4 * Copyright (c) 2003 Fabrice Bellard
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, see <http://www.gnu.org/licenses/>.
19 #include "qemu/osdep.h"
20 #include "qapi/error.h"
24 #include "qemu/cutils.h"
26 #include "exec/exec-all.h"
27 #include "exec/target_page.h"
29 #include "hw/qdev-core.h"
30 #if !defined(CONFIG_USER_ONLY)
31 #include "hw/boards.h"
32 #include "hw/xen/xen.h"
34 #include "sysemu/kvm.h"
35 #include "sysemu/sysemu.h"
36 #include "qemu/timer.h"
37 #include "qemu/config-file.h"
38 #include "qemu/error-report.h"
39 #if defined(CONFIG_USER_ONLY)
41 #else /* !CONFIG_USER_ONLY */
43 #include "exec/memory.h"
44 #include "exec/ioport.h"
45 #include "sysemu/dma.h"
46 #include "sysemu/numa.h"
47 #include "sysemu/hw_accel.h"
48 #include "exec/address-spaces.h"
49 #include "sysemu/xen-mapcache.h"
50 #include "trace-root.h"
52 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
54 #include <linux/falloc.h>
58 #include "exec/cpu-all.h"
59 #include "qemu/rcu_queue.h"
60 #include "qemu/main-loop.h"
61 #include "translate-all.h"
62 #include "sysemu/replay.h"
64 #include "exec/memory-internal.h"
65 #include "exec/ram_addr.h"
68 #include "migration/vmstate.h"
70 #include "qemu/range.h"
72 #include "qemu/mmap-alloc.h"
75 #include "monitor/monitor.h"
77 //#define DEBUG_SUBPAGE
79 #if !defined(CONFIG_USER_ONLY)
80 /* ram_list is read under rcu_read_lock()/rcu_read_unlock(). Writes
81 * are protected by the ramlist lock.
83 RAMList ram_list = { .blocks = QLIST_HEAD_INITIALIZER(ram_list.blocks) };
85 static MemoryRegion *system_memory;
86 static MemoryRegion *system_io;
88 AddressSpace address_space_io;
89 AddressSpace address_space_memory;
91 MemoryRegion io_mem_rom, io_mem_notdirty;
92 static MemoryRegion io_mem_unassigned;
94 /* RAM is pre-allocated and passed into qemu_ram_alloc_from_ptr */
95 #define RAM_PREALLOC (1 << 0)
97 /* RAM is mmap-ed with MAP_SHARED */
98 #define RAM_SHARED (1 << 1)
100 /* Only a portion of RAM (used_length) is actually used, and migrated.
101 * This used_length size can change across reboots.
103 #define RAM_RESIZEABLE (1 << 2)
107 #ifdef TARGET_PAGE_BITS_VARY
108 int target_page_bits;
109 bool target_page_bits_decided;
112 struct CPUTailQ cpus = QTAILQ_HEAD_INITIALIZER(cpus);
113 /* current CPU in the current thread. It is only valid inside
115 __thread CPUState *current_cpu;
116 /* 0 = Do not count executed instructions.
117 1 = Precise instruction counting.
118 2 = Adaptive rate instruction counting. */
121 bool set_preferred_target_page_bits(int bits)
123 /* The target page size is the lowest common denominator for all
124 * the CPUs in the system, so we can only make it smaller, never
125 * larger. And we can't make it smaller once we've committed to
128 #ifdef TARGET_PAGE_BITS_VARY
129 assert(bits >= TARGET_PAGE_BITS_MIN);
130 if (target_page_bits == 0 || target_page_bits > bits) {
131 if (target_page_bits_decided) {
134 target_page_bits = bits;
140 #if !defined(CONFIG_USER_ONLY)
142 static void finalize_target_page_bits(void)
144 #ifdef TARGET_PAGE_BITS_VARY
145 if (target_page_bits == 0) {
146 target_page_bits = TARGET_PAGE_BITS_MIN;
148 target_page_bits_decided = true;
152 typedef struct PhysPageEntry PhysPageEntry;
154 struct PhysPageEntry {
155 /* How many bits skip to next level (in units of L2_SIZE). 0 for a leaf. */
157 /* index into phys_sections (!skip) or phys_map_nodes (skip) */
161 #define PHYS_MAP_NODE_NIL (((uint32_t)~0) >> 6)
163 /* Size of the L2 (and L3, etc) page tables. */
164 #define ADDR_SPACE_BITS 64
167 #define P_L2_SIZE (1 << P_L2_BITS)
169 #define P_L2_LEVELS (((ADDR_SPACE_BITS - TARGET_PAGE_BITS - 1) / P_L2_BITS) + 1)
171 typedef PhysPageEntry Node[P_L2_SIZE];
173 typedef struct PhysPageMap {
176 unsigned sections_nb;
177 unsigned sections_nb_alloc;
179 unsigned nodes_nb_alloc;
181 MemoryRegionSection *sections;
184 struct AddressSpaceDispatch {
187 MemoryRegionSection *mru_section;
188 /* This is a multi-level map on the physical address space.
189 * The bottom level has pointers to MemoryRegionSections.
191 PhysPageEntry phys_map;
196 #define SUBPAGE_IDX(addr) ((addr) & ~TARGET_PAGE_MASK)
197 typedef struct subpage_t {
201 uint16_t sub_section[];
204 #define PHYS_SECTION_UNASSIGNED 0
205 #define PHYS_SECTION_NOTDIRTY 1
206 #define PHYS_SECTION_ROM 2
207 #define PHYS_SECTION_WATCH 3
209 static void io_mem_init(void);
210 static void memory_map_init(void);
211 static void tcg_commit(MemoryListener *listener);
213 static MemoryRegion io_mem_watch;
216 * CPUAddressSpace: all the information a CPU needs about an AddressSpace
217 * @cpu: the CPU whose AddressSpace this is
218 * @as: the AddressSpace itself
219 * @memory_dispatch: its dispatch pointer (cached, RCU protected)
220 * @tcg_as_listener: listener for tracking changes to the AddressSpace
222 struct CPUAddressSpace {
225 struct AddressSpaceDispatch *memory_dispatch;
226 MemoryListener tcg_as_listener;
229 struct DirtyBitmapSnapshot {
232 unsigned long dirty[];
237 #if !defined(CONFIG_USER_ONLY)
239 static void phys_map_node_reserve(PhysPageMap *map, unsigned nodes)
241 static unsigned alloc_hint = 16;
242 if (map->nodes_nb + nodes > map->nodes_nb_alloc) {
243 map->nodes_nb_alloc = MAX(map->nodes_nb_alloc, alloc_hint);
244 map->nodes_nb_alloc = MAX(map->nodes_nb_alloc, map->nodes_nb + nodes);
245 map->nodes = g_renew(Node, map->nodes, map->nodes_nb_alloc);
246 alloc_hint = map->nodes_nb_alloc;
250 static uint32_t phys_map_node_alloc(PhysPageMap *map, bool leaf)
257 ret = map->nodes_nb++;
259 assert(ret != PHYS_MAP_NODE_NIL);
260 assert(ret != map->nodes_nb_alloc);
262 e.skip = leaf ? 0 : 1;
263 e.ptr = leaf ? PHYS_SECTION_UNASSIGNED : PHYS_MAP_NODE_NIL;
264 for (i = 0; i < P_L2_SIZE; ++i) {
265 memcpy(&p[i], &e, sizeof(e));
270 static void phys_page_set_level(PhysPageMap *map, PhysPageEntry *lp,
271 hwaddr *index, hwaddr *nb, uint16_t leaf,
275 hwaddr step = (hwaddr)1 << (level * P_L2_BITS);
277 if (lp->skip && lp->ptr == PHYS_MAP_NODE_NIL) {
278 lp->ptr = phys_map_node_alloc(map, level == 0);
280 p = map->nodes[lp->ptr];
281 lp = &p[(*index >> (level * P_L2_BITS)) & (P_L2_SIZE - 1)];
283 while (*nb && lp < &p[P_L2_SIZE]) {
284 if ((*index & (step - 1)) == 0 && *nb >= step) {
290 phys_page_set_level(map, lp, index, nb, leaf, level - 1);
296 static void phys_page_set(AddressSpaceDispatch *d,
297 hwaddr index, hwaddr nb,
300 /* Wildly overreserve - it doesn't matter much. */
301 phys_map_node_reserve(&d->map, 3 * P_L2_LEVELS);
303 phys_page_set_level(&d->map, &d->phys_map, &index, &nb, leaf, P_L2_LEVELS - 1);
306 /* Compact a non leaf page entry. Simply detect that the entry has a single child,
307 * and update our entry so we can skip it and go directly to the destination.
309 static void phys_page_compact(PhysPageEntry *lp, Node *nodes)
311 unsigned valid_ptr = P_L2_SIZE;
316 if (lp->ptr == PHYS_MAP_NODE_NIL) {
321 for (i = 0; i < P_L2_SIZE; i++) {
322 if (p[i].ptr == PHYS_MAP_NODE_NIL) {
329 phys_page_compact(&p[i], nodes);
333 /* We can only compress if there's only one child. */
338 assert(valid_ptr < P_L2_SIZE);
340 /* Don't compress if it won't fit in the # of bits we have. */
341 if (lp->skip + p[valid_ptr].skip >= (1 << 3)) {
345 lp->ptr = p[valid_ptr].ptr;
346 if (!p[valid_ptr].skip) {
347 /* If our only child is a leaf, make this a leaf. */
348 /* By design, we should have made this node a leaf to begin with so we
349 * should never reach here.
