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.
6 * The V3VEE Project is a joint project between Northwestern University
7 * and the University of New Mexico. You can find out more at
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.
14 * Author: Jack Lange <jarusl@cs.northwestern.edu>
16 * This is free software. You are permitted to use,
17 * redistribute, and modify it as specified in the file "V3VEE_LICENSE".
20 #include <palacios/vmm_mem.h>
21 #include <palacios/vmm.h>
22 #include <palacios/vmm_util.h>
23 #include <palacios/vmm_emulator.h>
24 #include <palacios/vm_guest.h>
25 #include <palacios/vmm_debug.h>
27 #include <palacios/vmm_shadow_paging.h>
28 #include <palacios/vmm_direct_paging.h>
30 #include <interfaces/vmm_numa.h>
32 uint64_t v3_mem_block_size = V3_CONFIG_MEM_BLOCK_SIZE;
35 struct v3_mem_region * v3_get_base_region(struct v3_vm_info * vm, addr_t gpa) {
36 struct v3_mem_map * map = &(vm->mem_map);
37 uint32_t block_index = gpa / v3_mem_block_size;
39 if ((gpa >= (map->num_base_regions * v3_mem_block_size)) ||
40 (block_index >= map->num_base_regions)) {
41 PrintError(vm, VCORE_NONE, "Guest Address Exceeds Base Memory Size (ga=0x%p), (limit=0x%p)\n",
42 (void *)gpa, (void *)vm->mem_size);
49 return &(map->base_regions[block_index]);
53 static int mem_offset_hypercall(struct guest_info * info, uint_t hcall_id, void * private_data) {
55 PrintDebug(info->vm_info, info,"V3Vee: Memory offset hypercall (offset=%p)\n",
56 (void *)(info->vm_info->mem_map.base_region.host_addr));
58 info->vm_regs.rbx = info->vm_info->mem_map.base_region.host_addr;
63 static int unhandled_err(struct guest_info * core, addr_t guest_va, addr_t guest_pa,
64 struct v3_mem_region * reg, pf_error_t access_info) {
66 PrintError(core->vm_info, core, "Unhandled memory access error (gpa=%p, gva=%p, error_code=%d)\n",
67 (void *)guest_pa, (void *)guest_va, *(uint32_t *)&access_info);
69 v3_print_mem_map(core->vm_info);
71 v3_print_guest_state(core);
76 static int gpa_to_node_from_cfg(struct v3_vm_info * vm, addr_t gpa) {
77 v3_cfg_tree_t * layout_cfg = v3_cfg_subtree(vm->cfg_data->cfg, "mem_layout");
78 v3_cfg_tree_t * region_desc = v3_cfg_subtree(layout_cfg, "region");
81 char * start_addr_str = v3_cfg_val(region_desc, "start_addr");
82 char * end_addr_str = v3_cfg_val(region_desc, "end_addr");
83 char * node_id_str = v3_cfg_val(region_desc, "node");
85 addr_t start_addr = 0;
89 if ((!start_addr_str) || (!end_addr_str) || (!node_id_str)) {
90 PrintError(vm, VCORE_NONE, "Invalid memory layout in configuration\n");
94 start_addr = atox(start_addr_str);
95 end_addr = atox(end_addr_str);
96 node_id = atoi(node_id_str);
98 if ((gpa >= start_addr) && (gpa < end_addr)) {
102 region_desc = v3_cfg_next_branch(region_desc);
109 // This code parallels that in vmm_shadow_paging.c:v3_init_shdw_impl()
110 // and vmm_config.c:determine_paging_mode. The determination of which
111 // paging mode will be used is determined much later than the allocation of
112 // the guest memory regions, so we need to do this here to decide if they
113 // need to be below 4 GB or not.
