palacios/src/palacios/vmm_shadow_paging.c

   1 /*
   2  * This file is part of the Palacios Virtual Machine Monitor developed
   3  * by the V3VEE Project with funding from the United States National
   4  * Science Foundation and the Department of Energy.
   5  *
   6  * The V3VEE Project is a joint project between Northwestern University
   7  * and the University of New Mexico.  You can find out more at
   8  * http://www.v3vee.org
   9  *
  10  * Copyright (c) 2008, Jack Lange <jarusl@cs.northwestern.edu>
  11  * Copyright (c) 2008, The V3VEE Project <http://www.v3vee.org>
  12  * All rights reserved.
  13  *
  14  * Author: Jack Lange <jarusl@cs.northwestern.edu>
  15  *
  16  * This is free software.  You are permitted to use,
  17  * redistribute, and modify it as specified in the file "V3VEE_LICENSE".
  18  */
  19
  20
  21 #include <palacios/vmm_shadow_paging.h>
  22
  23
  24 #include <palacios/vmm.h>
  25 #include <palacios/vm_guest_mem.h>
  26 #include <palacios/vmm_decoder.h>
  27 #include <palacios/vmm_ctrl_regs.h>
  28
  29 #include <palacios/vmm_hashtable.h>
  30
  31 #include <palacios/vmm_direct_paging.h>
  32
  33
  34
  35
  36 #ifdef V3_CONFIG_SHADOW_PAGING_TELEMETRY
  37 #include <palacios/vmm_telemetry.h>
  38 #endif
  39
  40 #ifdef V3_CONFIG_SYMBIOTIC_SWAP
  41 #include <palacios/vmm_sym_swap.h>
  42 #endif
  43
  44 #ifndef V3_CONFIG_DEBUG_SHADOW_PAGING
  45 #undef PrintDebug
  46 #define PrintDebug(fmt, args...)
  47 #endif
  48
  49
  50 static const char default_strategy[] = "VTLB";
  51
  52
  53 static struct hashtable * master_shdw_pg_table = NULL;
  54
  55
  56 struct event_callback {
  57     int (*callback)(struct guest_info *core, struct v3_shdw_pg_event *event, void *priv_data);
  58     void *priv_data;
  59
  60     struct list_head node;
  61 };
  62
  63 static uint_t shdw_pg_hash_fn(addr_t key) {
  64     char * name = (char *)key;
  65     return v3_hash_buffer((uint8_t *)name, strlen(name));
  66 }
  67
  68 static int shdw_pg_eq_fn(addr_t key1, addr_t key2) {
  69     char * name1 = (char *)key1;
  70     char * name2 = (char *)key2;
  71
  72     return (strcmp(name1, name2) == 0);
  73 }
  74
  75 static int have_callbacks(struct guest_info *core)
  76 {
  77     return !list_empty(&(core->vm_info->shdw_impl.event_callback_list));
  78 }
  79
  80 static void dispatch_event(struct guest_info *core, struct v3_shdw_pg_event *event)
  81 {
  82     struct event_callback *cb,*temp;
  83
  84     list_for_each_entry_safe(cb,
  85                              temp,
  86                              &(core->vm_info->shdw_impl.event_callback_list),
  87                              node) {
  88         cb->callback(core,event,cb->priv_data);
  89     }
  90 }
  91
  92
  93 int V3_init_shdw_paging() {
  94     extern struct v3_shdw_pg_impl * __start__v3_shdw_pg_impls[];
  95     extern struct v3_shdw_pg_impl * __stop__v3_shdw_pg_impls[];
  96     struct v3_shdw_pg_impl ** tmp_impl = __start__v3_shdw_pg_impls;
  97     int i = 0;
  98
  99     master_shdw_pg_table = v3_create_htable(0, shdw_pg_hash_fn, shdw_pg_eq_fn);
 100
 101
 102     while (tmp_impl != __stop__v3_shdw_pg_impls) {
 103         V3_Print(VM_NONE, VCORE_NONE, "Registering Shadow Paging Impl (%s)\n", (*tmp_impl)->name);
 104
 105         if (v3_htable_search(master_shdw_pg_table, (addr_t)((*tmp_impl)->name))) {