350 * But since it's so simple to handle this, let's do it just in case we
355 lp->skip += p[valid_ptr].skip;
359 static void phys_page_compact_all(AddressSpaceDispatch *d, int nodes_nb)
361 if (d->phys_map.skip) {
362 phys_page_compact(&d->phys_map, d->map.nodes);
366 static inline bool section_covers_addr(const MemoryRegionSection *section,
369 /* Memory topology clips a memory region to [0, 2^64); size.hi > 0 means
370 * the section must cover the entire address space.
372 return int128_gethi(section->size) ||
373 range_covers_byte(section->offset_within_address_space,
374 int128_getlo(section->size), addr);
377 static MemoryRegionSection *phys_page_find(PhysPageEntry lp, hwaddr addr,
378 Node *nodes, MemoryRegionSection *sections)
381 hwaddr index = addr >> TARGET_PAGE_BITS;
384 for (i = P_L2_LEVELS; lp.skip && (i -= lp.skip) >= 0;) {
385 if (lp.ptr == PHYS_MAP_NODE_NIL) {
386 return §ions[PHYS_SECTION_UNASSIGNED];
389 lp = p[(index >> (i * P_L2_BITS)) & (P_L2_SIZE - 1)];
392 if (section_covers_addr(§ions[lp.ptr], addr)) {
393 return §ions[lp.ptr];
395 return §ions[PHYS_SECTION_UNASSIGNED];
399 bool memory_region_is_unassigned(MemoryRegion *mr)
401 return mr != &io_mem_rom && mr != &io_mem_notdirty && !mr->rom_device
402 && mr != &io_mem_watch;
405 /* Called from RCU critical section */
406 static MemoryRegionSection *address_space_lookup_region(AddressSpaceDispatch *d,
408 bool resolve_subpage)
410 MemoryRegionSection *section = atomic_read(&d->mru_section);
414 if (section && section != &d->map.sections[PHYS_SECTION_UNASSIGNED] &&
415 section_covers_addr(section, addr)) {
418 section = phys_page_find(d->phys_map, addr, d->map.nodes,
422 if (resolve_subpage && section->mr->subpage) {
423 subpage = container_of(section->mr, subpage_t, iomem);
424 section = &d->map.sections[subpage->sub_section[SUBPAGE_IDX(addr)]];
427 atomic_set(&d->mru_section, section);
432 /* Called from RCU critical section */
433 static MemoryRegionSection *
434 address_space_translate_internal(AddressSpaceDispatch *d, hwaddr addr, hwaddr *xlat,
435 hwaddr *plen, bool resolve_subpage)
437 MemoryRegionSection *section;
441 section = address_space_lookup_region(d, addr, resolve_subpage);
442 /* Compute offset within MemoryRegionSection */
443 addr -= section->offset_within_address_space;
445 /* Compute offset within MemoryRegion */
446 *xlat = addr + section->offset_within_region;
450 /* MMIO registers can be expected to perform full-width accesses based only
451 * on their address, without considering adjacent registers that could
452 * decode to completely different MemoryRegions. When such registers
453 * exist (e.g. I/O ports 0xcf8 and 0xcf9 on most PC chipsets), MMIO
454 * regions overlap wildly. For this reason we cannot clamp the accesses
457 * If the length is small (as is the case for address_space_ldl/stl),
458 * everything works fine. If the incoming length is large, however,
459 * the caller really has to do the clamping through memory_access_size.
461 if (memory_region_is_ram(mr)) {
462 diff = int128_sub(section->size, int128_make64(addr));
463 *plen = int128_get64(int128_min(diff, int128_make64(*plen)));
468 /* Called from RCU critical section */
469 static MemoryRegionSection address_space_do_translate(AddressSpace *as,
477 MemoryRegionSection *section;
481 AddressSpaceDispatch *d = atomic_rcu_read(&as->dispatch);
482 section = address_space_translate_internal(d, addr, &addr, plen, is_mmio);
485 if (!mr->iommu_ops) {
489 iotlb = mr->iommu_ops->translate(mr, addr, is_write ?
490 IOMMU_WO : IOMMU_RO);
491 addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
492 | (addr & iotlb.addr_mask));
493 *plen = MIN(*plen, (addr | iotlb.addr_mask) - addr + 1);
494 if (!(iotlb.perm & (1 << is_write))) {
498 as = iotlb.target_as;
506 return (MemoryRegionSection) { .mr = &io_mem_unassigned };
509 /* Called from RCU critical section */
510 IOMMUTLBEntry address_space_get_iotlb_entry(AddressSpace *as, hwaddr addr,
513 MemoryRegionSection section;
516 /* Try to get maximum page mask during translation. */
519 /* This can never be MMIO. */
520 section = address_space_do_translate(as, addr, &xlat, &plen,
523 /* Illegal translation */
524 if (section.mr == &io_mem_unassigned) {
528 /* Convert memory region offset into address space offset */
529 xlat += section.offset_within_address_space -
530 section.offset_within_region;
532 if (plen == (hwaddr)-1) {
534 * We use default page size here. Logically it only happens
535 * for identity mappings.
537 plen = TARGET_PAGE_SIZE;
540 /* Convert to address mask */
543 return (IOMMUTLBEntry) {
544 .target_as = section.address_space,
545 .iova = addr & ~plen,
546 .translated_addr = xlat & ~plen,
548 /* IOTLBs are for DMAs, and DMA only allows on RAMs. */
553 return (IOMMUTLBEntry) {0};
556 /* Called from RCU critical section */
557 MemoryRegion *address_space_translate(AddressSpace *as, hwaddr addr,
558 hwaddr *xlat, hwaddr *plen,
562 MemoryRegionSection section;
564 /* This can be MMIO, so setup MMIO bit. */
565 section = address_space_do_translate(as, addr, xlat, plen, is_write, true);
568 if (xen_enabled() && memory_access_is_direct(mr, is_write)) {
569 hwaddr page = ((addr & TARGET_PAGE_MASK) + TARGET_PAGE_SIZE) - addr;
570 *plen = MIN(page, *plen);
576 /* Called from RCU critical section */
577 MemoryRegionSection *
578 address_space_translate_for_iotlb(CPUState *cpu, int asidx, hwaddr addr,
579 hwaddr *xlat, hwaddr *plen)
581 MemoryRegionSection *section;
582 AddressSpaceDispatch *d = atomic_rcu_read(&cpu->cpu_ases[asidx].memory_dispatch);
584 section = address_space_translate_internal(d, addr, xlat, plen, false);
586 assert(!section->mr->iommu_ops);
591 #if !defined(CONFIG_USER_ONLY)
593 static int cpu_common_post_load(void *opaque, int version_id)
595 CPUState *cpu = opaque;
597 /* 0x01 was CPU_INTERRUPT_EXIT. This line can be removed when the
598 version_id is increased. */
599 cpu->interrupt_request &= ~0x01;
605 static int cpu_common_pre_load(void *opaque)
607 CPUState *cpu = opaque;
609 cpu->exception_index = -1;
614 static bool cpu_common_exception_index_needed(void *opaque)
616 CPUState *cpu = opaque;
618 return tcg_enabled() && cpu->exception_index != -1;
621 static const VMStateDescription vmstate_cpu_common_exception_index = {
622 .name = "cpu_common/exception_index",
624 .minimum_version_id = 1,
625 .needed = cpu_common_exception_index_needed,
626 .fields = (VMStateField[]) {
627 VMSTATE_INT32(exception_index, CPUState),
628 VMSTATE_END_OF_LIST()
632 static bool cpu_common_crash_occurred_needed(void *opaque)
634 CPUState *cpu = opaque;
636 return cpu->crash_occurred;
639 static const VMStateDescription vmstate_cpu_common_crash_occurred = {
640 .name = "cpu_common/crash_occurred",
642 .minimum_version_id = 1,
643 .needed = cpu_common_crash_occurred_needed,
644 .fields = (VMStateField[]) {
645 VMSTATE_BOOL(crash_occurred, CPUState),
646 VMSTATE_END_OF_LIST()
650 const VMStateDescription vmstate_cpu_common = {
651 .name = "cpu_common",
653 .minimum_version_id = 1,
654 .pre_load = cpu_common_pre_load,
655 .post_load = cpu_common_post_load,
656 .fields = (VMStateField[]) {
657 VMSTATE_UINT32(halted, CPUState),
658 VMSTATE_UINT32(interrupt_request, CPUState),
659 VMSTATE_END_OF_LIST()
661 .subsections = (const VMStateDescription*[]) {
662 &vmstate_cpu_common_exception_index,
663 &vmstate_cpu_common_crash_occurred,
670 CPUState *qemu_get_cpu(int index)
675 if (cpu->cpu_index == index) {
683 #if !defined(CONFIG_USER_ONLY)
684 void cpu_address_space_init(CPUState *cpu, AddressSpace *as, int asidx)
686 CPUAddressSpace *newas;
688 /* Target code should have set num_ases before calling us */
689 assert(asidx < cpu->num_ases);
692 /* address space 0 gets the convenience alias */
696 /* KVM cannot currently support multiple address spaces. */
697 assert(asidx == 0 || !kvm_enabled());
699 if (!cpu->cpu_ases) {
700 cpu->cpu_ases = g_new0(CPUAddressSpace, cpu->num_ases);
703 newas = &cpu->cpu_ases[asidx];
707 newas->tcg_as_listener.commit = tcg_commit;
708 memory_listener_register(&newas->tcg_as_listener, as);
712 AddressSpace *cpu_get_address_space(CPUState *cpu, int asidx)
714 /* Return the AddressSpace corresponding to the specified index */
715 return cpu->cpu_ases[asidx].as;
719 void cpu_exec_unrealizefn(CPUState *cpu)
721 CPUClass *cc = CPU_GET_CLASS(cpu);
723 cpu_list_remove(cpu);
725 if (cc->vmsd != NULL) {
726 vmstate_unregister(NULL, cc->vmsd, cpu);
728 if (qdev_get_vmsd(DEVICE(cpu)) == NULL) {
729 vmstate_unregister(NULL, &vmstate_cpu_common, cpu);
733 void cpu_exec_initfn(CPUState *cpu)
738 #ifndef CONFIG_USER_ONLY
739 cpu->thread_id = qemu_get_thread_id();
741 /* This is a softmmu CPU object, so create a property for it
742 * so users can wire up its memory. (This can't go in qom/cpu.c
743 * because that file is compiled only once for both user-mode
744 * and system builds.) The default if no link is set up is to use
745 * the system address space.
