_TYPE_MASK))
3318 if (!kvm_ghcb_rax_is_valid(svm))
3319 goto vmgexit_err;
3320 }
3321 break;
3322 case SVM_EXIT_MSR:
3323 if (!kvm_ghcb_rcx_is_valid(svm))
3324 goto vmgexit_err;
3325 if (control->exit_info_1) {
3326 if (!kvm_ghcb_rax_is_valid(svm) ||
3327 !kvm_ghcb_rdx_is_valid(svm))
3328 goto vmgexit_err;
3329 }
3330 break;
3331 case SVM_EXIT_VMMCALL:
3332 if (!kvm_ghcb_rax_is_valid(svm) ||
3333 !kvm_ghcb_cpl_is_valid(svm))
3334 goto vmgexit_err;
3335 break;
3336 case SVM_EXIT_RDTSCP:
3337 break;
3338 case SVM_EXIT_WBINVD:
3339 break;
3340 case SVM_EXIT_MONITOR:
3341 if (!kvm_ghcb_rax_is_valid(svm) ||
3342 !kvm_ghcb_rcx_is_valid(svm) ||
3343 !kvm_ghcb_rdx_is_valid(svm))
3344 goto vmgexit_err;
3345 break;
3346 case SVM_EXIT_MWAIT:
3347 if (!kvm_ghcb_rax_is_valid(svm) ||
3348 !kvm_ghcb_rcx_is_valid(svm))
3349 goto vmgexit_err;
3350 break;
3351 case SVM_VMGEXIT_MMIO_READ:
3352 case SVM_VMGEXIT_MMIO_WRITE:
3353 if (!kvm_ghcb_sw_scratch_is_valid(svm))
3354 goto vmgexit_err;
3355 break;
3356 case SVM_VMGEXIT_AP_CREATION:
3357 if (!sev_snp_guest(vcpu->kvm))
3358 goto vmgexit_err;
3359 if (lower_32_bits(control->exit_info_1) != SVM_VMGEXIT_AP_DESTROY)
3360 if (!kvm_ghcb_rax_is_valid(svm))
3361 goto vmgexit_err;
3362 break;
3363 case SVM_VMGEXIT_NMI_COMPLETE:
3364 case SVM_VMGEXIT_AP_HLT_LOOP:
3365 case SVM_VMGEXIT_AP_JUMP_TABLE:
3366 case SVM_VMGEXIT_UNSUPPORTED_EVENT:
3367 case SVM_VMGEXIT_HV_FEATURES:
3368 case SVM_VMGEXIT_TERM_REQUEST:
3369 break;
3370 case SVM_VMGEXIT_PSC:
3371 if (!sev_snp_guest(vcpu->kvm) || !kvm_ghcb_sw_scratch_is_valid(svm))
3372 goto vmgexit_err;
3373 break;
3374 case SVM_VMGEXIT_GUEST_REQUEST:
3375 case SVM_VMGEXIT_EXT_GUEST_REQUEST:
3376 if (!sev_snp_guest(vcpu->kvm) ||
3377 !PAGE_ALIGNED(control->exit_info_1) ||
3378 !PAGE_ALIGNED(control->exit_info_2) ||
3379 control->exit_info_1 == control->exit_info_2)
3380 goto vmgexit_err;
3381 break;
3382 default:
3383 reason = GHCB_ERR_INVALID_EVENT;
3384 goto vmgexit_err;
3385 }
3386
3387 return 0;
3388
3389 vmgexit_err:
3390 if (reason == GHCB_ERR_INVALID_USAGE) {
3391 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
3392 svm->sev_es.ghcb->ghcb_usage);
3393 } else if (reason == GHCB_ERR_INVALID_EVENT) {
3394 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx is not valid\n",
3395 exit_code);
3396 } else {
3397 vcpu_unimpl(vcpu, "vmgexit: exit code %#llx input is not valid\n",
3398 exit_code);
3399 dump_ghcb(svm);
3400 }
3401
3402 svm_vmgexit_bad_input(svm, reason);
3403
3404 /* Resume the guest to "return" the error code. */
3405 return 1;
3406 }
3407
sev_es_unmap_ghcb(struct vcpu_svm * svm)3408 void sev_es_unmap_ghcb(struct vcpu_svm *svm)
3409 {
3410 /* Clear any indication that the vCPU is in a type of AP Reset Hold */
3411 svm->sev_es.ap_reset_hold_type = AP_RESET_HOLD_NONE;
3412
3413 if (!svm->sev_es.ghcb)
3414 return;
3415
3416 if (svm->sev_es.ghcb_sa_free) {
3417 /*
3418 * The scratch area lives outside the GHCB, so there is a
3419 * buffer that, depending on the operation performed, may
3420 * need to be synced, then freed.
3421 */
3422 if (svm->sev_es.ghcb_sa_sync) {
3423 kvm_write_guest(svm->vcpu.kvm,
3424 svm->sev_es.sw_scratch,
3425 svm->sev_es.ghcb_sa,
3426 svm->sev_es.ghcb_sa_len);
3427 svm->sev_es.ghcb_sa_sync = false;
3428 }
3429
3430 kvfree(svm->sev_es.ghcb_sa);
3431 svm->sev_es.ghcb_sa = NULL;
3432 svm->sev_es.ghcb_sa_free = false;
3433 }
3434
3435 trace_kvm_vmgexit_exit(svm->vcpu.vcpu_id, svm->sev_es.ghcb);
3436
3437 sev_es_sync_to_ghcb(svm);
3438
3439 kvm_vcpu_unmap(&svm->vcpu, &svm->sev_es.ghcb_map);
3440 svm->sev_es.ghcb = NULL;
3441 }
3442
pre_sev_run(struct vcpu_svm * svm,int cpu)3443 int pre_sev_run(struct vcpu_svm *svm, int cpu)
3444 {
3445 struct svm_cpu_data *sd = per_cpu_ptr(&svm_data, cpu);
3446 struct kvm *kvm = svm->vcpu.kvm;
3447 unsigned int asid = sev_get_asid(kvm);
3448
3449 /*
3450 * Reject KVM_RUN if userspace attempts to run the vCPU with an invalid
3451 * VMSA, e.g. if userspace forces the vCPU to be RUNNABLE after an SNP
3452 * AP Destroy event.
3453 */
3454 if (sev_es_guest(kvm) && !VALID_PAGE(svm->vmcb->control.vmsa_pa))
3455 return -EINVAL;
3456
3457 /* Assign the asid allocated with this SEV guest */
3458 svm->asid = asid;
3459
3460 /*
3461 * Flush guest TLB:
3462 *
3463 * 1) when different VMCB for the same ASID is to be run on the same host CPU.
3464 * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
3465 */
3466 if (sd->sev_vmcbs[asid] == svm->vmcb &&
3467 svm->vcpu.arch.last_vmentry_cpu == cpu)
3468 return 0;
3469
3470 sd->sev_vmcbs[asid] = svm->vmcb;
3471 svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
3472 vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
3473 return 0;
3474 }
3475
3476 #define GHCB_SCRATCH_AREA_LIMIT (16ULL * PAGE_SIZE)
setup_vmgexit_scratch(struct vcpu_svm * svm,bool sync,u64 len)3477 static int setup_vmgexit_scratch(struct vcpu_svm *svm, bool sync, u64 len)
3478 {
3479 struct vmcb_control_area *control = &svm->vmcb->control;
3480 u64 ghcb_scratch_beg, ghcb_scratch_end;
3481 u64 scratch_gpa_beg, scratch_gpa_end;
3482 void *scratch_va;
3483
3484 scratch_gpa_beg = svm->sev_es.sw_scratch;
3485 if (!scratch_gpa_beg) {
3486 pr_err("vmgexit: scratch gpa not provided\n");
3487 goto e_scratch;
3488 }
3489
3490 scratch_gpa_end = scratch_gpa_beg + len;
3491 if (scratch_gpa_end < scratch_gpa_beg) {
3492 pr_err("vmgexit: scratch length (%#llx) not valid for scratch address (%#llx)\n",
3493 len, scratch_gpa_beg);
3494 goto e_scratch;
3495 }
3496
3497 if ((scratch_gpa_beg & PAGE_MASK) == control->ghcb_gpa) {
3498 /* Scratch area begins within GHCB */
3499 ghcb_scratch_beg = control->ghcb_gpa +
3500 offsetof(struct ghcb, shared_buffer);
3501 ghcb_scratch_end = control->ghcb_gpa +
3502 offsetof(struct ghcb, reserved_0xff0);
3503
3504 /*
3505 * If the scratch area begins within the GHCB, it must be
3506 * completely contained in the GHCB shared buffer area.
3507 */
3508 if (scratch_gpa_beg < ghcb_scratch_beg ||
3509 scratch_gpa_end > ghcb_scratch_end) {
3510 pr_err("vmgexit: scratch area is outside of GHCB shared buffer area (%#llx - %#llx)\n",
3511 scratch_gpa_beg, scratch_gpa_end);
3512 goto e_scratch;
3513 }
3514
3515 scratch_va = (void *)svm->sev_es.ghcb;
3516 scratch_va += (scratch_gpa_beg - control->ghcb_gpa);
3517 } else {
3518 /*
3519 * The guest memory must be read into a kernel buffer, so
3520 * limit the size
3521 */
3522 if (len > GHCB_SCRATCH_AREA_LIMIT) {
3523 pr_err("vmgexit: scratch area exceeds KVM limits (%#llx requested, %#llx limit)\n",
3524 len, GHCB_SCRATCH_AREA_LIMIT);
3525 goto e_scratch;
3526 }
3527 scratch_va = kvzalloc(len, GFP_KERNEL_ACCOUNT);
3528 if (!scratch_va)
3529 return -ENOMEM;
3530
3531 if (kvm_read_guest(svm->vcpu.kvm, scratch_gpa_beg, scratch_va, len)) {
3532 /* Unable to copy scratch area from guest */
3533 pr_err("vmgexit: kvm_read_guest for scratch area failed\n");
3534
3535 kvfree(scratch_va);
3536 return -EFAULT;
3537 }
3538
3539 /*
3540 * The scratch area is outside the GHCB. The operation will
3541 * dictate whether the buffer needs to be synced before running
3542 * the vCPU next time (i.e. a read was requested so the data
3543 * must be written back to the guest memory).