 106             PrintError(VM_NONE, VCORE_NONE, "Multiple instances of shadow paging impl (%s)\n", (*tmp_impl)->name);
 107             return -1;
 108         }
 109
 110         if (v3_htable_insert(master_shdw_pg_table,
 111                              (addr_t)((*tmp_impl)->name),
 112                              (addr_t)(*tmp_impl)) == 0) {
 113             PrintError(VM_NONE, VCORE_NONE, "Could not register shadow paging impl (%s)\n", (*tmp_impl)->name);
 114             return -1;
 115         }
 116
 117         tmp_impl = &(__start__v3_shdw_pg_impls[++i]);
 118     }
 119
 120     return 0;
 121 }
 122
 123 int V3_deinit_shdw_paging() {
 124     v3_free_htable(master_shdw_pg_table, 0, 0);
 125     return 0;
 126 }
 127
 128
 129
 130 /***
 131  ***  There be dragons
 132  ***/
 133
 134
 135 #ifdef V3_CONFIG_SHADOW_PAGING_TELEMETRY
 136 static void telemetry_cb(struct v3_vm_info * vm, void * private_data, char * hdr) {
 137     int i = 0;
 138     for (i = 0; i < vm->num_cores; i++) {
 139         struct guest_info * core = &(vm->cores[i]);
 140
 141         V3_Print(vm, core, "%s Guest Page faults: %d\n", hdr, core->shdw_pg_state.guest_faults);
 142     }
 143 }
 144 #endif
 145
 146
 147
 148 int v3_init_shdw_pg_state(struct guest_info * core) {
 149     struct v3_shdw_pg_state * state = &(core->shdw_pg_state);
 150     struct v3_shdw_pg_impl * impl = core->vm_info->shdw_impl.current_impl;
 151
 152
 153     state->guest_cr3 = 0;
 154     state->guest_cr0 = 0;
 155     state->guest_efer.value = 0x0LL;
 156
 157     if (impl->local_init(core) == -1) {
 158         PrintError(core->vm_info, core, "Error in Shadow paging local initialization (%s)\n", impl->name);
 159         return -1;
 160     }
 161
 162
 163 #ifdef V3_CONFIG_SHADOW_PAGING_TELEMETRY
 164     v3_add_telemetry_cb(core->vm_info, telemetry_cb, NULL);
 165 #endif
 166
 167
 168     return 0;
 169 }
 170
 171
 172 int v3_deinit_shdw_pg_state(struct guest_info * core) {
 173     struct v3_shdw_pg_impl * impl = core->vm_info->shdw_impl.current_impl;
 174
 175     if (impl->local_deinit(core) == -1) {
 176         PrintError(core->vm_info, core, "Error deinitializing shadow paging state\n");
 177         return -1;
 178     }
 179
 180
 181     return 0;
 182 }
 183
 184
 185
 186 int v3_init_shdw_impl(struct v3_vm_info * vm) {
 187     struct v3_shdw_impl_state * impl_state = &(vm->shdw_impl);
 188     v3_cfg_tree_t * pg_cfg = v3_cfg_subtree(vm->cfg_data->cfg, "paging");
 189     char * pg_mode = v3_cfg_val(pg_cfg, "mode");
 190     char * pg_strat = v3_cfg_val(pg_cfg, "strategy");
 191     struct v3_shdw_pg_impl * impl = NULL;
 192
 193     PrintDebug(vm, VCORE_NONE, "Checking if shadow paging requested.\n");
 194     if ((pg_mode != NULL) && (strcasecmp(pg_mode, "nested") == 0)) {
 195         PrintDebug(vm, VCORE_NONE, "Nested paging specified - not initializing shadow paging.\n");
 196         return 0;
 197     }
 198
 199     if (pg_strat == NULL) {
 200         pg_strat = (char *)default_strategy;
 201     }
 202
 203     V3_Print(vm, VCORE_NONE,"Initialization of Shadow Paging implementation\n");
 204
 205     impl = (struct v3_shdw_pg_impl *)v3_htable_search(master_shdw_pg_table, (addr_t)pg_strat);
 206
 207     if (impl == NULL) {