747 object_property_add_link(OBJECT(cpu), "memory", TYPE_MEMORY_REGION,
748 (Object **)&cpu->memory,
749 qdev_prop_allow_set_link_before_realize,
750 OBJ_PROP_LINK_UNREF_ON_RELEASE,
752 cpu->memory = system_memory;
753 object_ref(OBJECT(cpu->memory));
757 void cpu_exec_realizefn(CPUState *cpu, Error **errp)
759 CPUClass *cc ATTRIBUTE_UNUSED = CPU_GET_CLASS(cpu);
763 #ifndef CONFIG_USER_ONLY
764 if (qdev_get_vmsd(DEVICE(cpu)) == NULL) {
765 vmstate_register(NULL, cpu->cpu_index, &vmstate_cpu_common, cpu);
767 if (cc->vmsd != NULL) {
768 vmstate_register(NULL, cpu->cpu_index, cc->vmsd, cpu);
773 static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
775 /* Flush the whole TB as this will not have race conditions
776 * even if we don't have proper locking yet.
777 * Ideally we would just invalidate the TBs for the
783 #if defined(CONFIG_USER_ONLY)
784 void cpu_watchpoint_remove_all(CPUState *cpu, int mask)
789 int cpu_watchpoint_remove(CPUState *cpu, vaddr addr, vaddr len,
795 void cpu_watchpoint_remove_by_ref(CPUState *cpu, CPUWatchpoint *watchpoint)
799 int cpu_watchpoint_insert(CPUState *cpu, vaddr addr, vaddr len,
800 int flags, CPUWatchpoint **watchpoint)
805 /* Add a watchpoint. */
806 int cpu_watchpoint_insert(CPUState *cpu, vaddr addr, vaddr len,
807 int flags, CPUWatchpoint **watchpoint)
811 /* forbid ranges which are empty or run off the end of the address space */
812 if (len == 0 || (addr + len - 1) < addr) {
813 error_report("tried to set invalid watchpoint at %"
814 VADDR_PRIx ", len=%" VADDR_PRIu, addr, len);
817 wp = g_malloc(sizeof(*wp));
823 /* keep all GDB-injected watchpoints in front */
824 if (flags & BP_GDB) {
825 QTAILQ_INSERT_HEAD(&cpu->watchpoints, wp, entry);
827 QTAILQ_INSERT_TAIL(&cpu->watchpoints, wp, entry);
830 tlb_flush_page(cpu, addr);
837 /* Remove a specific watchpoint. */
838 int cpu_watchpoint_remove(CPUState *cpu, vaddr addr, vaddr len,
843 QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
844 if (addr == wp->vaddr && len == wp->len
845 && flags == (wp->flags & ~BP_WATCHPOINT_HIT)) {
846 cpu_watchpoint_remove_by_ref(cpu, wp);
853 /* Remove a specific watchpoint by reference. */
854 void cpu_watchpoint_remove_by_ref(CPUState *cpu, CPUWatchpoint *watchpoint)
856 QTAILQ_REMOVE(&cpu->watchpoints, watchpoint, entry);
858 tlb_flush_page(cpu, watchpoint->vaddr);
863 /* Remove all matching watchpoints. */
864 void cpu_watchpoint_remove_all(CPUState *cpu, int mask)
866 CPUWatchpoint *wp, *next;
868 QTAILQ_FOREACH_SAFE(wp, &cpu->watchpoints, entry, next) {
869 if (wp->flags & mask) {
870 cpu_watchpoint_remove_by_ref(cpu, wp);
875 /* Return true if this watchpoint address matches the specified
876 * access (ie the address range covered by the watchpoint overlaps
877 * partially or completely with the address range covered by the
880 static inline bool cpu_watchpoint_address_matches(CPUWatchpoint *wp,
884 /* We know the lengths are non-zero, but a little caution is
885 * required to avoid errors in the case where the range ends
886 * exactly at the top of the address space and so addr + len
887 * wraps round to zero.
889 vaddr wpend = wp->vaddr + wp->len - 1;
890 vaddr addrend = addr + len - 1;
892 return !(addr > wpend || wp->vaddr > addrend);
897 /* Add a breakpoint. */
898 int cpu_breakpoint_insert(CPUState *cpu, vaddr pc, int flags,
899 CPUBreakpoint **breakpoint)
903 bp = g_malloc(sizeof(*bp));
908 /* keep all GDB-injected breakpoints in front */
909 if (flags & BP_GDB) {
910 QTAILQ_INSERT_HEAD(&cpu->breakpoints, bp, entry);
912 QTAILQ_INSERT_TAIL(&cpu->breakpoints, bp, entry);
915 breakpoint_invalidate(cpu, pc);
923 /* Remove a specific breakpoint. */
924 int cpu_breakpoint_remove(CPUState *cpu, vaddr pc, int flags)
928 QTAILQ_FOREACH(bp, &cpu->breakpoints, entry) {
929 if (bp->pc == pc && bp->flags == flags) {
930 cpu_breakpoint_remove_by_ref(cpu, bp);
937 /* Remove a specific breakpoint by reference. */
938 void cpu_breakpoint_remove_by_ref(CPUState *cpu, CPUBreakpoint *breakpoint)
940 QTAILQ_REMOVE(&cpu->breakpoints, breakpoint, entry);
942 breakpoint_invalidate(cpu, breakpoint->pc);
947 /* Remove all matching breakpoints. */
948 void cpu_breakpoint_remove_all(CPUState *cpu, int mask)
950 CPUBreakpoint *bp, *next;
952 QTAILQ_FOREACH_SAFE(bp, &cpu->breakpoints, entry, next) {
953 if (bp->flags & mask) {
954 cpu_breakpoint_remove_by_ref(cpu, bp);
959 /* enable or disable single step mode. EXCP_DEBUG is returned by the
960 CPU loop after each instruction */
961 void cpu_single_step(CPUState *cpu, int enabled)
963 if (cpu->singlestep_enabled != enabled) {
964 cpu->singlestep_enabled = enabled;
966 kvm_update_guest_debug(cpu, 0);
968 /* must flush all the translated code to avoid inconsistencies */
969 /* XXX: only flush what is necessary */
975 void cpu_abort(CPUState *cpu, const char *fmt, ...)
982 fprintf(stderr, "qemu: fatal: ");
983 vfprintf(stderr, fmt, ap);
984 fprintf(stderr, "\n");
985 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_FPU | CPU_DUMP_CCOP);
986 if (qemu_log_separate()) {
988 qemu_log("qemu: fatal: ");
989 qemu_log_vprintf(fmt, ap2);
991 log_cpu_state(cpu, CPU_DUMP_FPU | CPU_DUMP_CCOP);
999 #if defined(CONFIG_USER_ONLY)
1001 struct sigaction act;
1002 sigfillset(&act.sa_mask);
1003 act.sa_handler = SIG_DFL;
1004 sigaction(SIGABRT, &act, NULL);
1010 #if !defined(CONFIG_USER_ONLY)
1011 /* Called from RCU critical section */
1012 static RAMBlock *qemu_get_ram_block(ram_addr_t addr)
1016 block = atomic_rcu_read(&ram_list.mru_block);
1017 if (block && addr - block->offset < block->max_length) {
1020 RAMBLOCK_FOREACH(block) {
1021 if (addr - block->offset < block->max_length) {
1026 fprintf(stderr, "Bad ram offset %" PRIx64 "\n", (uint64_t)addr);
1030 /* It is safe to write mru_block outside the iothread lock. This
1035 * xxx removed from list
1039 * call_rcu(reclaim_ramblock, xxx);
1042 * atomic_rcu_set is not needed here. The block was already published
1043 * when it was placed into the list. Here we're just making an extra
1044 * copy of the pointer.
