arch/x86/kvm/xen.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright © 2019 Oracle and/or its affiliates. All rights reserved.
   4  * Copyright © 2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
   5  *
   6  * KVM Xen emulation
   7  */
   8
   9 #include "x86.h"
  10 #include "xen.h"
  11 #include "hyperv.h"
  12
  13 #include <linux/kvm_host.h>
  14
  15 #include <trace/events/kvm.h>
  16 #include <xen/interface/xen.h>
  17
  18 #include "trace.h"
  19
  20 DEFINE_STATIC_KEY_DEFERRED_FALSE(kvm_xen_enabled, HZ);
  21
  22 static int kvm_xen_shared_info_init(struct kvm *kvm, gfn_t gfn)
  23 {
  24         gpa_t gpa = gfn_to_gpa(gfn);
  25         int wc_ofs, sec_hi_ofs;
  26         int ret;
  27         int idx = srcu_read_lock(&kvm->srcu);
  28
  29         ret = kvm_gfn_to_hva_cache_init(kvm, &kvm->arch.xen.shinfo_cache,
  30                                         gpa, PAGE_SIZE);
  31         if (ret)
  32                 goto out;
  33
  34         kvm->arch.xen.shinfo_set = true;
  35
  36         /* Paranoia checks on the 32-bit struct layout */
  37         BUILD_BUG_ON(offsetof(struct compat_shared_info, wc) != 0x900);
  38         BUILD_BUG_ON(offsetof(struct compat_shared_info, arch.wc_sec_hi) != 0x924);
  39         BUILD_BUG_ON(offsetof(struct pvclock_vcpu_time_info, version) != 0);
  40
  41         /* 32-bit location by default */
  42         wc_ofs = offsetof(struct compat_shared_info, wc);
  43         sec_hi_ofs = offsetof(struct compat_shared_info, arch.wc_sec_hi);
  44
  45 #ifdef CONFIG_X86_64
  46         /* Paranoia checks on the 64-bit struct layout */
  47         BUILD_BUG_ON(offsetof(struct shared_info, wc) != 0xc00);
  48         BUILD_BUG_ON(offsetof(struct shared_info, wc_sec_hi) != 0xc0c);
  49
  50         if (kvm->arch.xen.long_mode) {
  51                 wc_ofs = offsetof(struct shared_info, wc);
  52                 sec_hi_ofs = offsetof(struct shared_info, wc_sec_hi);
  53         }
  54 #endif
  55
  56         kvm_write_wall_clock(kvm, gpa + wc_ofs, sec_hi_ofs - wc_ofs);
  57         kvm_make_all_cpus_request(kvm, KVM_REQ_MASTERCLOCK_UPDATE);
  58
  59 out:
  60         srcu_read_unlock(&kvm->srcu, idx);
  61         return ret;
  62 }
  63
  64 int __kvm_xen_has_interrupt(struct kvm_vcpu *v)
  65 {
  66         u8 rc = 0;
  67
  68         /*
  69          * If the global upcall vector (HVMIRQ_callback_vector) is set and
  70          * the vCPU's evtchn_upcall_pending flag is set, the IRQ is pending.
  71          */
  72         struct gfn_to_hva_cache *ghc = &v->arch.xen.vcpu_info_cache;
  73         struct kvm_memslots *slots = kvm_memslots(v->kvm);
  74         unsigned int offset = offsetof(struct vcpu_info, evtchn_upcall_pending);
  75
  76         /* No need for compat handling here */
  77         BUILD_BUG_ON(offsetof(struct vcpu_info, evtchn_upcall_pending) !=
  78                      offsetof(struct compat_vcpu_info, evtchn_upcall_pending));
  79         BUILD_BUG_ON(sizeof(rc) !=
  80                      sizeof(((struct vcpu_info *)0)->evtchn_upcall_pending));
  81         BUILD_BUG_ON(sizeof(rc) !=
  82                      sizeof(((struct compat_vcpu_info *)0)->evtchn_upcall_pending));
  83
  84         /*
  85          * For efficiency, this mirrors the checks for using the valid
  86          * cache in kvm_read_guest_offset_cached(), but just uses
  87          * __get_user() instead. And falls back to the slow path.