114 static int will_use_shadow_paging(struct v3_vm_info *vm)
116 v3_cfg_tree_t * pg_cfg = v3_cfg_subtree(vm->cfg_data->cfg, "paging");
117 char * pg_mode = v3_cfg_val(pg_cfg, "mode");
119 if (pg_mode == NULL) {
120 return 1; // did not ask, get shadow
122 if (strcasecmp(pg_mode, "nested") == 0) {
123 extern v3_cpu_arch_t v3_mach_type;
124 if ((v3_mach_type == V3_SVM_REV3_CPU) ||
125 (v3_mach_type == V3_VMX_EPT_CPU) ||
126 (v3_mach_type == V3_VMX_EPT_UG_CPU)) {
127 return 0; // ask for nested, get nested
129 return 1; // ask for nested, get shadow
131 } else if (strcasecmp(pg_mode, "shadow") != 0) {
132 return 1; // ask for shadow, get shadow
134 return 1; // ask for something else, get shadow
140 int v3_init_mem_map(struct v3_vm_info * vm) {
141 struct v3_mem_map * map = &(vm->mem_map);
142 addr_t block_pages = v3_mem_block_size >> 12;
145 map->num_base_regions = (vm->mem_size / v3_mem_block_size) + \
146 ((vm->mem_size % v3_mem_block_size) > 0);
149 map->mem_regions.rb_node = NULL;
151 map->base_regions = V3_Malloc(sizeof(struct v3_mem_region) * map->num_base_regions);
153 if (map->base_regions == NULL) {
154 PrintError(vm, VCORE_NONE, "Could not allocate base region array\n");
158 memset(map->base_regions, 0, sizeof(struct v3_mem_region) * map->num_base_regions);
161 for (i = 0; i < map->num_base_regions; i++) {
162 struct v3_mem_region * region = &(map->base_regions[i]);
165 // 2MB page alignment needed for 2MB hardware nested paging
166 region->guest_start = v3_mem_block_size * i;
167 region->guest_end = region->guest_start + v3_mem_block_size;
169 // We assume that the xml config was smart enough to align the layout to the block size
170 // If they didn't we're going to ignore their settings
171 // and use whatever node the first byte of the block is assigned to
172 node_id = gpa_to_node_from_cfg(vm, region->guest_start);
174 V3_Print(vm, VCORE_NONE, "Allocating block %d on node %d\n", i, node_id);
176 region->host_addr = (addr_t)V3_AllocPagesExtended(block_pages,
179 0); // no constraints
181 if ((void *)region->host_addr == NULL) {
182 PrintError(vm, VCORE_NONE, "Could not allocate guest memory\n");
186 // Clear the memory...
187 memset(V3_VAddr((void *)region->host_addr), 0, v3_mem_block_size);
189 // Note assigned numa ID could be different than our request...
190 region->numa_id = v3_numa_hpa_to_node(region->host_addr);
192 region->flags.read = 1;
193 region->flags.write = 1;
194 region->flags.exec = 1;
195 region->flags.base = 1;
196 region->flags.alloced = 1;
197 region->flags.limit32 = will_use_shadow_paging(vm);
200 region->unhandled = unhandled_err;
203 v3_register_hypercall(vm, MEM_OFFSET_HCALL, mem_offset_hypercall, NULL);
209 void v3_delete_mem_map(struct v3_vm_info * vm) {
210 struct v3_mem_map * map = &(vm->mem_map);
211 struct rb_node * node = v3_rb_first(&(map->mem_regions));
212 struct v3_mem_region * reg;
213 struct rb_node * tmp_node = NULL;
214 addr_t block_pages = v3_mem_block_size >> 12;
218 reg = rb_entry(node, struct v3_mem_region, tree_node);
220 node = v3_rb_next(node);
222 v3_delete_mem_region(vm, reg);
225 for (i = 0; i < map->num_base_regions; i++) {
226 struct v3_mem_region * region = &(map->base_regions[i]);
227 V3_FreePages((void *)(region->host_addr), block_pages);
230 V3_Free(map->base_regions);
235 struct v3_mem_region * v3_create_mem_region(struct v3_vm_info * vm, uint16_t core_id,
236 addr_t guest_addr_start, addr_t guest_addr_end) {
237 struct v3_mem_region * entry = NULL;
239 if (guest_addr_start >= guest_addr_end) {
240 PrintError(vm, VCORE_NONE, "Region start is after region end\n");
244 entry = (struct v3_mem_region *)V3_Malloc(sizeof(struct v3_mem_region));