3544 */
3545 svm->sev_es.ghcb_sa_sync = sync;
3546 svm->sev_es.ghcb_sa_free = true;
3547 }
3548
3549 svm->sev_es.ghcb_sa = scratch_va;
3550 svm->sev_es.ghcb_sa_len = len;
3551
3552 return 0;
3553
3554 e_scratch:
3555 svm_vmgexit_bad_input(svm, GHCB_ERR_INVALID_SCRATCH_AREA);
3556
3557 return 1;
3558 }
3559
set_ghcb_msr_bits(struct vcpu_svm * svm,u64 value,u64 mask,unsigned int pos)3560 static void set_ghcb_msr_bits(struct vcpu_svm *svm, u64 value, u64 mask,
3561 unsigned int pos)
3562 {
3563 svm->vmcb->control.ghcb_gpa &= ~(mask << pos);
3564 svm->vmcb->control.ghcb_gpa |= (value & mask) << pos;
3565 }
3566
get_ghcb_msr_bits(struct vcpu_svm * svm,u64 mask,unsigned int pos)3567 static u64 get_ghcb_msr_bits(struct vcpu_svm *svm, u64 mask, unsigned int pos)
3568 {
3569 return (svm->vmcb->control.ghcb_gpa >> pos) & mask;
3570 }
3571
set_ghcb_msr(struct vcpu_svm * svm,u64 value)3572 static void set_ghcb_msr(struct vcpu_svm *svm, u64 value)
3573 {
3574 svm->vmcb->control.ghcb_gpa = value;
3575 }
3576
snp_rmptable_psmash(kvm_pfn_t pfn)3577 static int snp_rmptable_psmash(kvm_pfn_t pfn)
3578 {
3579 int ret;
3580
3581 pfn = pfn & ~(KVM_PAGES_PER_HPAGE(PG_LEVEL_2M) - 1);
3582
3583 /*
3584 * PSMASH_FAIL_INUSE indicates another processor is modifying the
3585 * entry, so retry until that's no longer the case.
3586 */
3587 do {
3588 ret = psmash(pfn);
3589 } while (ret == PSMASH_FAIL_INUSE);
3590
3591 return ret;
3592 }
3593
snp_complete_psc_msr(struct kvm_vcpu * vcpu)3594 static int snp_complete_psc_msr(struct kvm_vcpu *vcpu)
3595 {
3596 struct vcpu_svm *svm = to_svm(vcpu);
3597
3598 if (vcpu->run->hypercall.ret)
3599 set_ghcb_msr(svm, GHCB_MSR_PSC_RESP_ERROR);
3600 else
3601 set_ghcb_msr(svm, GHCB_MSR_PSC_RESP);
3602
3603 return 1; /* resume guest */
3604 }
3605
snp_begin_psc_msr(struct vcpu_svm * svm,u64 ghcb_msr)3606 static int snp_begin_psc_msr(struct vcpu_svm *svm, u64 ghcb_msr)
3607 {
3608 u64 gpa = gfn_to_gpa(GHCB_MSR_PSC_REQ_TO_GFN(ghcb_msr));
3609 u8 op = GHCB_MSR_PSC_REQ_TO_OP(ghcb_msr);
3610 struct kvm_vcpu *vcpu = &svm->vcpu;
3611
3612 if (op != SNP_PAGE_STATE_PRIVATE && op != SNP_PAGE_STATE_SHARED) {
3613 set_ghcb_msr(svm, GHCB_MSR_PSC_RESP_ERROR);
3614 return 1; /* resume guest */
3615 }
3616
3617 if (!user_exit_on_hypercall(vcpu->kvm, KVM_HC_MAP_GPA_RANGE)) {
3618 set_ghcb_msr(svm, GHCB_MSR_PSC_RESP_ERROR);
3619 return 1; /* resume guest */
3620 }
3621
3622 vcpu->run->exit_reason = KVM_EXIT_HYPERCALL;
3623 vcpu->run->hypercall.nr = KVM_HC_MAP_GPA_RANGE;
3624 /*
3625 * In principle this should have been -KVM_ENOSYS, but userspace (QEMU <=9.2)
3626 * assumed that vcpu->run->hypercall.ret is never changed by KVM and thus that
3627 * it was always zero on KVM_EXIT_HYPERCALL. Since KVM is now overwriting
3628 * vcpu->run->hypercall.ret, ensuring that it is zero to not break QEMU.
3629 */
3630 vcpu->run->hypercall.ret = 0;
3631 vcpu->run->hypercall.args[0] = gpa;
3632 vcpu->run->hypercall.args[1] = 1;
3633 vcpu->run->hypercall.args[2] = (op == SNP_PAGE_STATE_PRIVATE)
3634 ? KVM_MAP_GPA_RANGE_ENCRYPTED
3635 : KVM_MAP_GPA_RANGE_DECRYPTED;
3636 vcpu->run->hypercall.args[2] |= KVM_MAP_GPA_RANGE_PAGE_SZ_4K;
3637
3638 vcpu->arch.complete_userspace_io = snp_complete_psc_msr;
3639
3640 return 0; /* forward request to userspace */
3641 }
3642
3643 struct psc_buffer {
3644 struct psc_hdr hdr;
3645 struct psc_entry entries[];
3646 } __packed;
3647
3648 static int snp_begin_psc(struct vcpu_svm *svm, struct psc_buffer *psc);
3649
snp_complete_psc(struct vcpu_svm * svm,u64 psc_ret)3650 static void snp_complete_psc(struct vcpu_svm *svm, u64 psc_ret)
3651 {
3652 svm->sev_es.psc_inflight = 0;
3653 svm->sev_es.psc_idx = 0;
3654 svm->sev_es.psc_2m = false;
3655
3656 /*
3657 * PSC requests always get a "no action" response in SW_EXITINFO1, with
3658 * a PSC-specific return code in SW_EXITINFO2 that provides the "real"
3659 * return code. E.g. if the PSC request was interrupted, the need to
3660 * retry is communicated via SW_EXITINFO2, not SW_EXITINFO1.
3661 */
3662 svm_vmgexit_no_action(svm, psc_ret);
3663 }
3664
__snp_complete_one_psc(struct vcpu_svm * svm)3665 static void __snp_complete_one_psc(struct vcpu_svm *svm)
3666 {
3667 struct psc_buffer *psc = svm->sev_es.ghcb_sa;
3668 struct psc_entry *entries = psc->entries;
3669 struct psc_hdr *hdr = &psc->hdr;
3670 __u16 idx;
3671
3672 /*
3673 * Everything in-flight has been processed successfully. Update the
3674 * corresponding entries in the guest's PSC buffer and zero out the
3675 * count of in-flight PSC entries.
3676 */
3677 for (idx = svm->sev_es.psc_idx; svm->sev_es.psc_inflight;
3678 svm->sev_es.psc_inflight--, idx++) {
3679 struct psc_entry *entry = &entries[idx];
3680
3681 entry->cur_page = entry->pagesize ? 512 : 1;
3682 }
3683
3684 hdr->cur_entry = idx;
3685 }
3686
snp_complete_one_psc(struct kvm_vcpu * vcpu)3687 static int snp_complete_one_psc(struct kvm_vcpu *vcpu)
3688 {
3689 struct vcpu_svm *svm = to_svm(vcpu);
3690 struct psc_buffer *psc = svm->sev_es.ghcb_sa;
3691
3692 if (vcpu->run->hypercall.ret) {
3693 snp_complete_psc(svm, VMGEXIT_PSC_ERROR_GENERIC);
3694 return 1; /* resume guest */
3695 }
3696
3697 __snp_complete_one_psc(svm);
3698
3699 /* Handle the next range (if any). */
3700 return snp_begin_psc(svm, psc);
3701 }
3702
snp_begin_psc(struct vcpu_svm * svm,struct psc_buffer * psc)3703 static int snp_begin_psc(struct vcpu_svm *svm, struct psc_buffer *psc)
3704 {
3705 struct psc_entry *entries = psc->entries;
3706 struct kvm_vcpu *vcpu = &svm->vcpu;
3707 struct psc_hdr *hdr = &psc->hdr;
3708 struct psc_entry entry_start;
3709 u16 idx, idx_start, idx_end;
3710 int npages;
3711 bool huge;
3712 u64 gfn;
3713
3714 if (!user_exit_on_hypercall(vcpu->kvm, KVM_HC_MAP_GPA_RANGE)) {
3715 snp_complete_psc(svm, VMGEXIT_PSC_ERROR_GENERIC);
3716 return 1;
3717 }
3718
3719 next_range:
3720 /* There should be no other PSCs in-flight at this point. */
3721 if (WARN_ON_ONCE(svm->sev_es.psc_inflight)) {
3722 snp_complete_psc(svm, VMGEXIT_PSC_ERROR_GENERIC);
3723 return 1;
3724 }
3725
3726 /*
3727 * The PSC descriptor buffer can be modified by a misbehaved guest after
3728 * validation, so take care to only use validated copies of values used
3729 * for things like array indexing.