  88          */
  89         if (likely(slots->generation == ghc->generation &&
  90                    !kvm_is_error_hva(ghc->hva) && ghc->memslot)) {
  91                 /* Fast path */
  92                 __get_user(rc, (u8 __user *)ghc->hva + offset);
  93         } else {
  94                 /* Slow path */
  95                 kvm_read_guest_offset_cached(v->kvm, ghc, &rc, offset,
  96                                              sizeof(rc));
  97         }
  98
  99         return rc;
 100 }
 101
 102 int kvm_xen_hvm_set_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 103 {
 104         int r = -ENOENT;
 105
 106         mutex_lock(&kvm->lock);
 107
 108         switch (data->type) {
 109         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 110                 if (!IS_ENABLED(CONFIG_64BIT) && data->u.long_mode) {
 111                         r = -EINVAL;
 112                 } else {
 113                         kvm->arch.xen.long_mode = !!data->u.long_mode;
 114                         r = 0;
 115                 }
 116                 break;
 117
 118         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 119                 r = kvm_xen_shared_info_init(kvm, data->u.shared_info.gfn);
 120                 break;
 121
 122
 123         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 124                 if (data->u.vector < 0x10)
 125                         r = -EINVAL;
 126                 else {
 127                         kvm->arch.xen.upcall_vector = data->u.vector;
 128                         r = 0;
 129                 }
 130                 break;
 131
 132         default:
 133                 break;
 134         }
 135
 136         mutex_unlock(&kvm->lock);
 137         return r;
 138 }
 139
 140 int kvm_xen_hvm_get_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 141 {
 142         int r = -ENOENT;
 143
 144         mutex_lock(&kvm->lock);
 145
 146         switch (data->type) {
 147         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 148                 data->u.long_mode = kvm->arch.xen.long_mode;
 149                 r = 0;
 150                 break;
 151
 152         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 153                 if (kvm->arch.xen.shinfo_set) {
 154                         data->u.shared_info.gfn = gpa_to_gfn(kvm->arch.xen.shinfo_cache.gpa);
 155                         r = 0;
 156                 }
 157                 break;
 158
 159         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 160                 data->u.vector = kvm->arch.xen.upcall_vector;
 161                 r = 0;
 162                 break;
 163
 164         default:
 165                 break;
 166         }
 167
 168         mutex_unlock(&kvm->lock);
 169         return r;
 170 }
 171
 172 int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 173 {
 174         int idx, r = -ENOENT;
 175
 176         mutex_lock(&vcpu->kvm->lock);
 177         idx = srcu_read_lock(&vcpu->kvm->srcu);
 178
 179         switch (data->type) {
 180         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 181                 /* No compat necessary here. */
 182                 BUILD_BUG_ON(sizeof(struct vcpu_info) !=
 183                              sizeof(struct compat_vcpu_info));
 184
 185                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 186                                               &vcpu->arch.xen.vcpu_info_cache,
 187                                               data->u.gpa,
 188                                               sizeof(struct vcpu_info));
 189                 if (!r) {
 190                         vcpu->arch.xen.vcpu_info_set = true;
 191                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 192                 }
 193                 break;
 194
 195         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 196                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 197                                               &vcpu->arch.xen.vcpu_time_info_cache,
 198                                               data->u.gpa,
 199                                               sizeof(struct pvclock_vcpu_time_info));
 200                 if (!r) {
 201                         vcpu->arch.xen.vcpu_time_info_set = true;
 202                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 203                 }
 204                 break;
 205
