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         mutex_unlock(&kvm->lock);
 109
 110         switch (data->type) {
 111         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 112                 if (!IS_ENABLED(CONFIG_64BIT) && data->u.long_mode) {
 113                         r = -EINVAL;
 114                 } else {
 115                         kvm->arch.xen.long_mode = !!data->u.long_mode;
 116                         r = 0;
 117                 }
 118                 break;
 119
 120         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 121                 r = kvm_xen_shared_info_init(kvm, data->u.shared_info.gfn);
 122                 break;
 123
 124
 125         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 126                 if (data->u.vector < 0x10)
 127                         r = -EINVAL;
 128                 else {
 129                         kvm->arch.xen.upcall_vector = data->u.vector;
 130                         r = 0;
 131                 }
 132                 break;
 133
 134         default:
 135                 break;
 136         }
 137
 138         mutex_unlock(&kvm->lock);
 139         return r;
 140 }
 141
 142 int kvm_xen_hvm_get_attr(struct kvm *kvm, struct kvm_xen_hvm_attr *data)
 143 {
 144         int r = -ENOENT;
 145
 146         mutex_lock(&kvm->lock);
 147
 148         switch (data->type) {
 149         case KVM_XEN_ATTR_TYPE_LONG_MODE:
 150                 data->u.long_mode = kvm->arch.xen.long_mode;
 151                 r = 0;
 152                 break;
 153
 154         case KVM_XEN_ATTR_TYPE_SHARED_INFO:
 155                 if (kvm->arch.xen.shinfo_set) {
 156                         data->u.shared_info.gfn = gpa_to_gfn(kvm->arch.xen.shinfo_cache.gpa);
 157                         r = 0;
 158                 }
 159                 break;
 160
 161         case KVM_XEN_ATTR_TYPE_UPCALL_VECTOR:
 162                 data->u.vector = kvm->arch.xen.upcall_vector;
 163                 r = 0;
 164                 break;
 165
 166         default:
 167                 break;
 168         }
 169
 170         mutex_unlock(&kvm->lock);
 171         return r;
 172 }
 173
 174 int kvm_xen_vcpu_set_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 175 {
 176         int idx, r = -ENOENT;
 177
 178         mutex_lock(&vcpu->kvm->lock);
 179         idx = srcu_read_lock(&vcpu->kvm->srcu);
 180
 181         switch (data->type) {
 182         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 183                 /* No compat necessary here. */
 184                 BUILD_BUG_ON(sizeof(struct vcpu_info) !=
 185                              sizeof(struct compat_vcpu_info));
 186
 187                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 188                                               &vcpu->arch.xen.vcpu_info_cache,
 189                                               data->u.gpa,
 190                                               sizeof(struct vcpu_info));
 191                 if (!r) {
 192                         vcpu->arch.xen.vcpu_info_set = true;
 193                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 194                 }
 195                 break;
 196
 197         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 198                 r = kvm_gfn_to_hva_cache_init(vcpu->kvm,
 199                                               &vcpu->arch.xen.vcpu_time_info_cache,
 200                                               data->u.gpa,
 201                                               sizeof(struct pvclock_vcpu_time_info));
 202                 if (!r) {
 203                         vcpu->arch.xen.vcpu_time_info_set = true;
 204                         kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu);
 205                 }
 206                 break;
 207
 208         default:
 209                 break;
 210         }
 211
 212         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 213         mutex_unlock(&vcpu->kvm->lock);
 214         return r;