247 PrintError(vm, VCORE_NONE, "Cannot allocate in creating a memory region\n");
251 memset(entry, 0, sizeof(struct v3_mem_region));
253 entry->guest_start = guest_addr_start;
254 entry->guest_end = guest_addr_end;
255 entry->core_id = core_id;
256 entry->unhandled = unhandled_err;
264 int v3_add_shadow_mem( struct v3_vm_info * vm, uint16_t core_id,
265 addr_t guest_addr_start,
266 addr_t guest_addr_end,
269 struct v3_mem_region * entry = NULL;
271 entry = v3_create_mem_region(vm, core_id,
275 entry->host_addr = host_addr;
277 entry->flags.read = 1;
278 entry->flags.write = 1;
279 entry->flags.exec = 1;
280 entry->flags.alloced = 1;
282 if (v3_insert_mem_region(vm, entry) == -1) {
293 struct v3_mem_region * __insert_mem_region(struct v3_vm_info * vm,
294 struct v3_mem_region * region) {
295 struct rb_node ** p = &(vm->mem_map.mem_regions.rb_node);
296 struct rb_node * parent = NULL;
297 struct v3_mem_region * tmp_region;
301 tmp_region = rb_entry(parent, struct v3_mem_region, tree_node);
303 if (region->guest_end <= tmp_region->guest_start) {
305 } else if (region->guest_start >= tmp_region->guest_end) {
308 if ((region->guest_end != tmp_region->guest_end) ||
309 (region->guest_start != tmp_region->guest_start)) {
310 PrintError(vm, VCORE_NONE, "Trying to map a partial overlapped core specific page...\n");
311 return tmp_region; // This is ugly...
312 } else if (region->core_id == tmp_region->core_id) {
313 PrintError(vm, VCORE_NONE, "Trying to map a core-overlapping page\n");
315 } else if (region->core_id < tmp_region->core_id) {
323 rb_link_node(&(region->tree_node), parent, p);
330 int v3_insert_mem_region(struct v3_vm_info * vm, struct v3_mem_region * region) {
331 struct v3_mem_region * ret;
335 if ((ret = __insert_mem_region(vm, region))) {
336 PrintError(vm, VCORE_NONE, "Internal insert failed returned region is from 0x%p to 0x%p on vcore %d\n", (void*)(ret->guest_start), (void*)(ret->guest_end), ret->core_id);
340 v3_rb_insert_color(&(region->tree_node), &(vm->mem_map.mem_regions));
345 for (i = 0; i < vm->num_cores; i++) {
346 struct guest_info * info = &(vm->cores[i]);
348 // flush virtual page tables
349 // 3 cases shadow, shadow passthrough, and nested
351 if (info->shdw_pg_mode == SHADOW_PAGING) {
352 v3_mem_mode_t mem_mode = v3_get_vm_mem_mode(info);
354 if (mem_mode == PHYSICAL_MEM) {
355 rc |= v3_invalidate_passthrough_addr_range(info, region->guest_start, region->guest_end-1);
357 rc |= v3_invalidate_shadow_pts(info);
360 } else if (info->shdw_pg_mode == NESTED_PAGING) {
361 rc |= v3_invalidate_nested_addr_range(info, region->guest_start, region->guest_end-1);
371 struct v3_mem_region * v3_get_mem_region(struct v3_vm_info * vm, uint16_t core_id, addr_t guest_addr) {
372 struct rb_node * n = vm->mem_map.mem_regions.rb_node;
373 struct v3_mem_region * reg = NULL;
377 reg = rb_entry(n, struct v3_mem_region, tree_node);
379 if (guest_addr < reg->guest_start) {
381 } else if (guest_addr >= reg->guest_end) {
384 if (reg->core_id == V3_MEM_CORE_ANY) {
385 // found relevant region, it's available on all cores
387 } else if (core_id == reg->core_id) {
388 // found relevant region, it's available on the indicated core
390 } else if (core_id < reg->core_id) {
391 // go left, core too big
393 } else if (core_id > reg->core_id) {
394 // go right, core too small
397 PrintDebug(vm, VCORE_NONE, "v3_get_mem_region: Impossible!\n");
404 // There is not registered region, so we check if its a valid address in the base region
406 return v3_get_base_region(vm, guest_addr);
411 /* This returns the next memory region based on a given address.
412 * If the address falls inside a sub region, that region is returned.
