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 #ifdef V3_CONFIG_SWAPPING
33 #include <palacios/vmm_swapping.h>
36 uint64_t v3_mem_block_size = V3_CONFIG_MEM_BLOCK_SIZE;
41 struct v3_mem_region * v3_get_base_region(struct v3_vm_info * vm, addr_t gpa) {
43 //PrintDebug(VM_NONE, VCORE_NONE, "get_base_region called");
44 struct v3_mem_map * map = &(vm->mem_map);
45 uint32_t block_index = gpa / v3_mem_block_size;
46 struct v3_mem_region *reg;
47 if ((gpa >= (map->num_base_regions * v3_mem_block_size)) ||
48 (block_index >= map->num_base_regions)) {
49 PrintError(vm, VCORE_NONE, "Guest Address Exceeds Base Memory Size (ga=0x%p), (limit=0x%p)\n",
50 (void *)gpa, (void *)vm->mem_size);
56 reg = &(map->base_regions[block_index]);
58 #ifdef V3_CONFIG_SWAPPING
59 if(vm->swap_state.enable_swapping) {
60 if (reg->flags.swapped) {
61 if (v3_swap_in_region(vm,reg)) {
62 PrintError(vm, VCORE_NONE, "Unable to swap in region GPA=%p..%p!!!\n",(void*)reg->guest_start,(void*)reg->guest_end);
68 v3_touch_region(vm,reg);
76 static int mem_offset_hypercall(struct guest_info * info, uint_t hcall_id, void * private_data) {
78 PrintDebug(info->vm_info, info,"V3Vee: Memory offset hypercall (offset=%p)\n",
79 (void *)(info->vm_info->mem_map.base_region.host_addr));
81 info->vm_regs.rbx = info->vm_info->mem_map.base_region.host_addr;
86 static int unhandled_err(struct guest_info * core, addr_t guest_va, addr_t guest_pa,
87 struct v3_mem_region * reg, pf_error_t access_info) {
89 PrintError(core->vm_info, core, "Unhandled memory access error (gpa=%p, gva=%p, error_code=%d)\n",
90 (void *)guest_pa, (void *)guest_va, *(uint32_t *)&access_info);
92 v3_print_mem_map(core->vm_info);
94 v3_print_guest_state(core);
99 static int gpa_to_node_from_cfg(struct v3_vm_info * vm, addr_t gpa) {
100 v3_cfg_tree_t * layout_cfg = v3_cfg_subtree(vm->cfg_data->cfg, "mem_layout");
101 v3_cfg_tree_t * region_desc = v3_cfg_subtree(layout_cfg, "region");
103 while (region_desc) {
104 char * start_addr_str = v3_cfg_val(region_desc, "start_addr");
105 char * end_addr_str = v3_cfg_val(region_desc, "end_addr");
106 char * node_id_str = v3_cfg_val(region_desc, "node");
108 addr_t start_addr = 0;
112 if ((!start_addr_str) || (!end_addr_str) || (!node_id_str)) {
113 PrintError(vm, VCORE_NONE, "Invalid memory layout in configuration\n");
117 start_addr = atox(start_addr_str);
118 end_addr = atox(end_addr_str);
119 node_id = atoi(node_id_str);
121 if ((gpa >= start_addr) && (gpa < end_addr)) {
125 region_desc = v3_cfg_next_branch(region_desc);
132 // This code parallels that in vmm_shadow_paging.c:v3_init_shdw_impl()
133 // and vmm_config.c:determine_paging_mode. The determination of which
134 // paging mode will be used is determined much later than the allocation of
135 // the guest memory regions, so we need to do this here to decide if they
136 // need to be below 4 GB or not.