 215 }
 216
 217 int kvm_xen_vcpu_get_attr(struct kvm_vcpu *vcpu, struct kvm_xen_vcpu_attr *data)
 218 {
 219         int r = -ENOENT;
 220
 221         mutex_lock(&vcpu->kvm->lock);
 222
 223         switch (data->type) {
 224         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_INFO:
 225                 if (vcpu->arch.xen.vcpu_info_set) {
 226                         data->u.gpa = vcpu->arch.xen.vcpu_info_cache.gpa;
 227                         r = 0;
 228                 }
 229                 break;
 230
 231         case KVM_XEN_VCPU_ATTR_TYPE_VCPU_TIME_INFO:
 232                 if (vcpu->arch.xen.vcpu_time_info_set) {
 233                         data->u.gpa = vcpu->arch.xen.vcpu_time_info_cache.gpa;
 234                         r = 0;
 235                 }
 236                 break;
 237
 238         default:
 239                 break;
 240         }
 241
 242         mutex_unlock(&vcpu->kvm->lock);
 243         return r;
 244 }
 245
 246 int kvm_xen_write_hypercall_page(struct kvm_vcpu *vcpu, u64 data)
 247 {
 248         struct kvm *kvm = vcpu->kvm;
 249         u32 page_num = data & ~PAGE_MASK;
 250         u64 page_addr = data & PAGE_MASK;
 251         bool lm = is_long_mode(vcpu);
 252
 253         /* Latch long_mode for shared_info pages etc. */
 254         vcpu->kvm->arch.xen.long_mode = lm;
 255
 256         /*
 257          * If Xen hypercall intercept is enabled, fill the hypercall
 258          * page with VMCALL/VMMCALL instructions since that's what
 259          * we catch. Else the VMM has provided the hypercall pages
 260          * with instructions of its own choosing, so use those.
 261          */
 262         if (kvm_xen_hypercall_enabled(kvm)) {
 263                 u8 instructions[32];
 264                 int i;
 265
 266                 if (page_num)
 267                         return 1;
 268
 269                 /* mov imm32, %eax */
 270                 instructions[0] = 0xb8;
 271
 272                 /* vmcall / vmmcall */
 273                 kvm_x86_ops.patch_hypercall(vcpu, instructions + 5);
 274
 275                 /* ret */
 276                 instructions[8] = 0xc3;
 277
 278                 /* int3 to pad */
 279                 memset(instructions + 9, 0xcc, sizeof(instructions) - 9);
 280
 281                 for (i = 0; i < PAGE_SIZE / sizeof(instructions); i++) {
 282                         *(u32 *)&instructions[1] = i;
 283                         if (kvm_vcpu_write_guest(vcpu,
 284                                                  page_addr + (i * sizeof(instructions)),
 285                                                  instructions, sizeof(instructions)))
 286                                 return 1;
 287                 }
 288         } else {
 289                 u64 blob_addr = lm ? kvm->arch.xen_hvm_config.blob_addr_64
 290                                    : kvm->arch.xen_hvm_config.blob_addr_32;
 291                 u8 blob_size = lm ? kvm->arch.xen_hvm_config.blob_size_64
 292                                   : kvm->arch.xen_hvm_config.blob_size_32;
 293                 u8 *page;
 294
 295                 if (page_num >= blob_size)
 296                         return 1;
 297
 298                 blob_addr += page_num * PAGE_SIZE;
 299
 300                 page = memdup_user((u8 __user *)blob_addr, PAGE_SIZE);
 301                 if (IS_ERR(page))
 302                         return PTR_ERR(page);
 303
 304                 if (kvm_vcpu_write_guest(vcpu, page_addr, page, PAGE_SIZE)) {
 305                         kfree(page);
 306                         return 1;
 307                 }
 308         }
 309         return 0;
 310 }
 311
 312 int kvm_xen_hvm_config(struct kvm *kvm, struct kvm_xen_hvm_config *xhc)
 313 {
 314         if (xhc->flags & ~KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL)
 315                 return -EINVAL;
 316
 317         /*
 318          * With hypercall interception the kernel generates its own
 319          * hypercall page so it must not be provided.