 206         default:
 207                 break;
 208         }
 209
 210         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 211         mutex_unlock(&vcpu->kvm->lock);
 212         return r;
 213 }
 214
 215 int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 216 {
 217         int r = -ENOENT;
 218
 219         mutex_lock(&vcpu->kvm->lock);
 220
 221         switch (data->type) {
 222         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 223                 if (vcpu->arch.xen.vcpu_info_set) {
 224                         data->u.gpa = vcpu->arch.xen.vcpu_info_cache.gpa;
 225                         r = 0;
 226                 }
 227                 break;
 228
 229         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 230                 if (vcpu->arch.xen.vcpu_time_info_set) {
 231                         data->u.gpa = vcpu->arch.xen.vcpu_time_info_cache.gpa;
 232                         r = 0;
 233                 }
 234                 break;
 235
 236         default:
 237                 break;
 238         }
 239
 240         mutex_unlock(&vcpu->kvm->lock);
 241         return r;
 242 }
 243
 244 int kvm_xen_write_hypercall_page(struct kvm_vcpu *vcpu, u64 data)
 245 {
 246         struct kvm *kvm = vcpu->kvm;
 247         u32 page_num = data & ~PAGE_MASK;
 248         u64 page_addr = data & PAGE_MASK;
 249         bool lm = is_long_mode(vcpu);
 250
 251         /* Latch long_mode for shared_info pages etc. */
 252         vcpu->kvm->arch.xen.long_mode = lm;
 253
 254         /*
 255          * If Xen hypercall intercept is enabled, fill the hypercall
 256          * page with VMCALL/VMMCALL instructions since that's what
 257          * we catch. Else the VMM has provided the hypercall pages
 258          * with instructions of its own choosing, so use those.
 259          */
 260         if (kvm_xen_hypercall_enabled(kvm)) {
 261                 u8 instructions[32];
 262                 int i;
 263
 264                 if (page_num)
 265                         return 1;
 266
 267                 /* mov imm32, %eax */
 268                 instructions[0] = 0xb8;
 269
 270                 /* vmcall / vmmcall */
 271                 kvm_x86_ops.patch_hypercall(vcpu, instructions + 5);
 272
 273                 /* ret */
 274                 instructions[8] = 0xc3;
 275
 276                 /* int3 to pad */
 277                 memset(instructions + 9, 0xcc, sizeof(instructions) - 9);
 278
 279                 for (i = 0; i < PAGE_SIZE / sizeof(instructions); i++) {
 280                         *(u32 *)&instructions[1] = i;
 281                         if (kvm_vcpu_write_guest(vcpu,
 282                                                  page_addr + (i * sizeof(instructions)),
 283                                                  instructions, sizeof(instructions)))
 284                                 return 1;
 285                 }
 286         } else {
 287                 /*
 288                  * Note, truncation is a non-issue as 'lm' is guaranteed to be
 289                  * false for a 32-bit kernel, i.e. when hva_t is only 4 bytes.
 290                  */
 291                 hva_t blob_addr = lm ? kvm->arch.xen_hvm_config.blob_addr_64
 292                                      : kvm->arch.xen_hvm_config.blob_addr_32;
 293                 u8 blob_size = lm ? kvm->arch.xen_hvm_config.blob_size_64
 294                                   : kvm->arch.xen_hvm_config.blob_size_32;
 295                 u8 *page;
 296
 297                 if (page_num >= blob_size)
 298                         return 1;
 299
 300                 blob_addr += page_num * PAGE_SIZE;
 301
 302                 page = memdup_user((u8 __user *)blob_addr, PAGE_SIZE);
 303                 if (IS_ERR(page))
 304                         return PTR_ERR(page);
 305
 306                 if (kvm_vcpu_write_guest(vcpu, page_addr, page, PAGE_SIZE)) {
 307                         kfree(page);
 308                         return 1;
 309                 }
 310         }
 311         return 0;
 312 }
 313
 314 int kvm_xen_hvm_config(struct kvm *kvm, struct kvm_xen_hvm_config *xhc)
 315 {
 316         if (xhc->flags & ~KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL)
 317                 return -EINVAL;
 318
 319         /*
 320          * With hypercall interception the kernel generates its own
 321          * hypercall page so it must not be provided.