3730 */
3731 idx_start = hdr->cur_entry;
3732 idx_end = hdr->end_entry;
3733
3734 if (idx_end >= VMGEXIT_PSC_MAX_COUNT) {
3735 snp_complete_psc(svm, VMGEXIT_PSC_ERROR_INVALID_HDR);
3736 return 1;
3737 }
3738
3739 /* Find the start of the next range which needs processing. */
3740 for (idx = idx_start; idx <= idx_end; idx++, hdr->cur_entry++) {
3741 entry_start = entries[idx];
3742
3743 gfn = entry_start.gfn;
3744 huge = entry_start.pagesize;
3745 npages = huge ? 512 : 1;
3746
3747 if (entry_start.cur_page > npages || !IS_ALIGNED(gfn, npages)) {
3748 snp_complete_psc(svm, VMGEXIT_PSC_ERROR_INVALID_ENTRY);
3749 return 1;
3750 }
3751
3752 if (entry_start.cur_page) {
3753 /*
3754 * If this is a partially-completed 2M range, force 4K handling
3755 * for the remaining pages since they're effectively split at
3756 * this point. Subsequent code should ensure this doesn't get
3757 * combined with adjacent PSC entries where 2M handling is still
3758 * possible.
3759 */
3760 npages -= entry_start.cur_page;
3761 gfn += entry_start.cur_page;
3762 huge = false;
3763 }
3764
3765 if (npages)
3766 break;
3767 }
3768
3769 if (idx > idx_end) {
3770 /* Nothing more to process. */
3771 snp_complete_psc(svm, 0);
3772 return 1;
3773 }
3774
3775 svm->sev_es.psc_2m = huge;
3776 svm->sev_es.psc_idx = idx;
3777 svm->sev_es.psc_inflight = 1;
3778
3779 /*
3780 * Find all subsequent PSC entries that contain adjacent GPA
3781 * ranges/operations and can be combined into a single
3782 * KVM_HC_MAP_GPA_RANGE exit.
3783 */
3784 while (++idx <= idx_end) {
3785 struct psc_entry entry = entries[idx];
3786
3787 if (entry.operation != entry_start.operation ||
3788 entry.gfn != entry_start.gfn + npages ||
3789 entry.cur_page || !!entry.pagesize != huge)
3790 break;
3791
3792 svm->sev_es.psc_inflight++;
3793 npages += huge ? 512 : 1;
3794 }
3795
3796 switch (entry_start.operation) {
3797 case VMGEXIT_PSC_OP_PRIVATE:
3798 case VMGEXIT_PSC_OP_SHARED:
3799 vcpu->run->exit_reason = KVM_EXIT_HYPERCALL;
3800 vcpu->run->hypercall.nr = KVM_HC_MAP_GPA_RANGE;
3801 /*
3802 * In principle this should have been -KVM_ENOSYS, but userspace (QEMU <=9.2)
3803 * assumed that vcpu->run->hypercall.ret is never changed by KVM and thus that
3804 * it was always zero on KVM_EXIT_HYPERCALL. Since KVM is now overwriting
3805 * vcpu->run->hypercall.ret, ensuring that it is zero to not break QEMU.
3806 */
3807 vcpu->run->hypercall.ret = 0;
3808 vcpu->run->hypercall.args[0] = gfn_to_gpa(gfn);
3809 vcpu->run->hypercall.args[1] = npages;
3810 vcpu->run->hypercall.args[2] = entry_start.operation == VMGEXIT_PSC_OP_PRIVATE
3811 ? KVM_MAP_GPA_RANGE_ENCRYPTED
3812 : KVM_MAP_GPA_RANGE_DECRYPTED;
3813 vcpu->run->hypercall.args[2] |= entry_start.pagesize
3814 ? KVM_MAP_GPA_RANGE_PAGE_SZ_2M
3815 : KVM_MAP_GPA_RANGE_PAGE_SZ_4K;
3816 vcpu->arch.complete_userspace_io = snp_complete_one_psc;
3817 return 0; /* forward request to userspace */
3818 default:
3819 /*
3820 * Only shared/private PSC operations are currently supported, so if the
3821 * entire range consists of unsupported operations (e.g. SMASH/UNSMASH),
3822 * then consider the entire range completed and avoid exiting to
3823 * userspace. In theory snp_complete_psc() can always be called directly
3824 * at this point to complete the current range and start the next one,
3825 * but that could lead to unexpected levels of recursion.
3826 */
3827 __snp_complete_one_psc(svm);
3828 goto next_range;
3829 }
3830
3831 BUG();
3832 }
3833
3834 /*
3835 * Invoked as part of svm_vcpu_reset() processing of an init event.
3836 */
sev_snp_init_protected_guest_state(struct kvm_vcpu * vcpu)3837 void sev_snp_init_protected_guest_state(struct kvm_vcpu *vcpu)
3838 {
3839 struct vcpu_svm *svm = to_svm(vcpu);
3840 struct kvm_memory_slot *slot;
3841 struct page *page;
3842 kvm_pfn_t pfn;
3843 gfn_t gfn;
3844
3845 if (!sev_snp_guest(vcpu->kvm))
3846 return;
3847
3848 guard(mutex)(&svm->sev_es.snp_vmsa_mutex);
3849
3850 if (!svm->sev_es.snp_ap_waiting_for_reset)
3851 return;
3852
3853 svm->sev_es.snp_ap_waiting_for_reset = false;
3854
3855 /* Mark the vCPU as offline and not runnable */
3856 vcpu->arch.pv.pv_unhalted = false;
3857 kvm_set_mp_state(vcpu, KVM_MP_STATE_HALTED);
3858
3859 /* Clear use of the VMSA */
3860 svm->vmcb->control.vmsa_pa = INVALID_PAGE;
3861
3862 /*
3863 * When replacing the VMSA during SEV-SNP AP creation,
3864 * mark the VMCB dirty so that full state is always reloaded.
3865 */
3866 vmcb_mark_all_dirty(svm->vmcb);
3867
3868 if (!VALID_PAGE(svm->sev_es.snp_vmsa_gpa))
3869 return;
3870
3871 gfn = gpa_to_gfn(svm->sev_es.snp_vmsa_gpa);
3872 svm->sev_es.snp_vmsa_gpa = INVALID_PAGE;
3873
3874 slot = gfn_to_memslot(vcpu->kvm, gfn);
3875 if (!slot)
3876 return;
3877
3878 /*
3879 * The new VMSA will be private memory guest memory, so retrieve the
3880 * PFN from the gmem backend.
3881 */
3882 if (kvm_gmem_get_pfn(vcpu->kvm, slot, gfn, &pfn, &page, NULL))
3883 return;
3884
3885 /*
3886 * From this point forward, the VMSA will always be a guest-mapped page
3887 * rather than the initial one allocated by KVM in svm->sev_es.vmsa. In
3888 * theory, svm->sev_es.vmsa could be free'd and cleaned up here, but
3889 * that involves cleanups like wbinvd_on_all_cpus() which would ideally
3890 * be handled during teardown rather than guest boot. Deferring that
3891 * also allows the existing logic for SEV-ES VMSAs to be re-used with
3892 * minimal SNP-specific changes.
3893 */
3894 svm->sev_es.snp_has_guest_vmsa = true;
3895
3896 /* Use the new VMSA */
3897 svm->vmcb->control.vmsa_pa = pfn_to_hpa(pfn);
3898
3899 /* Mark the vCPU as runnable */
3900 kvm_set_mp_state(vcpu, KVM_MP_STATE_RUNNABLE);
3901
3902 /*
3903 * gmem pages aren't currently migratable, but if this ever changes
3904 * then care should be taken to ensure svm->sev_es.vmsa is pinned
3905 * through some other means.
3906 */
3907 kvm_release_page_clean(page);
3908 }
3909
sev_snp_ap_creation(struct vcpu_svm * svm)3910 static int sev_snp_ap_creation(struct vcpu_svm *svm)
3911 {
3912 struct kvm_sev_info *sev = to_kvm_sev_info(svm->vcpu.kvm);
3913 struct kvm_vcpu *vcpu = &svm->vcpu;
3914 struct kvm_vcpu *target_vcpu;
3915 struct vcpu_svm *target_svm;
3916 unsigned int request;
3917 unsigned int apic_id;
3918
3919 request = lower_32_bits(svm->vmcb->control.exit_info_1);
3920 apic_id = upper_32_bits(svm->vmcb->control.exit_info_1);
3921
3922 /* Validate the APIC ID */
3923 target_vcpu = kvm_get_vcpu_by_id(vcpu->kvm, apic_id);
3924 if (!target_vcpu) {
3925 vcpu_unimpl(vcpu, "vmgexit: invalid AP APIC ID [%#x] from guest\n",
3926 apic_id);
3927 return -EINVAL;
3928 }
3929
3930 target_svm = to_svm(target_vcpu);
3931
3932 guard(mutex)(&target_svm->sev_es.snp_vmsa_mutex);
3933
3934 switch (request) {
3935 case SVM_VMGEXIT_AP_CREATE_ON_INIT:
3936 case SVM_VMGEXIT_AP_CREATE:
3937 if (vcpu->arch.regs[VCPU_REGS_RAX] != sev->vmsa_features) {
3938 vcpu_unimpl(vcpu, "vmgexit: mismatched AP sev_features [%#lx] != [%#llx] from guest\n",
3939 vcpu->arch.regs[VCPU_REGS_RAX], sev->vmsa_features);
3940 return -EINVAL;
3941 }
3942
3943 if (!page_address_valid(vcpu, svm->vmcb->control.exit_info_2)) {
3944 vcpu_unimpl(vcpu, "vmgexit: invalid AP VMSA address [%#llx] from guest\n",
3945 svm->vmcb->control.exit_info_2);
3946 return -EINVAL;
3947 }
3948
3949 /*
3950 * Malicious guest can RMPADJUST a large page into VMSA which
3951 * will hit the SNP erratum where the CPU will incorrectly signal
3952 * an RMP violation #PF if a hugepage collides with the RMP entry
3953 * of VMSA page, reject the AP CREATE request if VMSA address from
3954 * guest is 2M aligned.