 208         PrintError(vm, VCORE_NONE, "Could not find shadow paging impl (%s)\n", pg_strat);
 209         return -1;
 210     }
 211
 212     INIT_LIST_HEAD(&(impl_state->event_callback_list));
 213
 214     impl_state->current_impl = impl;
 215
 216     if (impl->init(vm, pg_cfg) == -1) {
 217         PrintError(vm, VCORE_NONE, "Could not initialize Shadow paging implemenation (%s)\n", impl->name);
 218         return -1;
 219     }
 220
 221
 222
 223     return 0;
 224 }
 225
 226 int v3_deinit_shdw_impl(struct v3_vm_info * vm) {
 227     struct v3_shdw_pg_impl * impl = vm->shdw_impl.current_impl;
 228     struct event_callback *cb,*temp;
 229
 230     if (impl == NULL) {
 231         // Shadow paging not implemented
 232         return 0;
 233     }
 234
 235     if (impl->deinit(vm) == -1) {
 236         PrintError(vm, VCORE_NONE,"Error deinitializing shadow paging implementation\n");
 237         return -1;
 238     }
 239
 240     list_for_each_entry_safe(cb,
 241                              temp,
 242                              &(vm->shdw_impl.event_callback_list),
 243                              node) {
 244         list_del(&(cb->node));
 245         V3_Free(cb);
 246     }
 247
 248     return 0;
 249 }
 250
 251
 252 // Reads the guest CR3 register
 253 // creates new shadow page tables
 254 // updates the shadow CR3 register to point to the new pts
 255 int v3_activate_shadow_pt(struct guest_info * core) {
 256     struct v3_shdw_impl_state * state = &(core->vm_info->shdw_impl);
 257     struct v3_shdw_pg_impl * impl = state->current_impl;
 258
 259     if (!have_callbacks(core)) {
 260         return impl->activate_shdw_pt(core);
 261     } else {
 262         int rc;
 263         struct v3_shdw_pg_event event_pre={SHADOW_ACTIVATE,SHADOW_PREIMPL,0,{0,0,0,0,0,0}};
 264         struct v3_shdw_pg_event event_post={SHADOW_ACTIVATE,SHADOW_POSTIMPL,0,{0,0,0,0,0,0}};
 265
 266         dispatch_event(core,&event_pre);
 267
 268         rc =impl->activate_shdw_pt(core);
 269
 270         dispatch_event(core,&event_post);
 271
 272         return rc;
 273     }
 274 }
 275
 276
 277
 278 // This must flush any caches
 279 // and reset the cr3 value to the correct value
 280 int v3_invalidate_shadow_pts(struct guest_info * core) {
 281     struct v3_shdw_impl_state * state = &(core->vm_info->shdw_impl);
 282     struct v3_shdw_pg_impl * impl = state->current_impl;
 283
 284     if (!have_callbacks(core)) {
 285         return impl->invalidate_shdw_pt(core);
 286     } else {
 287         int rc;
 288         struct v3_shdw_pg_event event_pre={SHADOW_INVALIDATE,SHADOW_PREIMPL,0,{0,0,0,0,0,0}};
 289         struct v3_shdw_pg_event event_post={SHADOW_INVALIDATE,SHADOW_POSTIMPL,0,{0,0,0,0,0,0}};
 290
 291         dispatch_event(core,&event_pre);
 292
 293         rc = impl->invalidate_shdw_pt(core);
 294
 295         dispatch_event(core,&event_post);
 296
 297         return rc;
 298     }
 299 }
 300
 301
 302 int v3_handle_shadow_pagefault(struct guest_info * core, addr_t fault_addr, pf_error_t error_code)
 303 {
 304     int rc;
 305
 306
 307     if (have_callbacks(core)) {
 308         struct v3_shdw_pg_event event={SHADOW_PAGEFAULT,SHADOW_PREIMPL,fault_addr,error_code};
 309         dispatch_event(core,&event);
 310     }
 311
 312     if (v3_get_vm_mem_mode(core) == PHYSICAL_MEM) {