137 static int will_use_shadow_paging(struct v3_vm_info *vm)
139 v3_cfg_tree_t * pg_cfg = v3_cfg_subtree(vm->cfg_data->cfg, "paging");
140 char * pg_mode = v3_cfg_val(pg_cfg, "mode");
142 if (pg_mode == NULL) {
143 return 1; // did not ask, get shadow
145 if (strcasecmp(pg_mode, "nested") == 0) {
146 extern v3_cpu_arch_t v3_mach_type;
147 if ((v3_mach_type == V3_SVM_REV3_CPU) ||
148 (v3_mach_type == V3_VMX_EPT_CPU) ||
149 (v3_mach_type == V3_VMX_EPT_UG_CPU)) {
150 return 0; // ask for nested, get nested
152 return 1; // ask for nested, get shadow
154 } else if (strcasecmp(pg_mode, "shadow") != 0) {
155 return 1; // ask for shadow, get shadow
157 return 1; // ask for something else, get shadow
162 #define CEIL_DIV(x,y) (((x)/(y)) + !!((x)%(y)))
165 int v3_init_mem_map(struct v3_vm_info * vm) {
166 struct v3_mem_map * map = &(vm->mem_map);
167 addr_t block_pages = v3_mem_block_size >> 12;
169 uint64_t num_base_regions_host_mem;
171 map->num_base_regions = CEIL_DIV(vm->mem_size, v3_mem_block_size);
173 num_base_regions_host_mem=map->num_base_regions; // without swapping
175 PrintDebug(VM_NONE, VCORE_NONE, "v3_init_mem_map: num_base_regions:%d",map->num_base_regions);
177 map->mem_regions.rb_node = NULL;
179 #ifdef V3_CONFIG_SWAPPING
180 if (vm->swap_state.enable_swapping) {
181 num_base_regions_host_mem = CEIL_DIV(vm->swap_state.host_mem_size, v3_mem_block_size);
185 PrintDebug(VM_NONE, VCORE_NONE, "v3_init_mem_map: %llu base regions will be allocated of %llu base regions in guest\n",
186 (uint64_t)num_base_regions_host_mem, (uint64_t)map->num_base_regions);
188 map->base_regions = V3_AllocPages(CEIL_DIV(sizeof(struct v3_mem_region) * map->num_base_regions, PAGE_SIZE_4KB));
189 if (map->base_regions == NULL) {
190 PrintError(vm, VCORE_NONE, "Could not allocate base region array\n");
193 map->base_regions = V3_VAddr(map->base_regions);
195 memset(map->base_regions, 0, sizeof(struct v3_mem_region) * map->num_base_regions);
197 for (i = 0; i < map->num_base_regions; i++) {
200 struct v3_mem_region * region = &(map->base_regions[i]);
203 // 2MB page alignment needed for 2MB hardware nested paging
204 // If swapping is enabled, the host memory will be allocated to low address regions at initialization
205 region->guest_start = v3_mem_block_size * i;
206 region->guest_end = region->guest_start + v3_mem_block_size;
208 // We assume that the xml config was smart enough to align the layout to the block size
209 // If they didn't we're going to ignore their settings
210 // and use whatever node the first byte of the block is assigned to
211 node_id = gpa_to_node_from_cfg(vm, region->guest_start);
214 if (i < num_base_regions_host_mem) {
215 //The regions within num_base_regions_in_mem are allocated in host memory
216 V3_Print(vm, VCORE_NONE, "Allocating block %d on node %d\n", i, node_id);
218 #ifdef V3_CONFIG_SWAPPING
219 // nothing to do - memset will have done it.
222 region->host_addr = (addr_t)V3_AllocPagesExtended(block_pages,
225 0); // no constraints
227 if ((void *)region->host_addr == NULL) {
228 PrintError(vm, VCORE_NONE, "Could not allocate guest memory\n");
232 // Clear the memory...
233 memset(V3_VAddr((void *)region->host_addr), 0, v3_mem_block_size);
237 #ifdef V3_CONFIG_SWAPPING
238 if(vm->swap_state.enable_swapping) {
239 // The regions beyond num_base_regions_in_mem are allocated on disk to start
240 region->flags.swapped = 1;
241 region->host_addr=(addr_t) 0;
242 // other flags / state correctly set up by zeroing the region earlier
249 // Note assigned numa ID could be different than our request...