1046 ram_list.mru_block = block;
1050 static void tlb_reset_dirty_range_all(ram_addr_t start, ram_addr_t length)
1057 end = TARGET_PAGE_ALIGN(start + length);
1058 start &= TARGET_PAGE_MASK;
1061 block = qemu_get_ram_block(start);
1062 assert(block == qemu_get_ram_block(end - 1));
1063 start1 = (uintptr_t)ramblock_ptr(block, start - block->offset);
1065 tlb_reset_dirty(cpu, start1, length);
1070 /* Note: start and end must be within the same ram block. */
1071 bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start,
1075 DirtyMemoryBlocks *blocks;
1076 unsigned long end, page;
1083 end = TARGET_PAGE_ALIGN(start + length) >> TARGET_PAGE_BITS;
1084 page = start >> TARGET_PAGE_BITS;
1088 blocks = atomic_rcu_read(&ram_list.dirty_memory[client]);
1090 while (page < end) {
1091 unsigned long idx = page / DIRTY_MEMORY_BLOCK_SIZE;
1092 unsigned long offset = page % DIRTY_MEMORY_BLOCK_SIZE;
1093 unsigned long num = MIN(end - page, DIRTY_MEMORY_BLOCK_SIZE - offset);
1095 dirty |= bitmap_test_and_clear_atomic(blocks->blocks[idx],
1102 if (dirty && tcg_enabled()) {
1103 tlb_reset_dirty_range_all(start, length);
1109 DirtyBitmapSnapshot *cpu_physical_memory_snapshot_and_clear_dirty
1110 (ram_addr_t start, ram_addr_t length, unsigned client)
1112 DirtyMemoryBlocks *blocks;
1113 unsigned long align = 1UL << (TARGET_PAGE_BITS + BITS_PER_LEVEL);
1114 ram_addr_t first = QEMU_ALIGN_DOWN(start, align);
1115 ram_addr_t last = QEMU_ALIGN_UP(start + length, align);
1116 DirtyBitmapSnapshot *snap;
1117 unsigned long page, end, dest;
1119 snap = g_malloc0(sizeof(*snap) +
1120 ((last - first) >> (TARGET_PAGE_BITS + 3)));
1121 snap->start = first;
1124 page = first >> TARGET_PAGE_BITS;
1125 end = last >> TARGET_PAGE_BITS;
1130 blocks = atomic_rcu_read(&ram_list.dirty_memory[client]);
1132 while (page < end) {
1133 unsigned long idx = page / DIRTY_MEMORY_BLOCK_SIZE;
1134 unsigned long offset = page % DIRTY_MEMORY_BLOCK_SIZE;
1135 unsigned long num = MIN(end - page, DIRTY_MEMORY_BLOCK_SIZE - offset);
1137 assert(QEMU_IS_ALIGNED(offset, (1 << BITS_PER_LEVEL)));
1138 assert(QEMU_IS_ALIGNED(num, (1 << BITS_PER_LEVEL)));
1139 offset >>= BITS_PER_LEVEL;
1141 bitmap_copy_and_clear_atomic(snap->dirty + dest,
1142 blocks->blocks[idx] + offset,
1145 dest += num >> BITS_PER_LEVEL;
1150 if (tcg_enabled()) {
1151 tlb_reset_dirty_range_all(start, length);
1157 bool cpu_physical_memory_snapshot_get_dirty(DirtyBitmapSnapshot *snap,
1161 unsigned long page, end;
1163 assert(start >= snap->start);
1164 assert(start + length <= snap->end);
1166 end = TARGET_PAGE_ALIGN(start + length - snap->start) >> TARGET_PAGE_BITS;
1167 page = (start - snap->start) >> TARGET_PAGE_BITS;
1169 while (page < end) {
1170 if (test_bit(page, snap->dirty)) {
1178 /* Called from RCU critical section */
1179 hwaddr memory_region_section_get_iotlb(CPUState *cpu,
1180 MemoryRegionSection *section,
1182 hwaddr paddr, hwaddr xlat,
1184 target_ulong *address)
1189 if (memory_region_is_ram(section->mr)) {
1191 iotlb = memory_region_get_ram_addr(section->mr) + xlat;
1192 if (!section->readonly) {
1193 iotlb |= PHYS_SECTION_NOTDIRTY;
1195 iotlb |= PHYS_SECTION_ROM;
1198 AddressSpaceDispatch *d;
1200 d = atomic_rcu_read(§ion->address_space->dispatch);
1201 iotlb = section - d->map.sections;
1205 /* Make accesses to pages with watchpoints go via the
1206 watchpoint trap routines. */
1207 QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
1208 if (cpu_watchpoint_address_matches(wp, vaddr, TARGET_PAGE_SIZE)) {
1209 /* Avoid trapping reads of pages with a write breakpoint. */
1210 if ((prot & PAGE_WRITE) || (wp->flags & BP_MEM_READ)) {
1211 iotlb = PHYS_SECTION_WATCH + paddr;
1212 *address |= TLB_MMIO;
1220 #endif /* defined(CONFIG_USER_ONLY) */
1222 #if !defined(CONFIG_USER_ONLY)
1224 static int subpage_register (subpage_t *mmio, uint32_t start, uint32_t end,
1226 static subpage_t *subpage_init(AddressSpace *as, hwaddr base);
1228 static void *(*phys_mem_alloc)(size_t size, uint64_t *align) =
1229 qemu_anon_ram_alloc;
1232 * Set a custom physical guest memory alloator.
1233 * Accelerators with unusual needs may need this. Hopefully, we can
1234 * get rid of it eventually.
1236 void phys_mem_set_alloc(void *(*alloc)(size_t, uint64_t *align))
1238 phys_mem_alloc = alloc;
1241 static uint16_t phys_section_add(PhysPageMap *map,
1242 MemoryRegionSection *section)
1244 /* The physical section number is ORed with a page-aligned
1245 * pointer to produce the iotlb entries. Thus it should
1246 * never overflow into the page-aligned value.
1248 assert(map->sections_nb < TARGET_PAGE_SIZE);
1250 if (map->sections_nb == map->sections_nb_alloc) {
1251 map->sections_nb_alloc = MAX(map->sections_nb_alloc * 2, 16);
1252 map->sections = g_renew(MemoryRegionSection, map->sections,
1253 map->sections_nb_alloc);
1255 map->sections[map->sections_nb] = *section;
1256 memory_region_ref(section->mr);
1257 return map->sections_nb++;
1260 static void phys_section_destroy(MemoryRegion *mr)
1262 bool have_sub_page = mr->subpage;
1264 memory_region_unref(mr);
1266 if (have_sub_page) {
1267 subpage_t *subpage = container_of(mr, subpage_t, iomem);
1268 object_unref(OBJECT(&subpage->iomem));
1273 static void phys_sections_free(PhysPageMap *map)
1275 while (map->sections_nb > 0) {
1276 MemoryRegionSection *section = &map->sections[--map->sections_nb];
1277 phys_section_destroy(section->mr);
1279 g_free(map->sections);
1283 static void register_subpage(AddressSpaceDispatch *d, MemoryRegionSection *section)
1286 hwaddr base = section->offset_within_address_space
1288 MemoryRegionSection *existing = phys_page_find(d->phys_map, base,
1289 d->map.nodes, d->map.sections);
1290 MemoryRegionSection subsection = {
1291 .offset_within_address_space = base,
1292 .size = int128_make64(TARGET_PAGE_SIZE),
1296 assert(existing->mr->subpage || existing->mr == &io_mem_unassigned);
1298 if (!(existing->mr->subpage)) {
1299 subpage = subpage_init(d->as, base);
1300 subsection.address_space = d->as;
1301 subsection.mr = &subpage->iomem;
1302 phys_page_set(d, base >> TARGET_PAGE_BITS, 1,
1303 phys_section_add(&d->map, &subsection));
1305 subpage = container_of(existing->mr, subpage_t, iomem);
1307 start = section->offset_within_address_space & ~TARGET_PAGE_MASK;
1308 end = start + int128_get64(section->size) - 1;
1309 subpage_register(subpage, start, end,
1310 phys_section_add(&d->map, section));
1314 static void register_multipage(AddressSpaceDispatch *d,
1315 MemoryRegionSection *section)
1317 hwaddr start_addr = section->offset_within_address_space;
1318 uint16_t section_index = phys_section_add(&d->map, section);
1319 uint64_t num_pages = int128_get64(int128_rshift(section->size,
1323 phys_page_set(d, start_addr >> TARGET_PAGE_BITS, num_pages, section_index);
1326 static void mem_add(MemoryListener *listener, MemoryRegionSection *section)
1328 AddressSpace *as = container_of(listener, AddressSpace, dispatch_listener);
1329 AddressSpaceDispatch *d = as->next_dispatch;
1330 MemoryRegionSection now = *section, remain = *section;
1331 Int128 page_size = int128_make64(TARGET_PAGE_SIZE);
1333 if (now.offset_within_address_space & ~TARGET_PAGE_MASK) {
1334 uint64_t left = TARGET_PAGE_ALIGN(now.offset_within_address_space)
1335 - now.offset_within_address_space;
1337 now.size = int128_min(int128_make64(left), now.size);
1338 register_subpage(d, &now);
1340 now.size = int128_zero();
1342 while (int128_ne(remain.size, now.size)) {
1343 remain.size = int128_sub(remain.size, now.size);
1344 remain.offset_within_address_space += int128_get64(now.size);
1345 remain.offset_within_region += int128_get64(now.size);
1347 if (int128_lt(remain.size, page_size)) {
1348 register_subpage(d, &now);
1349 } else if (remain.offset_within_address_space & ~TARGET_PAGE_MASK) {
1350 now.size = page_size;
1351 register_subpage(d, &now);
1353 now.size = int128_and(now.size, int128_neg(page_size));
1354 register_multipage(d, &now);
1359 void qemu_flush_coalesced_mmio_buffer(void)
1362 kvm_flush_coalesced_mmio_buffer();
1365 void qemu_mutex_lock_ramlist(void)
1367 qemu_mutex_lock(&ram_list.mutex);
1370 void qemu_mutex_unlock_ramlist(void)
1372 qemu_mutex_unlock(&ram_list.mutex);
1375 void ram_block_dump(Monitor *mon)
1381 monitor_printf(mon, "%24s %8s %18s %18s %18s\n",
1382 "Block Name", "PSize", "Offset", "Used", "Total");
1383 RAMBLOCK_FOREACH(block) {
1384 psize = size_to_str(block->page_size);
1385 monitor_printf(mon, "%24s %8s 0x%016" PRIx64 " 0x%016" PRIx64
1386 " 0x%016" PRIx64 "\n", block->idstr, psize,
1387 (uint64_t)block->offset,
1388 (uint64_t)block->used_length,
1389 (uint64_t)block->max_length);
1397 * FIXME TOCTTOU: this iterates over memory backends' mem-path, which
1398 * may or may not name the same files / on the same filesystem now as
1399 * when we actually open and map them. Iterate over the file
1400 * descriptors instead, and use qemu_fd_getpagesize().