 313         // If paging is not turned on we need to handle the special cases
 314         rc = v3_handle_passthrough_pagefault(core, fault_addr, error_code);
 315     } else if (v3_get_vm_mem_mode(core) == VIRTUAL_MEM) {
 316         struct v3_shdw_impl_state * state = &(core->vm_info->shdw_impl);
 317         struct v3_shdw_pg_impl * impl = state->current_impl;
 318
 319         rc = impl->handle_pagefault(core, fault_addr, error_code);
 320     } else {
 321         PrintError(core->vm_info, core, "Invalid Memory mode\n");
 322         rc = -1;
 323     }
 324
 325     if (have_callbacks(core)) {
 326         struct v3_shdw_pg_event event={SHADOW_PAGEFAULT,SHADOW_POSTIMPL,fault_addr,error_code};
 327         dispatch_event(core,&event);
 328     }
 329
 330     return rc;
 331 }
 332
 333
 334 int v3_handle_shadow_invlpg(struct guest_info * core) {
 335     uchar_t instr[15];
 336     struct x86_instr dec_instr;
 337     int ret = 0;
 338     addr_t vaddr = 0;
 339
 340     if (v3_get_vm_mem_mode(core) != VIRTUAL_MEM) {
 341         // Paging must be turned on...
 342         // should handle with some sort of fault I think
 343         PrintError(core->vm_info, core, "ERROR: INVLPG called in non paged mode\n");
 344         return -1;
 345     }
 346
 347     if (v3_get_vm_mem_mode(core) == PHYSICAL_MEM) {
 348         ret = v3_read_gpa_memory(core, get_addr_linear(core, core->rip, &(core->segments.cs)), 15, instr);
 349     } else {
 350         ret = v3_read_gva_memory(core, get_addr_linear(core, core->rip, &(core->segments.cs)), 15, instr);
 351     }
 352
 353     if (ret == -1) {
 354         PrintError(core->vm_info, core, "Could not read instruction into buffer\n");
 355         return -1;
 356     }
 357
 358     if (v3_decode(core, (addr_t)instr, &dec_instr) == -1) {
 359         PrintError(core->vm_info, core, "Decoding Error\n");
 360         return -1;
 361     }
 362
 363     if ((dec_instr.op_type != V3_OP_INVLPG) ||
 364         (dec_instr.num_operands != 1) ||
 365         (dec_instr.dst_operand.type != MEM_OPERAND)) {
 366         PrintError(core->vm_info, core, "Decoder Error: Not a valid INVLPG instruction...\n");
 367         return -1;
 368     }
 369
 370     vaddr = dec_instr.dst_operand.operand;
 371
 372     core->rip += dec_instr.instr_length;
 373
 374     {
 375         struct v3_shdw_impl_state * state = &(core->vm_info->shdw_impl);
 376         struct v3_shdw_pg_impl * impl = state->current_impl;
 377         int rc;
 378
 379         if (have_callbacks(core)) {
 380             struct v3_shdw_pg_event event={SHADOW_INVLPG,SHADOW_PREIMPL,vaddr,{0,0,0,0,0,0}};
 381             dispatch_event(core,&event);
 382         }
 383
 384         rc=impl->handle_invlpg(core, vaddr);
 385
 386         if (have_callbacks(core)) {
 387             struct v3_shdw_pg_event event={SHADOW_INVLPG,SHADOW_POSTIMPL,vaddr,{0,0,0,0,0,0}};
 388             dispatch_event(core,&event);
 389         }
 390
 391         return rc;
 392     }
 393 }
 394
 395
 396
 397
 398
 399
 400 int v3_inject_guest_pf(struct guest_info * core, addr_t fault_addr, pf_error_t error_code) {
 401     core->ctrl_regs.cr2 = fault_addr;
 402
 403 #ifdef V3_CONFIG_SHADOW_PAGING_TELEMETRY
 404     core->shdw_pg_state.guest_faults++;
 405 #endif
 406
 407     return v3_raise_exception_with_error(core, PF_EXCEPTION, *(uint_t *)&error_code);