3955 */
3956 if (IS_ALIGNED(svm->vmcb->control.exit_info_2, PMD_SIZE)) {
3957 vcpu_unimpl(vcpu,
3958 "vmgexit: AP VMSA address [%llx] from guest is unsafe as it is 2M aligned\n",
3959 svm->vmcb->control.exit_info_2);
3960 return -EINVAL;
3961 }
3962
3963 target_svm->sev_es.snp_vmsa_gpa = svm->vmcb->control.exit_info_2;
3964 break;
3965 case SVM_VMGEXIT_AP_DESTROY:
3966 target_svm->sev_es.snp_vmsa_gpa = INVALID_PAGE;
3967 break;
3968 default:
3969 vcpu_unimpl(vcpu, "vmgexit: invalid AP creation request [%#x] from guest\n",
3970 request);
3971 return -EINVAL;
3972 }
3973
3974 target_svm->sev_es.snp_ap_waiting_for_reset = true;
3975
3976 /*
3977 * Unless Creation is deferred until INIT, signal the vCPU to update
3978 * its state.
3979 */
3980 if (request != SVM_VMGEXIT_AP_CREATE_ON_INIT)
3981 kvm_make_request_and_kick(KVM_REQ_UPDATE_PROTECTED_GUEST_STATE, target_vcpu);
3982
3983 return 0;
3984 }
3985
snp_handle_guest_req(struct vcpu_svm * svm,gpa_t req_gpa,gpa_t resp_gpa)3986 static int snp_handle_guest_req(struct vcpu_svm *svm, gpa_t req_gpa, gpa_t resp_gpa)
3987 {
3988 struct sev_data_snp_guest_request data = {0};
3989 struct kvm *kvm = svm->vcpu.kvm;
3990 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
3991 sev_ret_code fw_err = 0;
3992 int ret;
3993
3994 if (!sev_snp_guest(kvm))
3995 return -EINVAL;
3996
3997 mutex_lock(&sev->guest_req_mutex);
3998
3999 if (kvm_read_guest(kvm, req_gpa, sev->guest_req_buf, PAGE_SIZE)) {
4000 ret = -EIO;
4001 goto out_unlock;
4002 }
4003
4004 data.gctx_paddr = __psp_pa(sev->snp_context);
4005 data.req_paddr = __psp_pa(sev->guest_req_buf);
4006 data.res_paddr = __psp_pa(sev->guest_resp_buf);
4007
4008 /*
4009 * Firmware failures are propagated on to guest, but any other failure
4010 * condition along the way should be reported to userspace. E.g. if
4011 * the PSP is dead and commands are timing out.
4012 */
4013 ret = sev_issue_cmd(kvm, SEV_CMD_SNP_GUEST_REQUEST, &data, &fw_err);
4014 if (ret && !fw_err)
4015 goto out_unlock;
4016
4017 if (kvm_write_guest(kvm, resp_gpa, sev->guest_resp_buf, PAGE_SIZE)) {
4018 ret = -EIO;
4019 goto out_unlock;
4020 }
4021
4022 /* No action is requested *from KVM* if there was a firmware error. */
4023 svm_vmgexit_no_action(svm, SNP_GUEST_ERR(0, fw_err));
4024
4025 ret = 1; /* resume guest */
4026
4027 out_unlock:
4028 mutex_unlock(&sev->guest_req_mutex);
4029 return ret;
4030 }
4031
snp_handle_ext_guest_req(struct vcpu_svm * svm,gpa_t req_gpa,gpa_t resp_gpa)4032 static int snp_handle_ext_guest_req(struct vcpu_svm *svm, gpa_t req_gpa, gpa_t resp_gpa)
4033 {
4034 struct kvm *kvm = svm->vcpu.kvm;
4035 u8 msg_type;
4036
4037 if (!sev_snp_guest(kvm))
4038 return -EINVAL;
4039
4040 if (kvm_read_guest(kvm, req_gpa + offsetof(struct snp_guest_msg_hdr, msg_type),
4041 &msg_type, 1))
4042 return -EIO;
4043
4044 /*
4045 * As per GHCB spec, requests of type MSG_REPORT_REQ also allow for
4046 * additional certificate data to be provided alongside the attestation
4047 * report via the guest-provided data pages indicated by RAX/RBX. The
4048 * certificate data is optional and requires additional KVM enablement
4049 * to provide an interface for userspace to provide it, but KVM still
4050 * needs to be able to handle extended guest requests either way. So
4051 * provide a stub implementation that will always return an empty
4052 * certificate table in the guest-provided data pages.
4053 */
4054 if (msg_type == SNP_MSG_REPORT_REQ) {
4055 struct kvm_vcpu *vcpu = &svm->vcpu;
4056 u64 data_npages;
4057 gpa_t data_gpa;
4058
4059 if (!kvm_ghcb_rax_is_valid(svm) || !kvm_ghcb_rbx_is_valid(svm))
4060 goto request_invalid;
4061
4062 data_gpa = vcpu->arch.regs[VCPU_REGS_RAX];
4063 data_npages = vcpu->arch.regs[VCPU_REGS_RBX];
4064
4065 if (!PAGE_ALIGNED(data_gpa))
4066 goto request_invalid;
4067
4068 /*
4069 * As per GHCB spec (see "SNP Extended Guest Request"), the
4070 * certificate table is terminated by 24-bytes of zeroes.
4071 */
4072 if (data_npages && kvm_clear_guest(kvm, data_gpa, 24))
4073 return -EIO;
4074 }
4075
4076 return snp_handle_guest_req(svm, req_gpa, resp_gpa);
4077
4078 request_invalid:
4079 svm_vmgexit_bad_input(svm, GHCB_ERR_INVALID_INPUT);
4080 return 1; /* resume guest */
4081 }
4082
sev_handle_vmgexit_msr_protocol(struct vcpu_svm * svm)4083 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
4084 {
4085 struct vmcb_control_area *control = &svm->vmcb->control;
4086 struct kvm_vcpu *vcpu = &svm->vcpu;
4087 struct kvm_sev_info *sev = to_kvm_sev_info(vcpu->kvm);
4088 u64 ghcb_info;
4089 int ret = 1;
4090
4091 ghcb_info = control->ghcb_gpa & GHCB_MSR_INFO_MASK;
4092
4093 trace_kvm_vmgexit_msr_protocol_enter(svm->vcpu.vcpu_id,
4094 control->ghcb_gpa);
4095
4096 switch (ghcb_info) {
4097 case GHCB_MSR_SEV_INFO_REQ:
4098 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO((__u64)sev->ghcb_version,
4099 GHCB_VERSION_MIN,
4100 sev_enc_bit));
4101 break;
4102 case GHCB_MSR_CPUID_REQ: {
4103 u64 cpuid_fn, cpuid_reg, cpuid_value;
4104
4105 cpuid_fn = get_ghcb_msr_bits(svm,
4106 GHCB_MSR_CPUID_FUNC_MASK,
4107 GHCB_MSR_CPUID_FUNC_POS);
4108
4109 /* Initialize the registers needed by the CPUID intercept */
4110 vcpu->arch.regs[VCPU_REGS_RAX] = cpuid_fn;
4111 vcpu->arch.regs[VCPU_REGS_RCX] = 0;
4112
4113 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_CPUID);
4114 if (!ret) {
4115 /* Error, keep GHCB MSR value as-is */
4116 break;
4117 }
4118
4119 cpuid_reg = get_ghcb_msr_bits(svm,
4120 GHCB_MSR_CPUID_REG_MASK,
4121 GHCB_MSR_CPUID_REG_POS);
4122 if (cpuid_reg == 0)
4123 cpuid_value = vcpu->arch.regs[VCPU_REGS_RAX];
4124 else if (cpuid_reg == 1)
4125 cpuid_value = vcpu->arch.regs[VCPU_REGS_RBX];
4126 else if (cpuid_reg == 2)
4127 cpuid_value = vcpu->arch.regs[VCPU_REGS_RCX];
4128 else
4129 cpuid_value = vcpu->arch.regs[VCPU_REGS_RDX];
4130
4131 set_ghcb_msr_bits(svm, cpuid_value,
4132 GHCB_MSR_CPUID_VALUE_MASK,
4133 GHCB_MSR_CPUID_VALUE_POS);
4134
4135 set_ghcb_msr_bits(svm, GHCB_MSR_CPUID_RESP,
4136 GHCB_MSR_INFO_MASK,
4137 GHCB_MSR_INFO_POS);
4138 break;
4139 }
4140 case GHCB_MSR_AP_RESET_HOLD_REQ:
4141 svm->sev_es.ap_reset_hold_type = AP_RESET_HOLD_MSR_PROTO;
4142 ret = kvm_emulate_ap_reset_hold(&svm->vcpu);
4143
4144 /*
4145 * Preset the result to a non-SIPI return and then only set
4146 * the result to non-zero when delivering a SIPI.