1402 static int find_max_supported_pagesize(Object *obj, void *opaque)
1405 long *hpsize_min = opaque;
1407 if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
1408 mem_path = object_property_get_str(obj, "mem-path", NULL);
1410 long hpsize = qemu_mempath_getpagesize(mem_path);
1411 if (hpsize < *hpsize_min) {
1412 *hpsize_min = hpsize;
1415 *hpsize_min = getpagesize();
1422 long qemu_getrampagesize(void)
1424 long hpsize = LONG_MAX;
1425 long mainrampagesize;
1426 Object *memdev_root;
1429 mainrampagesize = qemu_mempath_getpagesize(mem_path);
1431 mainrampagesize = getpagesize();
1434 /* it's possible we have memory-backend objects with
1435 * hugepage-backed RAM. these may get mapped into system
1436 * address space via -numa parameters or memory hotplug
1437 * hooks. we want to take these into account, but we
1438 * also want to make sure these supported hugepage
1439 * sizes are applicable across the entire range of memory
1440 * we may boot from, so we take the min across all
1441 * backends, and assume normal pages in cases where a
1442 * backend isn't backed by hugepages.
1444 memdev_root = object_resolve_path("/objects", NULL);
1446 object_child_foreach(memdev_root, find_max_supported_pagesize, &hpsize);
1448 if (hpsize == LONG_MAX) {
1449 /* No additional memory regions found ==> Report main RAM page size */
1450 return mainrampagesize;
1453 /* If NUMA is disabled or the NUMA nodes are not backed with a
1454 * memory-backend, then there is at least one node using "normal" RAM,
1455 * so if its page size is smaller we have got to report that size instead.
1457 if (hpsize > mainrampagesize &&
1458 (nb_numa_nodes == 0 || numa_info[0].node_memdev == NULL)) {
1461 error_report("Huge page support disabled (n/a for main memory).");
1464 return mainrampagesize;
1470 long qemu_getrampagesize(void)
1472 return getpagesize();
1477 static int64_t get_file_size(int fd)
1479 int64_t size = lseek(fd, 0, SEEK_END);
1486 static void *file_ram_alloc(RAMBlock *block,
1491 bool unlink_on_error = false;
1493 char *sanitized_name;
1495 void *area = MAP_FAILED;
1499 if (kvm_enabled() && !kvm_has_sync_mmu()) {
1501 "host lacks kvm mmu notifiers, -mem-path unsupported");
1506 fd = open(path, O_RDWR);
1508 /* @path names an existing file, use it */
1511 if (errno == ENOENT) {
1512 /* @path names a file that doesn't exist, create it */
1513 fd = open(path, O_RDWR | O_CREAT | O_EXCL, 0644);
1515 unlink_on_error = true;
1518 } else if (errno == EISDIR) {
1519 /* @path names a directory, create a file there */
1520 /* Make name safe to use with mkstemp by replacing '/' with '_'. */
1521 sanitized_name = g_strdup(memory_region_name(block->mr));
1522 for (c = sanitized_name; *c != '\0'; c++) {
1528 filename = g_strdup_printf("%s/qemu_back_mem.%s.XXXXXX", path,
1530 g_free(sanitized_name);
1532 fd = mkstemp(filename);
1540 if (errno != EEXIST && errno != EINTR) {
1541 error_setg_errno(errp, errno,
1542 "can't open backing store %s for guest RAM",
1547 * Try again on EINTR and EEXIST. The latter happens when
1548 * something else creates the file between our two open().
1552 block->page_size = qemu_fd_getpagesize(fd);
1553 block->mr->align = block->page_size;
1554 #if defined(__s390x__)
1555 if (kvm_enabled()) {
1556 block->mr->align = MAX(block->mr->align, QEMU_VMALLOC_ALIGN);
1560 file_size = get_file_size(fd);
1562 if (memory < block->page_size) {
1563 error_setg(errp, "memory size 0x" RAM_ADDR_FMT " must be equal to "
1564 "or larger than page size 0x%zx",
1565 memory, block->page_size);
1569 if (file_size > 0 && file_size < memory) {
1570 error_setg(errp, "backing store %s size 0x%" PRIx64
1571 " does not match 'size' option 0x" RAM_ADDR_FMT,
1572 path, file_size, memory);
1576 memory = ROUND_UP(memory, block->page_size);
1579 * ftruncate is not supported by hugetlbfs in older
1580 * hosts, so don't bother bailing out on errors.
1581 * If anything goes wrong with it under other filesystems,
1584 * Do not truncate the non-empty backend file to avoid corrupting
1585 * the existing data in the file. Disabling shrinking is not
1586 * enough. For example, the current vNVDIMM implementation stores
1587 * the guest NVDIMM labels at the end of the backend file. If the
1588 * backend file is later extended, QEMU will not be able to find
1589 * those labels. Therefore, extending the non-empty backend file
1590 * is disabled as well.
1592 if (!file_size && ftruncate(fd, memory)) {
1593 perror("ftruncate");
1596 area = qemu_ram_mmap(fd, memory, block->mr->align,
1597 block->flags & RAM_SHARED);
1598 if (area == MAP_FAILED) {
1599 error_setg_errno(errp, errno,
1600 "unable to map backing store for guest RAM");
1605 os_mem_prealloc(fd, area, memory, smp_cpus, errp);
1606 if (errp && *errp) {
1615 if (area != MAP_FAILED) {
1616 qemu_ram_munmap(area, memory);
1618 if (unlink_on_error) {
1628 /* Called with the ramlist lock held. */
1629 static ram_addr_t find_ram_offset(ram_addr_t size)
1631 RAMBlock *block, *next_block;
1632 ram_addr_t offset = RAM_ADDR_MAX, mingap = RAM_ADDR_MAX;
1634 assert(size != 0); /* it would hand out same offset multiple times */
1636 if (QLIST_EMPTY_RCU(&ram_list.blocks)) {
1640 RAMBLOCK_FOREACH(block) {
1641 ram_addr_t end, next = RAM_ADDR_MAX;
1643 end = block->offset + block->max_length;
1645 RAMBLOCK_FOREACH(next_block) {
1646 if (next_block->offset >= end) {
1647 next = MIN(next, next_block->offset);
1650 if (next - end >= size && next - end < mingap) {
1652 mingap = next - end;
1656 if (offset == RAM_ADDR_MAX) {
1657 fprintf(stderr, "Failed to find gap of requested size: %" PRIu64 "\n",
1665 unsigned long last_ram_page(void)
1668 ram_addr_t last = 0;
1671 RAMBLOCK_FOREACH(block) {
1672 last = MAX(last, block->offset + block->max_length);
1675 return last >> TARGET_PAGE_BITS;
1678 static void qemu_ram_setup_dump(void *addr, ram_addr_t size)
1682 /* Use MADV_DONTDUMP, if user doesn't want the guest memory in the core */
1683 if (!machine_dump_guest_core(current_machine)) {
1684 ret = qemu_madvise(addr, size, QEMU_MADV_DONTDUMP);
1686 perror("qemu_madvise");
1687 fprintf(stderr, "madvise doesn't support MADV_DONTDUMP, "
1688 "but dump_guest_core=off specified\n");
1693 const char *qemu_ram_get_idstr(RAMBlock *rb)
1698 bool qemu_ram_is_shared(RAMBlock *rb)
1700 return rb->flags & RAM_SHARED;
1703 /* Called with iothread lock held. */
1704 void qemu_ram_set_idstr(RAMBlock *new_block, const char *name, DeviceState *dev)
1709 assert(!new_block->idstr[0]);
1712 char *id = qdev_get_dev_path(dev);
1714 snprintf(new_block->idstr, sizeof(new_block->idstr), "%s/", id);
1718 pstrcat(new_block->idstr, sizeof(new_block->idstr), name);
1721 RAMBLOCK_FOREACH(block) {
1722 if (block != new_block &&
1723 !strcmp(block->idstr, new_block->idstr)) {
1724 fprintf(stderr, "RAMBlock \"%s\" already registered, abort!\n",
1732 /* Called with iothread lock held. */
1733 void qemu_ram_unset_idstr(RAMBlock *block)
1735 /* FIXME: arch_init.c assumes that this is not called throughout
1736 * migration. Ignore the problem since hot-unplug during migration
1737 * does not work anyway.
1740 memset(block->idstr, 0, sizeof(block->idstr));
1744 size_t qemu_ram_pagesize(RAMBlock *rb)
1746 return rb->page_size;
1749 /* Returns the largest size of page in use */
1750 size_t qemu_ram_pagesize_largest(void)
1755 RAMBLOCK_FOREACH(block) {
1756 largest = MAX(largest, qemu_ram_pagesize(block));
1762 static int memory_try_enable_merging(void *addr, size_t len)
1764 if (!machine_mem_merge(current_machine)) {
1765 /* disabled by the user */
1769 return qemu_madvise(addr, len, QEMU_MADV_MERGEABLE);
1772 /* Only legal before guest might have detected the memory size: e.g. on
1773 * incoming migration, or right after reset.