413 * If the address falls outside a sub region, the next sub region is returned
414 * NOTE that we have to be careful about core_ids here...
416 static struct v3_mem_region * get_next_mem_region( struct v3_vm_info * vm, uint16_t core_id, addr_t guest_addr) {
417 struct rb_node * n = vm->mem_map.mem_regions.rb_node;
418 struct v3_mem_region * reg = NULL;
419 struct v3_mem_region * parent = NULL;
427 reg = rb_entry(n, struct v3_mem_region, tree_node);
429 if (guest_addr < reg->guest_start) {
431 } else if (guest_addr >= reg->guest_end) {
434 if (reg->core_id == V3_MEM_CORE_ANY) {
435 // found relevant region, it's available on all cores
437 } else if (core_id == reg->core_id) {
438 // found relevant region, it's available on the indicated core
440 } else if (core_id < reg->core_id) {
441 // go left, core too big
443 } else if (core_id > reg->core_id) {
444 // go right, core too small
447 PrintError(vm, VCORE_NONE, "v3_get_mem_region: Impossible!\n");
452 if ((reg->core_id == core_id) || (reg->core_id == V3_MEM_CORE_ANY)) {
458 if (parent->guest_start > guest_addr) {
460 } else if (parent->guest_end < guest_addr) {
461 struct rb_node * node = &(parent->tree_node);
463 while ((node = v3_rb_next(node)) != NULL) {
464 struct v3_mem_region * next_reg = rb_entry(node, struct v3_mem_region, tree_node);
466 if ((next_reg->core_id == V3_MEM_CORE_ANY) ||
467 (next_reg->core_id == core_id)) {
469 // This check is not strictly necessary, but it makes it clearer
470 if (next_reg->guest_start > guest_addr) {
483 /* Given an address region of memory, find if there are any regions that overlap with it.
484 * This checks that the range lies in a single region, and returns that region if it does,
485 * this can be either the base region or a sub region.
486 * IF there are multiple regions in the range then it returns NULL
488 static struct v3_mem_region * get_overlapping_region(struct v3_vm_info * vm, uint16_t core_id,
489 addr_t start_gpa, addr_t end_gpa) {
490 struct v3_mem_region * start_region = v3_get_mem_region(vm, core_id, start_gpa);
492 if (start_region == NULL) {
493 PrintError(vm, VCORE_NONE, "No overlapping region for core=%d, start_gpa=%p\n", core_id, (void*)start_gpa);
494 v3_print_mem_map(vm);
499 if (start_region->guest_end < end_gpa) {
500 // Region ends before range
502 } else if (start_region->flags.base == 0) {
503 // sub region overlaps range
506 // Base region, now we have to scan forward for the next sub region
507 struct v3_mem_region * next_reg = get_next_mem_region(vm, core_id, start_gpa);
509 if (next_reg == NULL) {
510 // no sub regions after start_addr, base region is ok
512 } else if (next_reg->guest_start >= end_gpa) {
513 // Next sub region begins outside range
521 // Should never get here
529 void v3_delete_mem_region(struct v3_vm_info * vm, struct v3_mem_region * reg) {
538 v3_rb_erase(&(reg->tree_node), &(vm->mem_map.mem_regions));
542 // If the guest isn't running then there shouldn't be anything to invalidate.