4147 */
4148 set_ghcb_msr_bits(svm, 0,
4149 GHCB_MSR_AP_RESET_HOLD_RESULT_MASK,
4150 GHCB_MSR_AP_RESET_HOLD_RESULT_POS);
4151
4152 set_ghcb_msr_bits(svm, GHCB_MSR_AP_RESET_HOLD_RESP,
4153 GHCB_MSR_INFO_MASK,
4154 GHCB_MSR_INFO_POS);
4155 break;
4156 case GHCB_MSR_HV_FT_REQ:
4157 set_ghcb_msr_bits(svm, GHCB_HV_FT_SUPPORTED,
4158 GHCB_MSR_HV_FT_MASK, GHCB_MSR_HV_FT_POS);
4159 set_ghcb_msr_bits(svm, GHCB_MSR_HV_FT_RESP,
4160 GHCB_MSR_INFO_MASK, GHCB_MSR_INFO_POS);
4161 break;
4162 case GHCB_MSR_PREF_GPA_REQ:
4163 if (!sev_snp_guest(vcpu->kvm))
4164 goto out_terminate;
4165
4166 set_ghcb_msr_bits(svm, GHCB_MSR_PREF_GPA_NONE, GHCB_MSR_GPA_VALUE_MASK,
4167 GHCB_MSR_GPA_VALUE_POS);
4168 set_ghcb_msr_bits(svm, GHCB_MSR_PREF_GPA_RESP, GHCB_MSR_INFO_MASK,
4169 GHCB_MSR_INFO_POS);
4170 break;
4171 case GHCB_MSR_REG_GPA_REQ: {
4172 u64 gfn;
4173
4174 if (!sev_snp_guest(vcpu->kvm))
4175 goto out_terminate;
4176
4177 gfn = get_ghcb_msr_bits(svm, GHCB_MSR_GPA_VALUE_MASK,
4178 GHCB_MSR_GPA_VALUE_POS);
4179
4180 svm->sev_es.ghcb_registered_gpa = gfn_to_gpa(gfn);
4181
4182 set_ghcb_msr_bits(svm, gfn, GHCB_MSR_GPA_VALUE_MASK,
4183 GHCB_MSR_GPA_VALUE_POS);
4184 set_ghcb_msr_bits(svm, GHCB_MSR_REG_GPA_RESP, GHCB_MSR_INFO_MASK,
4185 GHCB_MSR_INFO_POS);
4186 break;
4187 }
4188 case GHCB_MSR_PSC_REQ:
4189 if (!sev_snp_guest(vcpu->kvm))
4190 goto out_terminate;
4191
4192 ret = snp_begin_psc_msr(svm, control->ghcb_gpa);
4193 break;
4194 case GHCB_MSR_TERM_REQ: {
4195 u64 reason_set, reason_code;
4196
4197 reason_set = get_ghcb_msr_bits(svm,
4198 GHCB_MSR_TERM_REASON_SET_MASK,
4199 GHCB_MSR_TERM_REASON_SET_POS);
4200 reason_code = get_ghcb_msr_bits(svm,
4201 GHCB_MSR_TERM_REASON_MASK,
4202 GHCB_MSR_TERM_REASON_POS);
4203 pr_info("SEV-ES guest requested termination: %#llx:%#llx\n",
4204 reason_set, reason_code);
4205
4206 goto out_terminate;
4207 }
4208 default:
4209 /* Error, keep GHCB MSR value as-is */
4210 break;
4211 }
4212
4213 trace_kvm_vmgexit_msr_protocol_exit(svm->vcpu.vcpu_id,
4214 control->ghcb_gpa, ret);
4215
4216 return ret;
4217
4218 out_terminate:
4219 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
4220 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_SEV_TERM;
4221 vcpu->run->system_event.ndata = 1;
4222 vcpu->run->system_event.data[0] = control->ghcb_gpa;
4223
4224 return 0;
4225 }
4226
sev_handle_vmgexit(struct kvm_vcpu * vcpu)4227 int sev_handle_vmgexit(struct kvm_vcpu *vcpu)
4228 {
4229 struct vcpu_svm *svm = to_svm(vcpu);
4230 struct vmcb_control_area *control = &svm->vmcb->control;
4231 u64 ghcb_gpa, exit_code;
4232 int ret;
4233
4234 /* Validate the GHCB */
4235 ghcb_gpa = control->ghcb_gpa;
4236 if (ghcb_gpa & GHCB_MSR_INFO_MASK)
4237 return sev_handle_vmgexit_msr_protocol(svm);
4238
4239 if (!ghcb_gpa) {
4240 vcpu_unimpl(vcpu, "vmgexit: GHCB gpa is not set\n");
4241
4242 /* Without a GHCB, just return right back to the guest */
4243 return 1;
4244 }
4245
4246 if (kvm_vcpu_map(vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->sev_es.ghcb_map)) {
4247 /* Unable to map GHCB from guest */
4248 vcpu_unimpl(vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
4249 ghcb_gpa);
4250
4251 /* Without a GHCB, just return right back to the guest */
4252 return 1;
4253 }
4254
4255 svm->sev_es.ghcb = svm->sev_es.ghcb_map.hva;
4256
4257 trace_kvm_vmgexit_enter(vcpu->vcpu_id, svm->sev_es.ghcb);
4258
4259 sev_es_sync_from_ghcb(svm);
4260
4261 /* SEV-SNP guest requires that the GHCB GPA must be registered */
4262 if (sev_snp_guest(svm->vcpu.kvm) && !ghcb_gpa_is_registered(svm, ghcb_gpa)) {
4263 vcpu_unimpl(&svm->vcpu, "vmgexit: GHCB GPA [%#llx] is not registered.\n", ghcb_gpa);
4264 return -EINVAL;
4265 }
4266
4267 ret = sev_es_validate_vmgexit(svm);
4268 if (ret)
4269 return ret;
4270
4271 svm_vmgexit_success(svm, 0);
4272
4273 exit_code = kvm_ghcb_get_sw_exit_code(control);
4274 switch (exit_code) {
4275 case SVM_VMGEXIT_MMIO_READ:
4276 ret = setup_vmgexit_scratch(svm, true, control->exit_info_2);
4277 if (ret)
4278 break;
4279
4280 ret = kvm_sev_es_mmio_read(vcpu,
4281 control->exit_info_1,
4282 control->exit_info_2,
4283 svm->sev_es.ghcb_sa);
4284 break;
4285 case SVM_VMGEXIT_MMIO_WRITE:
4286 ret = setup_vmgexit_scratch(svm, false, control->exit_info_2);
4287 if (ret)
4288 break;
4289
4290 ret = kvm_sev_es_mmio_write(vcpu,
4291 control->exit_info_1,
4292 control->exit_info_2,
4293 svm->sev_es.ghcb_sa);
4294 break;
4295 case SVM_VMGEXIT_NMI_COMPLETE:
4296 ++vcpu->stat.nmi_window_exits;
4297 svm->nmi_masked = false;
4298 kvm_make_request(KVM_REQ_EVENT, vcpu);
4299 ret = 1;
4300 break;
4301 case SVM_VMGEXIT_AP_HLT_LOOP:
4302 svm->sev_es.ap_reset_hold_type = AP_RESET_HOLD_NAE_EVENT;
4303 ret = kvm_emulate_ap_reset_hold(vcpu);
4304 break;
4305 case SVM_VMGEXIT_AP_JUMP_TABLE: {
4306 struct kvm_sev_info *sev = to_kvm_sev_info(vcpu->kvm);
4307
4308 switch (control->exit_info_1) {
4309 case 0:
4310 /* Set AP jump table address */
4311 sev->ap_jump_table = control->exit_info_2;
4312 break;
4313 case 1:
4314 /* Get AP jump table address */
4315 svm_vmgexit_success(svm, sev->ap_jump_table);
4316 break;
4317 default:
4318 pr_err("svm: vmgexit: unsupported AP jump table request - exit_info_1=%#llx\n",
4319 control->exit_info_1);
4320 svm_vmgexit_bad_input(svm, GHCB_ERR_INVALID_INPUT);
4321 }
4322
4323 ret = 1;
4324 break;
4325 }
4326 case SVM_VMGEXIT_HV_FEATURES:
4327 svm_vmgexit_success(svm, GHCB_HV_FT_SUPPORTED);
4328 ret = 1;
4329 break;
4330 case SVM_VMGEXIT_TERM_REQUEST:
4331 pr_info("SEV-ES guest requested termination: reason %#llx info %#llx\n",
4332 control->exit_info_1, control->exit_info_2);
4333 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
4334 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_SEV_TERM;
4335 vcpu->run->system_event.ndata = 1;
4336 vcpu->run->system_event.data[0] = control->ghcb_gpa;
4337 break;
4338 case SVM_VMGEXIT_PSC:
4339 ret = setup_vmgexit_scratch(svm, true, control->exit_info_2);
4340 if (ret)
4341 break;
4342
4343 ret = snp_begin_psc(svm, svm->sev_es.ghcb_sa);
4344 break;
4345 case SVM_VMGEXIT_AP_CREATION:
4346 ret = sev_snp_ap_creation(svm);
4347 if (ret) {
4348 svm_vmgexit_bad_input(svm, GHCB_ERR_INVALID_INPUT);
4349 }
4350
4351 ret = 1;
4352 break;
4353 case SVM_VMGEXIT_GUEST_REQUEST:
4354 ret = snp_handle_guest_req(svm, control->exit_info_1, control->exit_info_2);
4355 break;
4356 case SVM_VMGEXIT_EXT_GUEST_REQUEST:
4357 ret = snp_handle_ext_guest_req(svm, control->exit_info_1, control->exit_info_2);
4358 break;
4359 case SVM_VMGEXIT_UNSUPPORTED_EVENT:
4360 vcpu_unimpl(vcpu,
4361 "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
4362 control->exit_info_1, control->exit_info_2);
4363 ret = -EINVAL;
4364 break;
4365 default:
4366 ret = svm_invoke_exit_handler(vcpu, exit_code);
4367 }
4368
4369 return ret;
4370 }
4371
sev_es_string_io(struct vcpu_svm * svm,int size,unsigned int port,int in)4372 int sev_es_string_io(struct vcpu_svm *svm, int size, unsigned int port, int in)
4373 {
4374 int count;
4375 int bytes;
4376 int r;
4377
4378 if (svm->vmcb->control.exit_info_2 > INT_MAX)
4379 return -EINVAL;
4380
4381 count = svm->vmcb->control.exit_info_2;
4382 if (unlikely(check_mul_overflow(count, size, &bytes)))
4383 return -EINVAL;
4384
4385 r = setup_vmgexit_scratch(svm, in, bytes);
4386 if (r)
4387 return r;
4388
4389 return kvm_sev_es_string_io(&svm->vcpu, size, port, svm->sev_es.ghcb_sa,
4390 count, in);
4391 }
4392
sev_es_vcpu_after_set_cpuid(struct vcpu_svm * svm)4393 static void sev_es_vcpu_after_set_cpuid(struct vcpu_svm *svm)
4394 {
4395 struct kvm_vcpu *vcpu = &svm->vcpu;
4396
4397 if (boot_cpu_has(X86_FEATURE_V_TSC_AUX)) {
4398 bool v_tsc_aux = guest_cpu_cap_has(vcpu, X86_FEATURE_RDTSCP) ||
4399 guest_cpu_cap_has(vcpu, X86_FEATURE_RDPID);
4400
4401 set_msr_interception(vcpu, svm->msrpm, MSR_TSC_AUX, v_tsc_aux, v_tsc_aux);
4402 }
4403
4404 /*
4405 * For SEV-ES, accesses to MSR_IA32_XSS should not be intercepted if
4406 * the host/guest supports its use.