1775 * As memory core doesn't know how is memory accessed, it is up to
1776 * resize callback to update device state and/or add assertions to detect
1777 * misuse, if necessary.
1779 int qemu_ram_resize(RAMBlock *block, ram_addr_t newsize, Error **errp)
1783 newsize = HOST_PAGE_ALIGN(newsize);
1785 if (block->used_length == newsize) {
1789 if (!(block->flags & RAM_RESIZEABLE)) {
1790 error_setg_errno(errp, EINVAL,
1791 "Length mismatch: %s: 0x" RAM_ADDR_FMT
1792 " in != 0x" RAM_ADDR_FMT, block->idstr,
1793 newsize, block->used_length);
1797 if (block->max_length < newsize) {
1798 error_setg_errno(errp, EINVAL,
1799 "Length too large: %s: 0x" RAM_ADDR_FMT
1800 " > 0x" RAM_ADDR_FMT, block->idstr,
1801 newsize, block->max_length);
1805 cpu_physical_memory_clear_dirty_range(block->offset, block->used_length);
1806 block->used_length = newsize;
1807 cpu_physical_memory_set_dirty_range(block->offset, block->used_length,
1809 memory_region_set_size(block->mr, newsize);
1810 if (block->resized) {
1811 block->resized(block->idstr, newsize, block->host);
1816 /* Called with ram_list.mutex held */
1817 static void dirty_memory_extend(ram_addr_t old_ram_size,
1818 ram_addr_t new_ram_size)
1820 ram_addr_t old_num_blocks = DIV_ROUND_UP(old_ram_size,
1821 DIRTY_MEMORY_BLOCK_SIZE);
1822 ram_addr_t new_num_blocks = DIV_ROUND_UP(new_ram_size,
1823 DIRTY_MEMORY_BLOCK_SIZE);
1826 /* Only need to extend if block count increased */
1827 if (new_num_blocks <= old_num_blocks) {
1831 for (i = 0; i < DIRTY_MEMORY_NUM; i++) {
1832 DirtyMemoryBlocks *old_blocks;
1833 DirtyMemoryBlocks *new_blocks;
1836 old_blocks = atomic_rcu_read(&ram_list.dirty_memory[i]);
1837 new_blocks = g_malloc(sizeof(*new_blocks) +
1838 sizeof(new_blocks->blocks[0]) * new_num_blocks);
1840 if (old_num_blocks) {
1841 memcpy(new_blocks->blocks, old_blocks->blocks,
1842 old_num_blocks * sizeof(old_blocks->blocks[0]));
1845 for (j = old_num_blocks; j < new_num_blocks; j++) {
1846 new_blocks->blocks[j] = bitmap_new(DIRTY_MEMORY_BLOCK_SIZE);
1849 atomic_rcu_set(&ram_list.dirty_memory[i], new_blocks);
1852 g_free_rcu(old_blocks, rcu);
1857 static void ram_block_add(RAMBlock *new_block, Error **errp)
1860 RAMBlock *last_block = NULL;
1861 ram_addr_t old_ram_size, new_ram_size;
1864 old_ram_size = last_ram_page();
1866 qemu_mutex_lock_ramlist();
1867 new_block->offset = find_ram_offset(new_block->max_length);
1869 if (!new_block->host) {
1870 if (xen_enabled()) {
1871 xen_ram_alloc(new_block->offset, new_block->max_length,
1872 new_block->mr, &err);
1874 error_propagate(errp, err);
1875 qemu_mutex_unlock_ramlist();
1879 new_block->host = phys_mem_alloc(new_block->max_length,
1880 &new_block->mr->align);
1881 if (!new_block->host) {
1882 error_setg_errno(errp, errno,
1883 "cannot set up guest memory '%s'",
1884 memory_region_name(new_block->mr));
1885 qemu_mutex_unlock_ramlist();
1888 memory_try_enable_merging(new_block->host, new_block->max_length);
1892 new_ram_size = MAX(old_ram_size,
1893 (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS);
1894 if (new_ram_size > old_ram_size) {
1895 dirty_memory_extend(old_ram_size, new_ram_size);
1897 /* Keep the list sorted from biggest to smallest block. Unlike QTAILQ,
1898 * QLIST (which has an RCU-friendly variant) does not have insertion at
1899 * tail, so save the last element in last_block.
1901 RAMBLOCK_FOREACH(block) {
1903 if (block->max_length < new_block->max_length) {
1908 QLIST_INSERT_BEFORE_RCU(block, new_block, next);
1909 } else if (last_block) {
1910 QLIST_INSERT_AFTER_RCU(last_block, new_block, next);
1911 } else { /* list is empty */
1912 QLIST_INSERT_HEAD_RCU(&ram_list.blocks, new_block, next);
1914 ram_list.mru_block = NULL;
1916 /* Write list before version */
1919 qemu_mutex_unlock_ramlist();
1921 cpu_physical_memory_set_dirty_range(new_block->offset,
1922 new_block->used_length,
1925 if (new_block->host) {
1926 qemu_ram_setup_dump(new_block->host, new_block->max_length);
1927 qemu_madvise(new_block->host, new_block->max_length, QEMU_MADV_HUGEPAGE);
1928 /* MADV_DONTFORK is also needed by KVM in absence of synchronous MMU */
1929 qemu_madvise(new_block->host, new_block->max_length, QEMU_MADV_DONTFORK);
1930 ram_block_notify_add(new_block->host, new_block->max_length);
1935 RAMBlock *qemu_ram_alloc_from_file(ram_addr_t size, MemoryRegion *mr,
1936 bool share, const char *mem_path,
1939 RAMBlock *new_block;
1940 Error *local_err = NULL;
1942 if (xen_enabled()) {
1943 error_setg(errp, "-mem-path not supported with Xen");
1947 if (phys_mem_alloc != qemu_anon_ram_alloc) {
1949 * file_ram_alloc() needs to allocate just like
1950 * phys_mem_alloc, but we haven't bothered to provide
1954 "-mem-path not supported with this accelerator");
1958 size = HOST_PAGE_ALIGN(size);
1959 new_block = g_malloc0(sizeof(*new_block));
1961 new_block->used_length = size;
1962 new_block->max_length = size;
1963 new_block->flags = share ? RAM_SHARED : 0;
1964 new_block->host = file_ram_alloc(new_block, size,
1966 if (!new_block->host) {
1971 ram_block_add(new_block, &local_err);
1974 error_propagate(errp, local_err);
1982 RAMBlock *qemu_ram_alloc_internal(ram_addr_t size, ram_addr_t max_size,
1983 void (*resized)(const char*,
1986 void *host, bool resizeable,
1987 MemoryRegion *mr, Error **errp)
1989 RAMBlock *new_block;
1990 Error *local_err = NULL;
1992 size = HOST_PAGE_ALIGN(size);
1993 max_size = HOST_PAGE_ALIGN(max_size);
1994 new_block = g_malloc0(sizeof(*new_block));
1996 new_block->resized = resized;
1997 new_block->used_length = size;
1998 new_block->max_length = max_size;
1999 assert(max_size >= size);
2001 new_block->page_size = getpagesize();
2002 new_block->host = host;
2004 new_block->flags |= RAM_PREALLOC;
2007 new_block->flags |= RAM_RESIZEABLE;
2009 ram_block_add(new_block, &local_err);
2012 error_propagate(errp, local_err);
2018 RAMBlock *qemu_ram_alloc_from_ptr(ram_addr_t size, void *host,
2019 MemoryRegion *mr, Error **errp)
2021 return qemu_ram_alloc_internal(size, size, NULL, host, false, mr, errp);
2024 RAMBlock *qemu_ram_alloc(ram_addr_t size, MemoryRegion *mr, Error **errp)
2026 return qemu_ram_alloc_internal(size, size, NULL, NULL, false, mr, errp);
2029 RAMBlock *qemu_ram_alloc_resizeable(ram_addr_t size, ram_addr_t maxsz,
2030 void (*resized)(const char*,
2033 MemoryRegion *mr, Error **errp)
2035 return qemu_ram_alloc_internal(size, maxsz, resized, NULL, true, mr, errp);
2038 static void reclaim_ramblock(RAMBlock *block)
2040 if (block->flags & RAM_PREALLOC) {
2042 } else if (xen_enabled()) {
2043 xen_invalidate_map_cache_entry(block->host);
2045 } else if (block->fd >= 0) {
2046 qemu_ram_munmap(block->host, block->max_length);
2050 qemu_anon_ram_free(block->host, block->max_length);
2055 void qemu_ram_free(RAMBlock *block)
2062 ram_block_notify_remove(block->host, block->max_length);
2065 qemu_mutex_lock_ramlist();
2066 QLIST_REMOVE_RCU(block, next);
2067 ram_list.mru_block = NULL;
2068 /* Write list before version */
2071 call_rcu(block, reclaim_ramblock, rcu);
2072 qemu_mutex_unlock_ramlist();
2076 void qemu_ram_remap(ram_addr_t addr, ram_addr_t length)
2083 RAMBLOCK_FOREACH(block) {
2084 offset = addr - block->offset;
2085 if (offset < block->max_length) {
2086 vaddr = ramblock_ptr(block, offset);
2087 if (block->flags & RAM_PREALLOC) {
2089 } else if (xen_enabled()) {
2093 if (block->fd >= 0) {
2094 flags |= (block->flags & RAM_SHARED ?