 322          */
 323         if ((xhc->flags & KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL) &&
 324             (xhc->blob_addr_32 || xhc->blob_addr_64 ||
 325              xhc->blob_size_32 || xhc->blob_size_64))
 326                 return -EINVAL;
 327
 328         mutex_lock(&kvm->lock);
 329
 330         if (xhc->msr && !kvm->arch.xen_hvm_config.msr)
 331                 static_branch_inc(&kvm_xen_enabled.key);
 332         else if (!xhc->msr && kvm->arch.xen_hvm_config.msr)
 333                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
 334
 335         memcpy(&kvm->arch.xen_hvm_config, xhc, sizeof(*xhc));
 336
 337         mutex_unlock(&kvm->lock);
 338         return 0;
 339 }
 340
 341 void kvm_xen_destroy_vm(struct kvm *kvm)
 342 {
 343         if (kvm->arch.xen_hvm_config.msr)
 344                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
 345 }
 346
 347 static int kvm_xen_hypercall_set_result(struct kvm_vcpu *vcpu, u64 result)
 348 {
 349         kvm_rax_write(vcpu, result);
 350         return kvm_skip_emulated_instruction(vcpu);
 351 }
 352
 353 static int kvm_xen_hypercall_complete_userspace(struct kvm_vcpu *vcpu)
 354 {
 355         struct kvm_run *run = vcpu->run;
 356
 357         if (unlikely(!kvm_is_linear_rip(vcpu, vcpu->arch.xen.hypercall_rip)))
 358                 return 1;
 359
 360         return kvm_xen_hypercall_set_result(vcpu, run->xen.u.hcall.result);
 361 }
 362
 363 int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
 364 {
 365         bool longmode;
 366         u64 input, params[6];
 367
 368         input = (u64)kvm_register_read(vcpu, VCPU_REGS_RAX);
 369
 370         /* Hyper-V hypercalls get bit 31 set in EAX */
 371         if ((input & 0x80000000) &&
 372             kvm_hv_hypercall_enabled(vcpu))
 373                 return kvm_hv_hypercall(vcpu);
 374
 375         longmode = is_64_bit_mode(vcpu);
 376         if (!longmode) {
 377                 params[0] = (u32)kvm_rbx_read(vcpu);
 378                 params[1] = (u32)kvm_rcx_read(vcpu);
 379                 params[2] = (u32)kvm_rdx_read(vcpu);
 380                 params[3] = (u32)kvm_rsi_read(vcpu);
 381                 params[4] = (u32)kvm_rdi_read(vcpu);
 382                 params[5] = (u32)kvm_rbp_read(vcpu);
 383         }
 384 #ifdef CONFIG_X86_64
 385         else {
 386                 params[0] = (u64)kvm_rdi_read(vcpu);
 387                 params[1] = (u64)kvm_rsi_read(vcpu);
 388                 params[2] = (u64)kvm_rdx_read(vcpu);
 389                 params[3] = (u64)kvm_r10_read(vcpu);
 390                 params[4] = (u64)kvm_r8_read(vcpu);
 391                 params[5] = (u64)kvm_r9_read(vcpu);
 392         }
 393 #endif
 394         trace_kvm_xen_hypercall(input, params[0], params[1], params[2],
 395                                 params[3], params[4], params[5]);
 396
 397         vcpu->run->exit_reason = KVM_EXIT_XEN;
 398         vcpu->run->xen.type = KVM_EXIT_XEN_HCALL;
 399         vcpu->run->xen.u.hcall.longmode = longmode;
 400         vcpu->run->xen.u.hcall.cpl = kvm_x86_ops.get_cpl(vcpu);
 401         vcpu->run->xen.u.hcall.input = input;
 402         vcpu->run->xen.u.hcall.params[0] = params[0];
 403         vcpu->run->xen.u.hcall.params[1] = params[1];
 404         vcpu->run->xen.u.hcall.params[2] = params[2];
 405         vcpu->run->xen.u.hcall.params[3] = params[3];
 406         vcpu->run->xen.u.hcall.params[4] = params[4];
 407         vcpu->run->xen.u.hcall.params[5] = params[5];
 408         vcpu->arch.xen.hypercall_rip = kvm_get_linear_rip(vcpu);
 409         vcpu->arch.complete_userspace_io =
 410                 kvm_xen_hypercall_complete_userspace;
 411
 412         return 0;
 413 }