4407 *
4408 * KVM treats the guest as being capable of using XSAVES even if XSAVES
4409 * isn't enabled in guest CPUID as there is no intercept for XSAVES,
4410 * i.e. the guest can use XSAVES/XRSTOR to read/write XSS if XSAVE is
4411 * exposed to the guest and XSAVES is supported in hardware. Condition
4412 * full XSS passthrough on the guest being able to use XSAVES *and*
4413 * XSAVES being exposed to the guest so that KVM can at least honor
4414 * guest CPUID for RDMSR and WRMSR.
4415 */
4416 if (guest_cpu_cap_has(vcpu, X86_FEATURE_XSAVES) &&
4417 guest_cpuid_has(vcpu, X86_FEATURE_XSAVES))
4418 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_XSS, 1, 1);
4419 else
4420 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_XSS, 0, 0);
4421 }
4422
sev_vcpu_after_set_cpuid(struct vcpu_svm * svm)4423 void sev_vcpu_after_set_cpuid(struct vcpu_svm *svm)
4424 {
4425 struct kvm_vcpu *vcpu = &svm->vcpu;
4426 struct kvm_cpuid_entry2 *best;
4427
4428 /* For sev guests, the memory encryption bit is not reserved in CR3. */
4429 best = kvm_find_cpuid_entry(vcpu, 0x8000001F);
4430 if (best)
4431 vcpu->arch.reserved_gpa_bits &= ~(1UL << (best->ebx & 0x3f));
4432
4433 if (sev_es_guest(svm->vcpu.kvm))
4434 sev_es_vcpu_after_set_cpuid(svm);
4435 }
4436
sev_es_init_vmcb(struct vcpu_svm * svm)4437 static void sev_es_init_vmcb(struct vcpu_svm *svm)
4438 {
4439 struct kvm_sev_info *sev = to_kvm_sev_info(svm->vcpu.kvm);
4440 struct vmcb *vmcb = svm->vmcb01.ptr;
4441 struct kvm_vcpu *vcpu = &svm->vcpu;
4442
4443 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ES_ENABLE;
4444
4445 /*
4446 * An SEV-ES guest requires a VMSA area that is a separate from the
4447 * VMCB page. Do not include the encryption mask on the VMSA physical
4448 * address since hardware will access it using the guest key. Note,
4449 * the VMSA will be NULL if this vCPU is the destination for intrahost
4450 * migration, and will be copied later.
4451 */
4452 if (!svm->sev_es.snp_has_guest_vmsa) {
4453 if (svm->sev_es.vmsa)
4454 svm->vmcb->control.vmsa_pa = __pa(svm->sev_es.vmsa);
4455 else
4456 svm->vmcb->control.vmsa_pa = INVALID_PAGE;
4457 }
4458
4459 if (cpu_feature_enabled(X86_FEATURE_ALLOWED_SEV_FEATURES))
4460 svm->vmcb->control.allowed_sev_features = sev->vmsa_features |
4461 VMCB_ALLOWED_SEV_FEATURES_VALID;
4462
4463 /* Can't intercept CR register access, HV can't modify CR registers */
4464 svm_clr_intercept(svm, INTERCEPT_CR0_READ);
4465 svm_clr_intercept(svm, INTERCEPT_CR4_READ);
4466 svm_clr_intercept(svm, INTERCEPT_CR8_READ);
4467 svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
4468 svm_clr_intercept(svm, INTERCEPT_CR4_WRITE);
4469 svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
4470
4471 svm_clr_intercept(svm, INTERCEPT_SELECTIVE_CR0);
4472
4473 /* Track EFER/CR register changes */
4474 svm_set_intercept(svm, TRAP_EFER_WRITE);
4475 svm_set_intercept(svm, TRAP_CR0_WRITE);
4476 svm_set_intercept(svm, TRAP_CR4_WRITE);
4477 svm_set_intercept(svm, TRAP_CR8_WRITE);
4478
4479 vmcb->control.intercepts[INTERCEPT_DR] = 0;
4480 if (!sev_vcpu_has_debug_swap(svm)) {
4481 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_READ);
4482 vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_WRITE);
4483 recalc_intercepts(svm);
4484 } else {
4485 /*
4486 * Disable #DB intercept iff DebugSwap is enabled. KVM doesn't
4487 * allow debugging SEV-ES guests, and enables DebugSwap iff
4488 * NO_NESTED_DATA_BP is supported, so there's no reason to
4489 * intercept #DB when DebugSwap is enabled. For simplicity
4490 * with respect to guest debug, intercept #DB for other VMs
4491 * even if NO_NESTED_DATA_BP is supported, i.e. even if the
4492 * guest can't DoS the CPU with infinite #DB vectoring.
4493 */
4494 clr_exception_intercept(svm, DB_VECTOR);
4495 }
4496
4497 /* Can't intercept XSETBV, HV can't modify XCR0 directly */
4498 svm_clr_intercept(svm, INTERCEPT_XSETBV);
4499
4500 /* Clear intercepts on selected MSRs */
4501 set_msr_interception(vcpu, svm->msrpm, MSR_EFER, 1, 1);
4502 set_msr_interception(vcpu, svm->msrpm, MSR_IA32_CR_PAT, 1, 1);
4503 }
4504
sev_init_vmcb(struct vcpu_svm * svm)4505 void sev_init_vmcb(struct vcpu_svm *svm)
4506 {
4507 svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
4508 clr_exception_intercept(svm, UD_VECTOR);
4509
4510 /*
4511 * Don't intercept #GP for SEV guests, e.g. for the VMware backdoor, as
4512 * KVM can't decrypt guest memory to decode the faulting instruction.
4513 */
4514 clr_exception_intercept(svm, GP_VECTOR);
4515
4516 if (sev_es_guest(svm->vcpu.kvm))
4517 sev_es_init_vmcb(svm);
4518 }
4519
sev_es_vcpu_reset(struct vcpu_svm * svm)4520 void sev_es_vcpu_reset(struct vcpu_svm *svm)
4521 {
4522 struct kvm_vcpu *vcpu = &svm->vcpu;
4523 struct kvm_sev_info *sev = to_kvm_sev_info(vcpu->kvm);
4524
4525 /*
4526 * Set the GHCB MSR value as per the GHCB specification when emulating
4527 * vCPU RESET for an SEV-ES guest.
4528 */
4529 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO((__u64)sev->ghcb_version,
4530 GHCB_VERSION_MIN,
4531 sev_enc_bit));
4532
4533 mutex_init(&svm->sev_es.snp_vmsa_mutex);
4534 }
4535
sev_es_prepare_switch_to_guest(struct vcpu_svm * svm,struct sev_es_save_area * hostsa)4536 void sev_es_prepare_switch_to_guest(struct vcpu_svm *svm, struct sev_es_save_area *hostsa)
4537 {
4538 struct kvm *kvm = svm->vcpu.kvm;
4539
4540 /*
4541 * All host state for SEV-ES guests is categorized into three swap types
4542 * based on how it is handled by hardware during a world switch:
4543 *
4544 * A: VMRUN: Host state saved in host save area
4545 * VMEXIT: Host state loaded from host save area
4546 *
4547 * B: VMRUN: Host state _NOT_ saved in host save area
4548 * VMEXIT: Host state loaded from host save area
4549 *
4550 * C: VMRUN: Host state _NOT_ saved in host save area
4551 * VMEXIT: Host state initialized to default(reset) values
4552 *
4553 * Manually save type-B state, i.e. state that is loaded by VMEXIT but
4554 * isn't saved by VMRUN, that isn't already saved by VMSAVE (performed
4555 * by common SVM code).
4556 */
4557 hostsa->xcr0 = kvm_host.xcr0;
4558 hostsa->pkru = read_pkru();
4559 hostsa->xss = kvm_host.xss;
4560
4561 /*
4562 * If DebugSwap is enabled, debug registers are loaded but NOT saved by
4563 * the CPU (Type-B). If DebugSwap is disabled/unsupported, the CPU does
4564 * not save or load debug registers. Sadly, KVM can't prevent SNP
4565 * guests from lying about DebugSwap on secondary vCPUs, i.e. the
4566 * SEV_FEATURES provided at "AP Create" isn't guaranteed to match what
4567 * the guest has actually enabled (or not!) in the VMSA.
4568 *
4569 * If DebugSwap is *possible*, save the masks so that they're restored
4570 * if the guest enables DebugSwap. But for the DRs themselves, do NOT
4571 * rely on the CPU to restore the host values; KVM will restore them as
4572 * needed in common code, via hw_breakpoint_restore(). Note, KVM does
4573 * NOT support virtualizing Breakpoint Extensions, i.e. the mask MSRs
4574 * don't need to be restored per se, KVM just needs to ensure they are
4575 * loaded with the correct values *if* the CPU writes the MSRs.