 408 }
 409
 410
 411 int v3_is_guest_pf(pt_access_status_t guest_access, pt_access_status_t shadow_access) {
 412     /* basically the reasoning is that there can be multiple reasons for a page fault:
 413        If there is a permissions failure for a page present in the guest _BUT_
 414        the reason for the fault was that the page is not present in the shadow,
 415        _THEN_ we have to map the shadow page in and reexecute, this will generate
 416        a permissions fault which is _THEN_ valid to send to the guest
 417        _UNLESS_ both the guest and shadow have marked the page as not present
 418
 419        whew...
 420     */
 421     if (guest_access != PT_ACCESS_OK) {
 422         // Guest Access Error
 423
 424         if ((shadow_access != PT_ACCESS_NOT_PRESENT) &&
 425             (guest_access != PT_ACCESS_NOT_PRESENT)) {
 426             // aka (guest permission error)
 427             return 1;
 428         }
 429
 430         /*
 431           if ((shadow_access == PT_ACCESS_NOT_PRESENT) &&
 432           (guest_access == PT_ACCESS_NOT_PRESENT)) {
 433           // Page tables completely blank, handle guest first
 434           return 1;
 435           }
 436         */
 437
 438         if (guest_access == PT_ACCESS_NOT_PRESENT) {
 439             // Page tables completely blank, handle guest first
 440             return 1;
 441         }
 442
 443         // Otherwise we'll handle the guest fault later...?
 444     }
 445
 446     return 0;
 447 }
 448
 449
 450 int v3_register_shadow_paging_event_callback(struct v3_vm_info *vm,
 451                                              int (*callback)(struct guest_info *core,
 452                                                              struct v3_shdw_pg_event *event,
 453                                                              void      *priv_data),
 454                                              void *priv_data)
 455 {
 456     struct event_callback *ec = V3_Malloc(sizeof(struct event_callback));
 457
 458     if (!ec) {
 459         PrintError(vm, VCORE_NONE, "Unable to allocate for a shadow paging event callback\n");
 460         return -1;
 461     }
 462
 463     ec->callback = callback;
 464     ec->priv_data = priv_data;
 465
 466     list_add(&(ec->node),&(vm->shdw_impl.event_callback_list));
 467
 468     return 0;
 469
 470 }
 471
 472 int v3_unregister_shadow_paging_event_callback(struct v3_vm_info *vm,
 473                                                int (*callback)(struct guest_info *core,
 474                                                                struct v3_shdw_pg_event *event,
 475                                                                void      *priv_data),
 476                                                void *priv_data)
 477 {
 478     struct event_callback *cb,*temp;
 479
 480     list_for_each_entry_safe(cb,
 481                              temp,
 482                              &(vm->shdw_impl.event_callback_list),
 483                              node) {
 484         if ((callback == cb->callback) && (priv_data == cb->priv_data)) {
 485             list_del(&(cb->node));
 486             V3_Free(cb);
 487             return 0;
 488         }
 489     }
 490
 491     PrintError(vm, VCORE_NONE, "No callback found!\n");
 492
 493     return -1;
 494 }
 495
 496