2095 MAP_SHARED : MAP_PRIVATE);
2096 area = mmap(vaddr, length, PROT_READ | PROT_WRITE,
2097 flags, block->fd, offset);
2100 * Remap needs to match alloc. Accelerators that
2101 * set phys_mem_alloc never remap. If they did,
2102 * we'd need a remap hook here.
2104 assert(phys_mem_alloc == qemu_anon_ram_alloc);
2106 flags |= MAP_PRIVATE | MAP_ANONYMOUS;
2107 area = mmap(vaddr, length, PROT_READ | PROT_WRITE,
2110 if (area != vaddr) {
2111 fprintf(stderr, "Could not remap addr: "
2112 RAM_ADDR_FMT "@" RAM_ADDR_FMT "\n",
2116 memory_try_enable_merging(vaddr, length);
2117 qemu_ram_setup_dump(vaddr, length);
2122 #endif /* !_WIN32 */
2124 /* Return a host pointer to ram allocated with qemu_ram_alloc.
2125 * This should not be used for general purpose DMA. Use address_space_map
2126 * or address_space_rw instead. For local memory (e.g. video ram) that the
2127 * device owns, use memory_region_get_ram_ptr.
2129 * Called within RCU critical section.
2131 void *qemu_map_ram_ptr(RAMBlock *ram_block, ram_addr_t addr)
2133 RAMBlock *block = ram_block;
2135 if (block == NULL) {
2136 block = qemu_get_ram_block(addr);
2137 addr -= block->offset;
2140 if (xen_enabled() && block->host == NULL) {
2141 /* We need to check if the requested address is in the RAM
2142 * because we don't want to map the entire memory in QEMU.
2143 * In that case just map until the end of the page.
2145 if (block->offset == 0) {
2146 return xen_map_cache(addr, 0, 0, false);
2149 block->host = xen_map_cache(block->offset, block->max_length, 1, false);
2151 return ramblock_ptr(block, addr);
2154 /* Return a host pointer to guest's ram. Similar to qemu_map_ram_ptr
2155 * but takes a size argument.
2157 * Called within RCU critical section.
2159 static void *qemu_ram_ptr_length(RAMBlock *ram_block, ram_addr_t addr,
2162 RAMBlock *block = ram_block;
2167 if (block == NULL) {
2168 block = qemu_get_ram_block(addr);
2169 addr -= block->offset;
2171 *size = MIN(*size, block->max_length - addr);
2173 if (xen_enabled() && block->host == NULL) {
2174 /* We need to check if the requested address is in the RAM
2175 * because we don't want to map the entire memory in QEMU.
2176 * In that case just map the requested area.
2178 if (block->offset == 0) {
2179 return xen_map_cache(addr, *size, 1, true);
2182 block->host = xen_map_cache(block->offset, block->max_length, 1, true);
2185 return ramblock_ptr(block, addr);
2189 * Translates a host ptr back to a RAMBlock, a ram_addr and an offset
2192 * ptr: Host pointer to look up
2193 * round_offset: If true round the result offset down to a page boundary
2194 * *ram_addr: set to result ram_addr
2195 * *offset: set to result offset within the RAMBlock
2197 * Returns: RAMBlock (or NULL if not found)
2199 * By the time this function returns, the returned pointer is not protected
2200 * by RCU anymore. If the caller is not within an RCU critical section and
2201 * does not hold the iothread lock, it must have other means of protecting the
2202 * pointer, such as a reference to the region that includes the incoming
2205 RAMBlock *qemu_ram_block_from_host(void *ptr, bool round_offset,
2209 uint8_t *host = ptr;
2211 if (xen_enabled()) {
2212 ram_addr_t ram_addr;
2214 ram_addr = xen_ram_addr_from_mapcache(ptr);
2215 block = qemu_get_ram_block(ram_addr);
2217 *offset = ram_addr - block->offset;
2224 block = atomic_rcu_read(&ram_list.mru_block);
2225 if (block && block->host && host - block->host < block->max_length) {
2229 RAMBLOCK_FOREACH(block) {
2230 /* This case append when the block is not mapped. */
2231 if (block->host == NULL) {
2234 if (host - block->host < block->max_length) {
2243 *offset = (host - block->host);
2245 *offset &= TARGET_PAGE_MASK;
2252 * Finds the named RAMBlock
2254 * name: The name of RAMBlock to find
2256 * Returns: RAMBlock (or NULL if not found)
2258 RAMBlock *qemu_ram_block_by_name(const char *name)
2262 RAMBLOCK_FOREACH(block) {
2263 if (!strcmp(name, block->idstr)) {
2271 /* Some of the softmmu routines need to translate from a host pointer
2272 (typically a TLB entry) back to a ram offset. */
2273 ram_addr_t qemu_ram_addr_from_host(void *ptr)
2278 block = qemu_ram_block_from_host(ptr, false, &offset);
2280 return RAM_ADDR_INVALID;
2283 return block->offset + offset;
2286 /* Called within RCU critical section. */
2287 static void notdirty_mem_write(void *opaque, hwaddr ram_addr,
2288 uint64_t val, unsigned size)
2290 bool locked = false;
2292 if (!cpu_physical_memory_get_dirty_flag(ram_addr, DIRTY_MEMORY_CODE)) {
2295 tb_invalidate_phys_page_fast(ram_addr, size);
2299 stb_p(qemu_map_ram_ptr(NULL, ram_addr), val);
2302 stw_p(qemu_map_ram_ptr(NULL, ram_addr), val);
2305 stl_p(qemu_map_ram_ptr(NULL, ram_addr), val);
2315 /* Set both VGA and migration bits for simplicity and to remove
2316 * the notdirty callback faster.
2318 cpu_physical_memory_set_dirty_range(ram_addr, size,
2319 DIRTY_CLIENTS_NOCODE);
2320 /* we remove the notdirty callback only if the code has been
2322 if (!cpu_physical_memory_is_clean(ram_addr)) {
2323 tlb_set_dirty(current_cpu, current_cpu->mem_io_vaddr);
2327 static bool notdirty_mem_accepts(void *opaque, hwaddr addr,
2328 unsigned size, bool is_write)
2333 static const MemoryRegionOps notdirty_mem_ops = {
2334 .write = notdirty_mem_write,
2335 .valid.accepts = notdirty_mem_accepts,
2336 .endianness = DEVICE_NATIVE_ENDIAN,
2339 /* Generate a debug exception if a watchpoint has been hit. */
2340 static void check_watchpoint(int offset, int len, MemTxAttrs attrs, int flags)
2342 CPUState *cpu = current_cpu;
2343 CPUClass *cc = CPU_GET_CLASS(cpu);
2344 CPUArchState *env = cpu->env_ptr;
2345 target_ulong pc, cs_base;
2350 if (cpu->watchpoint_hit) {
2351 /* We re-entered the check after replacing the TB. Now raise
2352 * the debug interrupt so that is will trigger after the
2353 * current instruction. */
2354 cpu_interrupt(cpu, CPU_INTERRUPT_DEBUG);
2357 vaddr = (cpu->mem_io_vaddr & TARGET_PAGE_MASK) + offset;
2358 vaddr = cc->adjust_watchpoint_address(cpu, vaddr, len);
2359 QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
2360 if (cpu_watchpoint_address_matches(wp, vaddr, len)
2361 && (wp->flags & flags)) {
2362 if (flags == BP_MEM_READ) {
2363 wp->flags |= BP_WATCHPOINT_HIT_READ;
2365 wp->flags |= BP_WATCHPOINT_HIT_WRITE;
2367 wp->hitaddr = vaddr;
2368 wp->hitattrs = attrs;
2369 if (!cpu->watchpoint_hit) {
2370 if (wp->flags & BP_CPU &&
2371 !cc->debug_check_watchpoint(cpu, wp)) {
2372 wp->flags &= ~BP_WATCHPOINT_HIT;
2375 cpu->watchpoint_hit = wp;
2377 /* Both tb_lock and iothread_mutex will be reset when
2378 * cpu_loop_exit or cpu_loop_exit_noexc longjmp
2379 * back into the cpu_exec main loop.