250 // Also note that when swapping is used, the numa info will
251 // reflect the numa id of address 0x0 for unallocated regions
253 region->numa_id = v3_numa_hpa_to_node(region->host_addr);
255 region->flags.read = 1;
256 region->flags.write = 1;
257 region->flags.exec = 1;
258 region->flags.base = 1;
259 region->flags.alloced = 1;
260 region->flags.limit32 = will_use_shadow_paging(vm);
262 region->unhandled = unhandled_err;
265 v3_register_hypercall(vm, MEM_OFFSET_HCALL, mem_offset_hypercall, NULL);
271 void v3_delete_mem_map(struct v3_vm_info * vm) {
272 struct v3_mem_map * map = &(vm->mem_map);
273 struct rb_node * node = v3_rb_first(&(map->mem_regions));
274 struct v3_mem_region * reg;
275 struct rb_node * tmp_node = NULL;
276 addr_t block_pages = v3_mem_block_size >> 12;
280 reg = rb_entry(node, struct v3_mem_region, tree_node);
282 node = v3_rb_next(node);
284 v3_delete_mem_region(vm, reg);
287 for (i = 0; i < map->num_base_regions; i++) {
288 struct v3_mem_region * region = &(map->base_regions[i]);
289 #ifdef V3_CONFIG_SWAPPING
290 if (vm->swap_state.enable_swapping) {
291 if (!region->flags.swapped) {
292 V3_FreePages((void *)(region->host_addr), block_pages);
293 } // otherwise this is not allocated space
296 V3_FreePages((void *)(region->host_addr), block_pages);
300 V3_FreePages(V3_PAddr(map->base_regions),
301 CEIL_DIV(sizeof(struct v3_mem_region) * map->num_base_regions, PAGE_SIZE_4KB));
305 struct v3_mem_region * v3_create_mem_region(struct v3_vm_info * vm, uint16_t core_id,
306 addr_t guest_addr_start, addr_t guest_addr_end) {
307 struct v3_mem_region * entry = NULL;
309 if (guest_addr_start >= guest_addr_end) {
310 PrintError(vm, VCORE_NONE, "Region start is after region end\n");
314 entry = (struct v3_mem_region *)V3_Malloc(sizeof(struct v3_mem_region));
317 PrintError(vm, VCORE_NONE, "Cannot allocate in creating a memory region\n");
321 memset(entry, 0, sizeof(struct v3_mem_region));
323 entry->guest_start = guest_addr_start;
324 entry->guest_end = guest_addr_end;
325 entry->core_id = core_id;
326 entry->unhandled = unhandled_err;
334 int v3_add_shadow_mem( struct v3_vm_info * vm, uint16_t core_id,
335 addr_t guest_addr_start,
336 addr_t guest_addr_end,
339 struct v3_mem_region * entry = NULL;
341 entry = v3_create_mem_region(vm, core_id,
345 entry->host_addr = host_addr;
347 entry->flags.read = 1;
348 entry->flags.write = 1;
349 entry->flags.exec = 1;
350 entry->flags.alloced = 1;
352 if (v3_insert_mem_region(vm, entry) == -1) {
363 struct v3_mem_region * __insert_mem_region(struct v3_vm_info * vm,
364 struct v3_mem_region * region) {
365 struct rb_node ** p = &(vm->mem_map.mem_regions.rb_node);
366 struct rb_node * parent = NULL;
367 struct v3_mem_region * tmp_region;
371 tmp_region = rb_entry(parent, struct v3_mem_region, tree_node);
373 if (region->guest_end <= tmp_region->guest_start) {
375 } else if (region->guest_start >= tmp_region->guest_end) {
378 if ((region->guest_end != tmp_region->guest_end) ||
379 (region->guest_start != tmp_region->guest_start)) {
380 PrintError(vm, VCORE_NONE, "Trying to map a partial overlapped core specific page...\n");
381 return tmp_region; // This is ugly...
382 } else if (region->core_id == tmp_region->core_id) {
383 PrintError(vm, VCORE_NONE, "Trying to map a core-overlapping page\n");
385 } else if (region->core_id < tmp_region->core_id) {
393 rb_link_node(&(region->tree_node), parent, p);
400 int v3_insert_mem_region(struct v3_vm_info * vm, struct v3_mem_region * region) {
401 struct v3_mem_region * ret;
405 if ((ret = __insert_mem_region(vm, region))) {
406 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);
410 v3_rb_insert_color(&(region->tree_node), &(vm->mem_map.mem_regions));
415 for (i = 0; i < vm->num_cores; i++) {
416 struct guest_info * info = &(vm->cores[i]);
418 // flush virtual page tables
419 // 3 cases shadow, shadow passthrough, and nested
421 if (info->shdw_pg_mode == SHADOW_PAGING) {
422 v3_mem_mode_t mem_mode = v3_get_vm_mem_mode(info);
424 if (mem_mode == PHYSICAL_MEM) {
425 rc |= v3_invalidate_passthrough_addr_range(info, region->guest_start, region->guest_end-1,NULL,NULL);
427 rc |= v3_invalidate_shadow_pts(info);
430 } else if (info->shdw_pg_mode == NESTED_PAGING) {
431 rc |= v3_invalidate_nested_addr_range(info, region->guest_start, region->guest_end-1,NULL,NULL);
441 struct v3_mem_region * v3_get_mem_region(struct v3_vm_info * vm, uint16_t core_id, addr_t guest_addr) {
442 struct rb_node * n = vm->mem_map.mem_regions.rb_node;
443 struct v3_mem_region * reg = NULL;
447 reg = rb_entry(n, struct v3_mem_region, tree_node);
449 if (guest_addr < reg->guest_start) {
451 } else if (guest_addr >= reg->guest_end) {
454 if (reg->core_id == V3_MEM_CORE_ANY) {
455 // found relevant region, it's available on all cores
457 } else if (core_id == reg->core_id) {
458 // found relevant region, it's available on the indicated core
460 } else if (core_id < reg->core_id) {
461 // go left, core too big
463 } else if (core_id > reg->core_id) {
464 // go right, core too small
467 PrintDebug(vm, VCORE_NONE, "v3_get_mem_region: Impossible!\n");
474 // There is not registered region, so we check if its a valid address in the base region
476 return v3_get_base_region(vm, guest_addr);
481 /* This returns the next memory region based on a given address.