543 // Page tables should __always__ be created on demand during execution
544 // NOTE: This is a sanity check, and can be removed if that assumption changes
545 if (vm->run_state != VM_RUNNING) {
552 for (i = 0; i < vm->num_cores; i++) {
553 struct guest_info * info = &(vm->cores[i]);
555 // flush virtual page tables
556 // 3 cases shadow, shadow passthrough, and nested
558 if (info->shdw_pg_mode == SHADOW_PAGING) {
559 v3_mem_mode_t mem_mode = v3_get_vm_mem_mode(info);
561 if (mem_mode == PHYSICAL_MEM) {
562 rc |= v3_invalidate_passthrough_addr_range(info,reg->guest_start, reg->guest_end-1);
564 rc |= v3_invalidate_shadow_pts(info);
567 } else if (info->shdw_pg_mode == NESTED_PAGING) {
568 rc |= v3_invalidate_nested_addr_range(info,reg->guest_start, reg->guest_end-1);
574 // flush virtual page tables
575 // 3 cases shadow, shadow passthrough, and nested
577 if (rc) { PrintError(vm, VCORE_NONE, "Error in deleting memory region\n"); }
580 // Determine if a given address can be handled by a large page of the requested size
581 uint32_t v3_get_max_page_size(struct guest_info * core, addr_t page_addr, v3_cpu_mode_t mode) {
584 uint32_t page_size = PAGE_SIZE_4KB;
585 struct v3_mem_region * reg = NULL;
589 if (core->use_large_pages == 1) {
590 pg_start = PAGE_ADDR_4MB(page_addr);
591 pg_end = (pg_start + PAGE_SIZE_4MB);
593 reg = get_overlapping_region(core->vm_info, core->vcpu_id, pg_start, pg_end);
595 if ((reg) && ((reg->host_addr % PAGE_SIZE_4MB) == 0)) {
596 page_size = PAGE_SIZE_4MB;
601 if (core->use_large_pages == 1) {
602 pg_start = PAGE_ADDR_2MB(page_addr);
603 pg_end = (pg_start + PAGE_SIZE_2MB);
605 reg = get_overlapping_region(core->vm_info, core->vcpu_id, pg_start, pg_end);
607 if ((reg) && ((reg->host_addr % PAGE_SIZE_2MB) == 0)) {
608 page_size = PAGE_SIZE_2MB;
615 if (core->use_giant_pages == 1) {
616 pg_start = PAGE_ADDR_1GB(page_addr);
617 pg_end = (pg_start + PAGE_SIZE_1GB);
619 reg = get_overlapping_region(core->vm_info, core->vcpu_id, pg_start, pg_end);
621 if ((reg) && ((reg->host_addr % PAGE_SIZE_1GB) == 0)) {
622 page_size = PAGE_SIZE_1GB;
627 if (core->use_large_pages == 1) {
628 pg_start = PAGE_ADDR_2MB(page_addr);
629 pg_end = (pg_start + PAGE_SIZE_2MB);
631 reg = get_overlapping_region(core->vm_info, core->vcpu_id, pg_start, pg_end);
633 if ((reg) && ((reg->host_addr % PAGE_SIZE_2MB) == 0)) {
634 page_size = PAGE_SIZE_2MB;
639 PrintError(core->vm_info, core, "Invalid CPU mode: %s\n", v3_cpu_mode_to_str(v3_get_vm_cpu_mode(core)));
648 void v3_print_mem_map(struct v3_vm_info * vm) {
649 struct v3_mem_map * map = &(vm->mem_map);
650 struct rb_node * node = v3_rb_first(&(vm->mem_map.mem_regions));
651 struct v3_mem_region * reg = NULL;
654 V3_Print(vm, VCORE_NONE, "Memory Layout (all cores):\n");
656 V3_Print(vm, VCORE_NONE, "Base Memory: (%d regions)\n", map->num_base_regions);
658 for (i = 0; i < map->num_base_regions; i++) {
659 reg = &(map->base_regions[i]);
661 V3_Print(vm, VCORE_NONE, "Base Region[%d] (all cores): 0x%p - 0x%p -> 0x%p\n",
663 (void *)(reg->guest_start),
664 (void *)(reg->guest_end - 1),
665 (void *)(reg->host_addr));
669 // If the memory map is empty, don't print it
675 reg = rb_entry(node, struct v3_mem_region, tree_node);
677 V3_Print(vm, VCORE_NONE, "%d: 0x%p - 0x%p -> 0x%p\n", i,
678 (void *)(reg->guest_start),
679 (void *)(reg->guest_end - 1),
680 (void *)(reg->host_addr));
682 V3_Print(vm, VCORE_NONE, "\t(flags=0x%x) (core=0x%x) (unhandled = 0x%p)\n",
688 } while ((node = v3_rb_next(node)));
694 char *arg = v3_lookup_option("mem_block_size");
697 v3_mem_block_size = atoi(arg);
698 V3_Print(VM_NONE,VCORE_NONE,"memory block size set to %llu bytes\n",v3_mem_block_size);
700 V3_Print(VM_NONE,VCORE_NONE,"default memory block size of %llu bytes is in use\n",v3_mem_block_size);