 320          */
 321         if ((xhc->flags & KVM_XEN_HVM_CONFIG_INTERCEPT_HCALL) &&
 322             (xhc->blob_addr_32 || xhc->blob_addr_64 ||
 323              xhc->blob_size_32 || xhc->blob_size_64))
 324                 return -EINVAL;
 325
 326         mutex_lock(&kvm->lock);
 327
 328         if (xhc->msr && !kvm->arch.xen_hvm_config.msr)
 329                 static_branch_inc(&kvm_xen_enabled.key);
 330         else if (!xhc->msr && kvm->arch.xen_hvm_config.msr)
 331                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
 332
 333         memcpy(&kvm->arch.xen_hvm_config, xhc, sizeof(*xhc));
 334
 335         mutex_unlock(&kvm->lock);
 336         return 0;
 337 }
 338
 339 void kvm_xen_destroy_vm(struct kvm *kvm)
 340 {
 341         if (kvm->arch.xen_hvm_config.msr)
 342                 static_branch_slow_dec_deferred(&kvm_xen_enabled);
 343 }
 344
 345 static int kvm_xen_hypercall_set_result(struct kvm_vcpu *vcpu, u64 result)
 346 {
 347         kvm_rax_write(vcpu, result);
 348         return kvm_skip_emulated_instruction(vcpu);
 349 }
 350
 351 static int kvm_xen_hypercall_complete_userspace(struct kvm_vcpu *vcpu)
 352 {
 353         struct kvm_run *run = vcpu->run;
 354
 355         if (unlikely(!kvm_is_linear_rip(vcpu, vcpu->arch.xen.hypercall_rip)))
 356                 return 1;
 357
 358         return kvm_xen_hypercall_set_result(vcpu, run->xen.u.hcall.result);
 359 }
 360
 361 int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
 362 {
 363         bool longmode;
 364         u64 input, params[6];
 365
 366         input = (u64)kvm_register_read(vcpu, VCPU_REGS_RAX);
 367
 368         /* Hyper-V hypercalls get bit 31 set in EAX */
 369         if ((input & 0x80000000) &&
 370             kvm_hv_hypercall_enabled(vcpu->kvm))
 371                 return kvm_hv_hypercall(vcpu);
 372
 373         longmode = is_64_bit_mode(vcpu);
 374         if (!longmode) {
 375                 params[0] = (u32)kvm_rbx_read(vcpu);
 376                 params[1] = (u32)kvm_rcx_read(vcpu);
 377                 params[2] = (u32)kvm_rdx_read(vcpu);
 378                 params[3] = (u32)kvm_rsi_read(vcpu);
 379                 params[4] = (u32)kvm_rdi_read(vcpu);
 380                 params[5] = (u32)kvm_rbp_read(vcpu);
 381         }
 382 #ifdef CONFIG_X86_64
 383         else {
 384                 params[0] = (u64)kvm_rdi_read(vcpu);
 385                 params[1] = (u64)kvm_rsi_read(vcpu);
 386                 params[2] = (u64)kvm_rdx_read(vcpu);
 387                 params[3] = (u64)kvm_r10_read(vcpu);
 388                 params[4] = (u64)kvm_r8_read(vcpu);
 389                 params[5] = (u64)kvm_r9_read(vcpu);
 390         }
 391 #endif
 392         trace_kvm_xen_hypercall(input, params[0], params[1], params[2],
 393                                 params[3], params[4], params[5]);
 394
 395         vcpu->run->exit_reason = KVM_EXIT_XEN;
 396         vcpu->run->xen.type = KVM_EXIT_XEN_HCALL;
 397         vcpu->run->xen.u.hcall.longmode = longmode;
 398         vcpu->run->xen.u.hcall.cpl = kvm_x86_ops.get_cpl(vcpu);
 399         vcpu->run->xen.u.hcall.input = input;
 400         vcpu->run->xen.u.hcall.params[0] = params[0];
 401         vcpu->run->xen.u.hcall.params[1] = params[1];
 402         vcpu->run->xen.u.hcall.params[2] = params[2];
 403         vcpu->run->xen.u.hcall.params[3] = params[3];
 404         vcpu->run->xen.u.hcall.params[4] = params[4];
 405         vcpu->run->xen.u.hcall.params[5] = params[5];
 406         vcpu->arch.xen.hypercall_rip = kvm_get_linear_rip(vcpu);
 407         vcpu->arch.complete_userspace_io =
 408                 kvm_xen_hypercall_complete_userspace;
 409
 410         return 0;
 411 }