482 * If the address falls inside a sub region, that region is returned.
483 * If the address falls outside a sub region, the next sub region is returned
484 * NOTE that we have to be careful about core_ids here...
486 static struct v3_mem_region * get_next_mem_region( struct v3_vm_info * vm, uint16_t core_id, addr_t guest_addr) {
487 struct rb_node * n = vm->mem_map.mem_regions.rb_node;
488 struct v3_mem_region * reg = NULL;
489 struct v3_mem_region * parent = NULL;
497 reg = rb_entry(n, struct v3_mem_region, tree_node);
499 if (guest_addr < reg->guest_start) {
501 } else if (guest_addr >= reg->guest_end) {
504 if (reg->core_id == V3_MEM_CORE_ANY) {
505 // found relevant region, it's available on all cores
507 } else if (core_id == reg->core_id) {
508 // found relevant region, it's available on the indicated core
510 } else if (core_id < reg->core_id) {
511 // go left, core too big
513 } else if (core_id > reg->core_id) {
514 // go right, core too small
517 PrintError(vm, VCORE_NONE, "v3_get_mem_region: Impossible!\n");
522 if ((reg->core_id == core_id) || (reg->core_id == V3_MEM_CORE_ANY)) {
528 if (parent->guest_start > guest_addr) {
530 } else if (parent->guest_end < guest_addr) {
531 struct rb_node * node = &(parent->tree_node);
533 while ((node = v3_rb_next(node)) != NULL) {
534 struct v3_mem_region * next_reg = rb_entry(node, struct v3_mem_region, tree_node);
536 if ((next_reg->core_id == V3_MEM_CORE_ANY) ||
537 (next_reg->core_id == core_id)) {
539 // This check is not strictly necessary, but it makes it clearer
540 if (next_reg->guest_start > guest_addr) {
553 /* Given an address region of memory, find if there are any regions that overlap with it.
554 * This checks that the range lies in a single region, and returns that region if it does,
555 * this can be either the base region or a sub region.
556 * IF there are multiple regions in the range then it returns NULL
558 static struct v3_mem_region * get_overlapping_region(struct v3_vm_info * vm, uint16_t core_id,
559 addr_t start_gpa, addr_t end_gpa) {
560 struct v3_mem_region * start_region = v3_get_mem_region(vm, core_id, start_gpa);
562 if (start_region == NULL) {
563 PrintError(vm, VCORE_NONE, "No overlapping region for core=%d, start_gpa=%p\n", core_id, (void*)start_gpa);
564 v3_print_mem_map(vm);
569 if (start_region->guest_end < end_gpa) {
570 // Region ends before range
572 } else if (start_region->flags.base == 0) {
573 // sub region overlaps range
576 // Base region, now we have to scan forward for the next sub region
577 struct v3_mem_region * next_reg = get_next_mem_region(vm, core_id, start_gpa);
579 if (next_reg == NULL) {
580 // no sub regions after start_addr, base region is ok
582 } else if (next_reg->guest_start >= end_gpa) {
583 // Next sub region begins outside range
591 // Should never get here
599 void v3_delete_mem_region(struct v3_vm_info * vm, struct v3_mem_region * reg) {
608 v3_rb_erase(&(reg->tree_node), &(vm->mem_map.mem_regions));
612 // If the guest isn't running then there shouldn't be anything to invalidate.