4576 */
4577 if (sev_vcpu_has_debug_swap(svm) ||
4578 (sev_snp_guest(kvm) && cpu_feature_enabled(X86_FEATURE_DEBUG_SWAP))) {
4579 hostsa->dr0_addr_mask = amd_get_dr_addr_mask(0);
4580 hostsa->dr1_addr_mask = amd_get_dr_addr_mask(1);
4581 hostsa->dr2_addr_mask = amd_get_dr_addr_mask(2);
4582 hostsa->dr3_addr_mask = amd_get_dr_addr_mask(3);
4583 }
4584 }
4585
sev_vcpu_deliver_sipi_vector(struct kvm_vcpu * vcpu,u8 vector)4586 void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
4587 {
4588 struct vcpu_svm *svm = to_svm(vcpu);
4589
4590 /* First SIPI: Use the values as initially set by the VMM */
4591 if (!svm->sev_es.received_first_sipi) {
4592 svm->sev_es.received_first_sipi = true;
4593 return;
4594 }
4595
4596 /* Subsequent SIPI */
4597 switch (svm->sev_es.ap_reset_hold_type) {
4598 case AP_RESET_HOLD_NAE_EVENT:
4599 /*
4600 * Return from an AP Reset Hold VMGEXIT, where the guest will
4601 * set the CS and RIP. Set SW_EXIT_INFO_2 to a non-zero value.
4602 */
4603 svm_vmgexit_success(svm, 1);
4604 break;
4605 case AP_RESET_HOLD_MSR_PROTO:
4606 /*
4607 * Return from an AP Reset Hold VMGEXIT, where the guest will
4608 * set the CS and RIP. Set GHCB data field to a non-zero value.
4609 */
4610 set_ghcb_msr_bits(svm, 1,
4611 GHCB_MSR_AP_RESET_HOLD_RESULT_MASK,
4612 GHCB_MSR_AP_RESET_HOLD_RESULT_POS);
4613
4614 set_ghcb_msr_bits(svm, GHCB_MSR_AP_RESET_HOLD_RESP,
4615 GHCB_MSR_INFO_MASK,
4616 GHCB_MSR_INFO_POS);
4617 break;
4618 default:
4619 break;
4620 }
4621 }
4622
snp_safe_alloc_page_node(int node,gfp_t gfp)4623 struct page *snp_safe_alloc_page_node(int node, gfp_t gfp)
4624 {
4625 unsigned long pfn;
4626 struct page *p;
4627
4628 if (!cc_platform_has(CC_ATTR_HOST_SEV_SNP))
4629 return alloc_pages_node(node, gfp | __GFP_ZERO, 0);
4630
4631 /*
4632 * Allocate an SNP-safe page to workaround the SNP erratum where
4633 * the CPU will incorrectly signal an RMP violation #PF if a
4634 * hugepage (2MB or 1GB) collides with the RMP entry of a
4635 * 2MB-aligned VMCB, VMSA, or AVIC backing page.
4636 *
4637 * Allocate one extra page, choose a page which is not
4638 * 2MB-aligned, and free the other.
4639 */
4640 p = alloc_pages_node(node, gfp | __GFP_ZERO, 1);
4641 if (!p)
4642 return NULL;
4643
4644 split_page(p, 1);
4645
4646 pfn = page_to_pfn(p);
4647 if (IS_ALIGNED(pfn, PTRS_PER_PMD))
4648 __free_page(p++);
4649 else
4650 __free_page(p + 1);
4651
4652 return p;
4653 }
4654
sev_handle_rmp_fault(struct kvm_vcpu * vcpu,gpa_t gpa,u64 error_code)4655 void sev_handle_rmp_fault(struct kvm_vcpu *vcpu, gpa_t gpa, u64 error_code)
4656 {
4657 struct kvm_memory_slot *slot;
4658 struct kvm *kvm = vcpu->kvm;
4659 int order, rmp_level, ret;
4660 struct page *page;
4661 bool assigned;
4662 kvm_pfn_t pfn;
4663 gfn_t gfn;
4664
4665 gfn = gpa >> PAGE_SHIFT;
4666
4667 /*
4668 * The only time RMP faults occur for shared pages is when the guest is
4669 * triggering an RMP fault for an implicit page-state change from
4670 * shared->private. Implicit page-state changes are forwarded to
4671 * userspace via KVM_EXIT_MEMORY_FAULT events, however, so RMP faults
4672 * for shared pages should not end up here.
4673 */
4674 if (!kvm_mem_is_private(kvm, gfn)) {
4675 pr_warn_ratelimited("SEV: Unexpected RMP fault for non-private GPA 0x%llx\n",
4676 gpa);
4677 return;
4678 }
4679
4680 slot = gfn_to_memslot(kvm, gfn);
4681 if (!kvm_slot_can_be_private(slot)) {
4682 pr_warn_ratelimited("SEV: Unexpected RMP fault, non-private slot for GPA 0x%llx\n",
4683 gpa);
4684 return;
4685 }
4686
4687 ret = kvm_gmem_get_pfn(kvm, slot, gfn, &pfn, &page, &order);
4688 if (ret) {
4689 pr_warn_ratelimited("SEV: Unexpected RMP fault, no backing page for private GPA 0x%llx\n",
4690 gpa);
4691 return;
4692 }
4693
4694 ret = snp_lookup_rmpentry(pfn, &assigned, &rmp_level);
4695 if (ret || !assigned) {
4696 pr_warn_ratelimited("SEV: Unexpected RMP fault, no assigned RMP entry found for GPA 0x%llx PFN 0x%llx error %d\n",
4697 gpa, pfn, ret);
4698 goto out_no_trace;
4699 }
4700
4701 /*
4702 * There are 2 cases where a PSMASH may be needed to resolve an #NPF
4703 * with PFERR_GUEST_RMP_BIT set:
4704 *
4705 * 1) RMPADJUST/PVALIDATE can trigger an #NPF with PFERR_GUEST_SIZEM
4706 * bit set if the guest issues them with a smaller granularity than
4707 * what is indicated by the page-size bit in the 2MB RMP entry for
4708 * the PFN that backs the GPA.
4709 *
4710 * 2) Guest access via NPT can trigger an #NPF if the NPT mapping is
4711 * smaller than what is indicated by the 2MB RMP entry for the PFN
4712 * that backs the GPA.
4713 *
4714 * In both these cases, the corresponding 2M RMP entry needs to
4715 * be PSMASH'd to 512 4K RMP entries. If the RMP entry is already
4716 * split into 4K RMP entries, then this is likely a spurious case which
4717 * can occur when there are concurrent accesses by the guest to a 2MB
4718 * GPA range that is backed by a 2MB-aligned PFN who's RMP entry is in
4719 * the process of being PMASH'd into 4K entries. These cases should
4720 * resolve automatically on subsequent accesses, so just ignore them
4721 * here.
4722 */
4723 if (rmp_level == PG_LEVEL_4K)
4724 goto out;
4725
4726 ret = snp_rmptable_psmash(pfn);
4727 if (ret) {
4728 /*
4729 * Look it up again. If it's 4K now then the PSMASH may have
4730 * raced with another process and the issue has already resolved
4731 * itself.
4732 */
4733 if (!snp_lookup_rmpentry(pfn, &assigned, &rmp_level) &&
4734 assigned && rmp_level == PG_LEVEL_4K)
4735 goto out;
4736
4737 pr_warn_ratelimited("SEV: Unable to split RMP entry for GPA 0x%llx PFN 0x%llx ret %d\n",
4738 gpa, pfn, ret);
4739 }
4740
4741 kvm_zap_gfn_range(kvm, gfn, gfn + PTRS_PER_PMD);
4742 out:
4743 trace_kvm_rmp_fault(vcpu, gpa, pfn, error_code, rmp_level, ret);
4744 out_no_trace:
4745 kvm_release_page_unused(page);
4746 }
4747
is_pfn_range_shared(kvm_pfn_t start,kvm_pfn_t end)4748 static bool is_pfn_range_shared(kvm_pfn_t start, kvm_pfn_t end)
4749 {
4750 kvm_pfn_t pfn = start;
4751
4752 while (pfn < end) {
4753 int ret, rmp_level;
4754 bool assigned;
4755
4756 ret = snp_lookup_rmpentry(pfn, &assigned, &rmp_level);
4757 if (ret) {
4758 pr_warn_ratelimited("SEV: Failed to retrieve RMP entry: PFN 0x%llx GFN start 0x%llx GFN end 0x%llx RMP level %d error %d\n",
4759 pfn, start, end, rmp_level, ret);
4760 return false;
4761 }
4762
4763 if (assigned) {
4764 pr_debug("%s: overlap detected, PFN 0x%llx start 0x%llx end 0x%llx RMP level %d\n",
4765 __func__, pfn, start, end, rmp_level);
4766 return false;
4767 }
4768
4769 pfn++;
4770 }
4771
4772 return true;
4773 }
4774
max_level_for_order(int order)4775 static u8 max_level_for_order(int order)
4776 {
4777 if (order >= KVM_HPAGE_GFN_SHIFT(PG_LEVEL_2M))
4778 return PG_LEVEL_2M;
4779
4780 return PG_LEVEL_4K;
4781 }
4782
is_large_rmp_possible(struct kvm * kvm,kvm_pfn_t pfn,int order)4783 static bool is_large_rmp_possible(struct kvm *kvm, kvm_pfn_t pfn, int order)
4784 {
4785 kvm_pfn_t pfn_aligned = ALIGN_DOWN(pfn, PTRS_PER_PMD);
4786
4787 /*
4788 * If this is a large folio, and the entire 2M range containing the
4789 * PFN is currently shared, then the entire 2M-aligned range can be
4790 * set to private via a single 2M RMP entry.