613 // Page tables should __always__ be created on demand during execution
614 // NOTE: This is a sanity check, and can be removed if that assumption changes
615 if (vm->run_state != VM_RUNNING) {
622 for (i = 0; i < vm->num_cores; i++) {
623 struct guest_info * info = &(vm->cores[i]);
625 // flush virtual page tables
626 // 3 cases shadow, shadow passthrough, and nested
628 if (info->shdw_pg_mode == SHADOW_PAGING) {
629 v3_mem_mode_t mem_mode = v3_get_vm_mem_mode(info);
631 if (mem_mode == PHYSICAL_MEM) {
632 rc |= v3_invalidate_passthrough_addr_range(info,reg->guest_start, reg->guest_end-1,NULL,NULL);
634 rc |= v3_invalidate_shadow_pts(info);
637 } else if (info->shdw_pg_mode == NESTED_PAGING) {
638 rc |= v3_invalidate_nested_addr_range(info,reg->guest_start, reg->guest_end-1,NULL,NULL);
644 // flush virtual page tables
645 // 3 cases shadow, shadow passthrough, and nested
647 if (rc) { PrintError(vm, VCORE_NONE, "Error in deleting memory region\n"); }
650 // Determine if a given address can be handled by a large page of the requested size
651 uint32_t v3_get_max_page_size(struct guest_info * core, addr_t page_addr, v3_cpu_mode_t mode) {
654 uint32_t page_size = PAGE_SIZE_4KB;
655 struct v3_mem_region * reg = NULL;
659 if (core->use_large_pages == 1) {
660 pg_start = PAGE_ADDR_4MB(page_addr);
661 pg_end = (pg_start + PAGE_SIZE_4MB);
663 reg = get_overlapping_region(core->vm_info, core->vcpu_id, pg_start, pg_end);
665 if ((reg) && ((reg->host_addr % PAGE_SIZE_4MB) == 0)) {
666 page_size = PAGE_SIZE_4MB;
671 if (core->use_large_pages == 1) {
672 pg_start = PAGE_ADDR_2MB(page_addr);
673 pg_end = (pg_start + PAGE_SIZE_2MB);
675 reg = get_overlapping_region(core->vm_info, core->vcpu_id, pg_start, pg_end);
677 if ((reg) && ((reg->host_addr % PAGE_SIZE_2MB) == 0)) {
678 page_size = PAGE_SIZE_2MB;
685 if (core->use_giant_pages == 1) {
686 pg_start = PAGE_ADDR_1GB(page_addr);
687 pg_end = (pg_start + PAGE_SIZE_1GB);
689 reg = get_overlapping_region(core->vm_info, core->vcpu_id, pg_start, pg_end);
691 if ((reg) && ((reg->host_addr % PAGE_SIZE_1GB) == 0)) {
692 page_size = PAGE_SIZE_1GB;
697 if (core->use_large_pages == 1) {
698 pg_start = PAGE_ADDR_2MB(page_addr);
699 pg_end = (pg_start + PAGE_SIZE_2MB);
701 reg = get_overlapping_region(core->vm_info, core->vcpu_id, pg_start, pg_end);
703 if ((reg) && ((reg->host_addr % PAGE_SIZE_2MB) == 0)) {
704 page_size = PAGE_SIZE_2MB;
709 PrintError(core->vm_info, core, "Invalid CPU mode: %s\n", v3_cpu_mode_to_str(v3_get_vm_cpu_mode(core)));
718 void v3_print_mem_map(struct v3_vm_info * vm) {
719 struct v3_mem_map * map = &(vm->mem_map);
720 struct rb_node * node = v3_rb_first(&(vm->mem_map.mem_regions));
721 struct v3_mem_region * reg = NULL;
724 V3_Print(vm, VCORE_NONE, "Memory Layout (all cores):\n");
726 V3_Print(vm, VCORE_NONE, "Base Memory: (%d regions)\n", map->num_base_regions);
728 for (i = 0; i < map->num_base_regions; i++) {
729 reg = &(map->base_regions[i]);
731 V3_Print(vm, VCORE_NONE, "Base Region[%d] (all cores): 0x%p - 0x%p -> 0x%p\n",
733 (void *)(reg->guest_start),
734 (void *)(reg->guest_end - 1),
735 (void *)(reg->host_addr));
739 // If the memory map is empty, don't print it
745 reg = rb_entry(node, struct v3_mem_region, tree_node);
747 V3_Print(vm, VCORE_NONE, "%d: 0x%p - 0x%p -> 0x%p\n", i,
748 (void *)(reg->guest_start),
749 (void *)(reg->guest_end - 1),
750 (void *)(reg->host_addr));
752 V3_Print(vm, VCORE_NONE, "\t(flags=0x%x) (core=0x%x) (unhandled = 0x%p)\n",
758 } while ((node = v3_rb_next(node)));
764 char *arg = v3_lookup_option("mem_block_size");
767 v3_mem_block_size = atoi(arg);
768 V3_Print(VM_NONE,VCORE_NONE,"memory block size set to %llu bytes\n",v3_mem_block_size);
770 V3_Print(VM_NONE,VCORE_NONE,"default memory block size of %llu bytes is in use\n",v3_mem_block_size);