4791 */
4792 if (max_level_for_order(order) > PG_LEVEL_4K &&
4793 is_pfn_range_shared(pfn_aligned, pfn_aligned + PTRS_PER_PMD))
4794 return true;
4795
4796 return false;
4797 }
4798
sev_gmem_prepare(struct kvm * kvm,kvm_pfn_t pfn,gfn_t gfn,int max_order)4799 int sev_gmem_prepare(struct kvm *kvm, kvm_pfn_t pfn, gfn_t gfn, int max_order)
4800 {
4801 struct kvm_sev_info *sev = to_kvm_sev_info(kvm);
4802 kvm_pfn_t pfn_aligned;
4803 gfn_t gfn_aligned;
4804 int level, rc;
4805 bool assigned;
4806
4807 if (!sev_snp_guest(kvm))
4808 return 0;
4809
4810 rc = snp_lookup_rmpentry(pfn, &assigned, &level);
4811 if (rc) {
4812 pr_err_ratelimited("SEV: Failed to look up RMP entry: GFN %llx PFN %llx error %d\n",
4813 gfn, pfn, rc);
4814 return -ENOENT;
4815 }
4816
4817 if (assigned) {
4818 pr_debug("%s: already assigned: gfn %llx pfn %llx max_order %d level %d\n",
4819 __func__, gfn, pfn, max_order, level);
4820 return 0;
4821 }
4822
4823 if (is_large_rmp_possible(kvm, pfn, max_order)) {
4824 level = PG_LEVEL_2M;
4825 pfn_aligned = ALIGN_DOWN(pfn, PTRS_PER_PMD);
4826 gfn_aligned = ALIGN_DOWN(gfn, PTRS_PER_PMD);
4827 } else {
4828 level = PG_LEVEL_4K;
4829 pfn_aligned = pfn;
4830 gfn_aligned = gfn;
4831 }
4832
4833 rc = rmp_make_private(pfn_aligned, gfn_to_gpa(gfn_aligned), level, sev->asid, false);
4834 if (rc) {
4835 pr_err_ratelimited("SEV: Failed to update RMP entry: GFN %llx PFN %llx level %d error %d\n",
4836 gfn, pfn, level, rc);
4837 return -EINVAL;
4838 }
4839
4840 pr_debug("%s: updated: gfn %llx pfn %llx pfn_aligned %llx max_order %d level %d\n",
4841 __func__, gfn, pfn, pfn_aligned, max_order, level);
4842
4843 return 0;
4844 }
4845
sev_gmem_invalidate(kvm_pfn_t start,kvm_pfn_t end)4846 void sev_gmem_invalidate(kvm_pfn_t start, kvm_pfn_t end)
4847 {
4848 kvm_pfn_t pfn;
4849
4850 if (!cc_platform_has(CC_ATTR_HOST_SEV_SNP))
4851 return;
4852
4853 pr_debug("%s: PFN start 0x%llx PFN end 0x%llx\n", __func__, start, end);
4854
4855 for (pfn = start; pfn < end;) {
4856 bool use_2m_update = false;
4857 int rc, rmp_level;
4858 bool assigned;
4859
4860 rc = snp_lookup_rmpentry(pfn, &assigned, &rmp_level);
4861 if (rc || !assigned)
4862 goto next_pfn;
4863
4864 use_2m_update = IS_ALIGNED(pfn, PTRS_PER_PMD) &&
4865 end >= (pfn + PTRS_PER_PMD) &&
4866 rmp_level > PG_LEVEL_4K;
4867
4868 /*
4869 * If an unaligned PFN corresponds to a 2M region assigned as a
4870 * large page in the RMP table, PSMASH the region into individual
4871 * 4K RMP entries before attempting to convert a 4K sub-page.
4872 */
4873 if (!use_2m_update && rmp_level > PG_LEVEL_4K) {
4874 /*
4875 * This shouldn't fail, but if it does, report it, but
4876 * still try to update RMP entry to shared and pray this
4877 * was a spurious error that can be addressed later.
4878 */
4879 rc = snp_rmptable_psmash(pfn);
4880 WARN_ONCE(rc, "SEV: Failed to PSMASH RMP entry for PFN 0x%llx error %d\n",
4881 pfn, rc);
4882 }
4883
4884 rc = rmp_make_shared(pfn, use_2m_update ? PG_LEVEL_2M : PG_LEVEL_4K);
4885 if (WARN_ONCE(rc, "SEV: Failed to update RMP entry for PFN 0x%llx error %d\n",
4886 pfn, rc))
4887 goto next_pfn;
4888
4889 /*
4890 * SEV-ES avoids host/guest cache coherency issues through
4891 * WBINVD hooks issued via MMU notifiers during run-time, and
4892 * KVM's VM destroy path at shutdown. Those MMU notifier events
4893 * don't cover gmem since there is no requirement to map pages
4894 * to a HVA in order to use them for a running guest. While the
4895 * shutdown path would still likely cover things for SNP guests,
4896 * userspace may also free gmem pages during run-time via
4897 * hole-punching operations on the guest_memfd, so flush the
4898 * cache entries for these pages before free'ing them back to
4899 * the host.
4900 */
4901 clflush_cache_range(__va(pfn_to_hpa(pfn)),
4902 use_2m_update ? PMD_SIZE : PAGE_SIZE);
4903 next_pfn:
4904 pfn += use_2m_update ? PTRS_PER_PMD : 1;
4905 cond_resched();
4906 }
4907 }
4908
sev_private_max_mapping_level(struct kvm * kvm,kvm_pfn_t pfn)4909 int sev_private_max_mapping_level(struct kvm *kvm, kvm_pfn_t pfn)
4910 {
4911 int level, rc;
4912 bool assigned;
4913
4914 if (!sev_snp_guest(kvm))
4915 return 0;
4916
4917 rc = snp_lookup_rmpentry(pfn, &assigned, &level);
4918 if (rc || !assigned)
4919 return PG_LEVEL_4K;
4920
4921 return level;
4922 }
4923
sev_decrypt_vmsa(struct kvm_vcpu * vcpu)4924 struct vmcb_save_area *sev_decrypt_vmsa(struct kvm_vcpu *vcpu)
4925 {
4926 struct vcpu_svm *svm = to_svm(vcpu);
4927 struct vmcb_save_area *vmsa;
4928 struct kvm_sev_info *sev;
4929 int error = 0;
4930 int ret;
4931
4932 if (!sev_es_guest(vcpu->kvm))
4933 return NULL;
4934
4935 /*
4936 * If the VMSA has not yet been encrypted, return a pointer to the
4937 * current un-encrypted VMSA.
4938 */
4939 if (!vcpu->arch.guest_state_protected)
4940 return (struct vmcb_save_area *)svm->sev_es.vmsa;
4941
4942 sev = to_kvm_sev_info(vcpu->kvm);
4943
4944 /* Check if the SEV policy allows debugging */
4945 if (sev_snp_guest(vcpu->kvm)) {
4946 if (!(sev->policy & SNP_POLICY_DEBUG))
4947 return NULL;
4948 } else {
4949 if (sev->policy & SEV_POLICY_NODBG)
4950 return NULL;
4951 }
4952
4953 if (sev_snp_guest(vcpu->kvm)) {
4954 struct sev_data_snp_dbg dbg = {0};
4955
4956 vmsa = snp_alloc_firmware_page(__GFP_ZERO);
4957 if (!vmsa)
4958 return NULL;
4959
4960 dbg.gctx_paddr = __psp_pa(sev->snp_context);
4961 dbg.src_addr = svm->vmcb->control.vmsa_pa;
4962 dbg.dst_addr = __psp_pa(vmsa);
4963
4964 ret = sev_do_cmd(SEV_CMD_SNP_DBG_DECRYPT, &dbg, &error);
4965
4966 /*
4967 * Return the target page to a hypervisor page no matter what.
4968 * If this fails, the page can't be used, so leak it and don't
4969 * try to use it.
4970 */
4971 if (snp_page_reclaim(vcpu->kvm, PHYS_PFN(__pa(vmsa))))
4972 return NULL;
4973
4974 if (ret) {
4975 pr_err("SEV: SNP_DBG_DECRYPT failed ret=%d, fw_error=%d (%#x)\n",
4976 ret, error, error);
4977 free_page((unsigned long)vmsa);
4978
4979 return NULL;
4980 }
4981 } else {
4982 struct sev_data_dbg dbg = {0};
4983 struct page *vmsa_page;
4984
4985 vmsa_page = alloc_page(GFP_KERNEL);
4986 if (!vmsa_page)
4987 return NULL;
4988
4989 vmsa = page_address(vmsa_page);
4990
4991 dbg.handle = sev->handle;
4992 dbg.src_addr = svm->vmcb->control.vmsa_pa;
4993 dbg.dst_addr = __psp_pa(vmsa);
4994 dbg.len = PAGE_SIZE;
4995
4996 ret = sev_do_cmd(SEV_CMD_DBG_DECRYPT, &dbg, &error);
4997 if (ret) {
4998 pr_err("SEV: SEV_CMD_DBG_DECRYPT failed ret=%d, fw_error=%d (0x%x)\n",
4999 ret, error, error);
5000 __free_page(vmsa_page);
5001
5002 return NULL;
5003 }
5004 }
5005
5006 return vmsa;
5007 }
5008
sev_free_decrypted_vmsa(struct kvm_vcpu * vcpu,struct vmcb_save_area * vmsa)5009 void sev_free_decrypted_vmsa(struct kvm_vcpu *vcpu, struct vmcb_save_area *vmsa)
5010 {
5011 /* If the VMSA has not yet been encrypted, nothing was allocated */
5012 if (!vcpu->arch.guest_state_protected || !vmsa)
5013 return;
5014
5015 free_page((unsigned long)vmsa);
5016 }
5017