2382 tb_check_watchpoint(cpu);
2383 if (wp->flags & BP_STOP_BEFORE_ACCESS) {
2384 cpu->exception_index = EXCP_DEBUG;
2387 cpu_get_tb_cpu_state(env, &pc, &cs_base, &cpu_flags);
2388 tb_gen_code(cpu, pc, cs_base, cpu_flags, 1);
2389 cpu_loop_exit_noexc(cpu);
2393 wp->flags &= ~BP_WATCHPOINT_HIT;
2398 /* Watchpoint access routines. Watchpoints are inserted using TLB tricks,
2399 so these check for a hit then pass through to the normal out-of-line
2401 static MemTxResult watch_mem_read(void *opaque, hwaddr addr, uint64_t *pdata,
2402 unsigned size, MemTxAttrs attrs)
2406 int asidx = cpu_asidx_from_attrs(current_cpu, attrs);
2407 AddressSpace *as = current_cpu->cpu_ases[asidx].as;
2409 check_watchpoint(addr & ~TARGET_PAGE_MASK, size, attrs, BP_MEM_READ);
2412 data = address_space_ldub(as, addr, attrs, &res);
2415 data = address_space_lduw(as, addr, attrs, &res);
2418 data = address_space_ldl(as, addr, attrs, &res);
2426 static MemTxResult watch_mem_write(void *opaque, hwaddr addr,
2427 uint64_t val, unsigned size,
2431 int asidx = cpu_asidx_from_attrs(current_cpu, attrs);
2432 AddressSpace *as = current_cpu->cpu_ases[asidx].as;
2434 check_watchpoint(addr & ~TARGET_PAGE_MASK, size, attrs, BP_MEM_WRITE);
2437 address_space_stb(as, addr, val, attrs, &res);
2440 address_space_stw(as, addr, val, attrs, &res);
2443 address_space_stl(as, addr, val, attrs, &res);
2450 static const MemoryRegionOps watch_mem_ops = {
2451 .read_with_attrs = watch_mem_read,
2452 .write_with_attrs = watch_mem_write,
2453 .endianness = DEVICE_NATIVE_ENDIAN,
2456 static MemTxResult subpage_read(void *opaque, hwaddr addr, uint64_t *data,
2457 unsigned len, MemTxAttrs attrs)
2459 subpage_t *subpage = opaque;
2463 #if defined(DEBUG_SUBPAGE)
2464 printf("%s: subpage %p len %u addr " TARGET_FMT_plx "\n", __func__,
2465 subpage, len, addr);
2467 res = address_space_read(subpage->as, addr + subpage->base,
2474 *data = ldub_p(buf);
2477 *data = lduw_p(buf);
2490 static MemTxResult subpage_write(void *opaque, hwaddr addr,
2491 uint64_t value, unsigned len, MemTxAttrs attrs)
2493 subpage_t *subpage = opaque;
2496 #if defined(DEBUG_SUBPAGE)
2497 printf("%s: subpage %p len %u addr " TARGET_FMT_plx
2498 " value %"PRIx64"\n",
2499 __func__, subpage, len, addr, value);
2517 return address_space_write(subpage->as, addr + subpage->base,
2521 static bool subpage_accepts(void *opaque, hwaddr addr,
2522 unsigned len, bool is_write)
2524 subpage_t *subpage = opaque;
2525 #if defined(DEBUG_SUBPAGE)
2526 printf("%s: subpage %p %c len %u addr " TARGET_FMT_plx "\n",
2527 __func__, subpage, is_write ? 'w' : 'r', len, addr);
2530 return address_space_access_valid(subpage->as, addr + subpage->base,
2534 static const MemoryRegionOps subpage_ops = {
2535 .read_with_attrs = subpage_read,
2536 .write_with_attrs = subpage_write,
2537 .impl.min_access_size = 1,
2538 .impl.max_access_size = 8,
2539 .valid.min_access_size = 1,
2540 .valid.max_access_size = 8,
2541 .valid.accepts = subpage_accepts,
2542 .endianness = DEVICE_NATIVE_ENDIAN,
2545 static int subpage_register (subpage_t *mmio, uint32_t start, uint32_t end,
2550 if (start >= TARGET_PAGE_SIZE || end >= TARGET_PAGE_SIZE)
2552 idx = SUBPAGE_IDX(start);
2553 eidx = SUBPAGE_IDX(end);
2554 #if defined(DEBUG_SUBPAGE)
2555 printf("%s: %p start %08x end %08x idx %08x eidx %08x section %d\n",
2556 __func__, mmio, start, end, idx, eidx, section);
2558 for (; idx <= eidx; idx++) {
2559 mmio->sub_section[idx] = section;
2565 static subpage_t *subpage_init(AddressSpace *as, hwaddr base)
2569 mmio = g_malloc0(sizeof(subpage_t) + TARGET_PAGE_SIZE * sizeof(uint16_t));
2572 memory_region_init_io(&mmio->iomem, NULL, &subpage_ops, mmio,
2573 NULL, TARGET_PAGE_SIZE);
2574 mmio->iomem.subpage = true;
2575 #if defined(DEBUG_SUBPAGE)
2576 printf("%s: %p base " TARGET_FMT_plx " len %08x\n", __func__,
2577 mmio, base, TARGET_PAGE_SIZE);
2579 subpage_register(mmio, 0, TARGET_PAGE_SIZE-1, PHYS_SECTION_UNASSIGNED);
2584 static uint16_t dummy_section(PhysPageMap *map, AddressSpace *as,
2588 MemoryRegionSection section = {
2589 .address_space = as,
2591 .offset_within_address_space = 0,
2592 .offset_within_region = 0,
2593 .size = int128_2_64(),
2596 return phys_section_add(map, §ion);
2599 MemoryRegion *iotlb_to_region(CPUState *cpu, hwaddr index, MemTxAttrs attrs)
2601 int asidx = cpu_asidx_from_attrs(cpu, attrs);
2602 CPUAddressSpace *cpuas = &cpu->cpu_ases[asidx];
2603 AddressSpaceDispatch *d = atomic_rcu_read(&cpuas->memory_dispatch);
2604 MemoryRegionSection *sections = d->map.sections;
2606 return sections[index & ~TARGET_PAGE_MASK].mr;
2609 static void io_mem_init(void)
2611 memory_region_init_io(&io_mem_rom, NULL, &unassigned_mem_ops, NULL, NULL, UINT64_MAX);
2612 memory_region_init_io(&io_mem_unassigned, NULL, &unassigned_mem_ops, NULL,
2615 /* io_mem_notdirty calls tb_invalidate_phys_page_fast,
2616 * which can be called without the iothread mutex.
2618 memory_region_init_io(&io_mem_notdirty, NULL, ¬dirty_mem_ops, NULL,
2620 memory_region_clear_global_locking(&io_mem_notdirty);
2622 memory_region_init_io(&io_mem_watch, NULL, &watch_mem_ops, NULL,
2626 static void mem_begin(MemoryListener *listener)
2628 AddressSpace *as = container_of(listener, AddressSpace, dispatch_listener);
2629 AddressSpaceDispatch *d = g_new0(AddressSpaceDispatch, 1);
2632 n = dummy_section(&d->map, as, &io_mem_unassigned);
2633 assert(n == PHYS_SECTION_UNASSIGNED);
2634 n = dummy_section(&d->map, as, &io_mem_notdirty);
2635 assert(n == PHYS_SECTION_NOTDIRTY);
2636 n = dummy_section(&d->map, as, &io_mem_rom);
2637 assert(n == PHYS_SECTION_ROM);
2638 n = dummy_section(&d->map, as, &io_mem_watch);
2639 assert(n == PHYS_SECTION_WATCH);
2641 d->phys_map = (PhysPageEntry) { .ptr = PHYS_MAP_NODE_NIL, .skip = 1 };
2643 as->next_dispatch = d;
2646 static void address_space_dispatch_free(AddressSpaceDispatch *d)
2648 phys_sections_free(&d->map);
2652 static void mem_commit(MemoryListener *listener)
2654 AddressSpace *as = container_of(listener, AddressSpace, dispatch_listener);
2655 AddressSpaceDispatch *cur = as->dispatch;
2656 AddressSpaceDispatch *next = as->next_dispatch;
2658 phys_page_compact_all(next, next->map.nodes_nb);
2660 atomic_rcu_set(&as->dispatch, next);
2662 call_rcu(cur, address_space_dispatch_free, rcu);
2666 static void tcg_commit(MemoryListener *listener)
2668 CPUAddressSpace *cpuas;
2669 AddressSpaceDispatch *d;
2671 /* since each CPU stores ram addresses in its TLB cache, we must
2672 reset the modified entries */
2673 cpuas = container_of(listener, CPUAddressSpace, tcg_as_listener);
2674 cpu_reloading_memory_map();
2675 /* The CPU and TLB are protected by the iothread lock.
2676 * We reload the dispatch pointer now because cpu_reloading_memory_map()
2677 * may have split the RCU critical section.
2679 d = atomic_rcu_read(&cpuas->as->dispatch);
2680 atomic_rcu_set(&cpuas->memory_dispatch, d);
2681 tlb_flush(cpuas->cpu);
2684 void address_space_init_dispatch(AddressSpace *as)
2686 as->dispatch = NULL;
2687 as->dispatch_listener = (MemoryListener) {
2689 .commit = mem_commit,
2690 .region_add = mem_add,
2691 .region_nop = mem_add,
2694 memory_listener_register(&as->dispatch_listener, as);
2697 void address_space_unregister(AddressSpace *as)
2699 memory_listener_unregister(&as->dispatch_listener);
2702 void address_space_destroy_dispatch(AddressSpace *as)
2704 AddressSpaceDispatch *d = as->dispatch;
2706 atomic_rcu_set(&as->dispatch, NULL);
2708 call_rcu(d, address_space_dispatch_free, rcu);
2712 static void memory_map_init(void)
2714 system_memory = g_malloc(sizeof(*system_memory));
2716 memory_region_init(system_memory, NULL, "system", UINT64_MAX);
2717 address_space_init(&address_space_memory, system_memory, "memory");
2719 system_io = g_malloc(sizeof(*system_io));
2720 memory_region_init_io(system_io, NULL, &unassigned_io_ops, NULL, "io",
2722 address_space_init(&address_space_io, system_io, "I/O");
2725 MemoryRegion *get_system_memory(void)
2727 return system_memory;
2730 MemoryRegion *get_system_io(void)
2735 #endif /* !defined(CONFIG_USER_ONLY) */
2737 /* physical memory access (slow version, mainly for debug) */
2738 #if defined(CONFIG_USER_ONLY)
2739 int cpu_memory_rw_debug(CPUState *cpu, target_ulong addr,
2740 uint8_t *buf, int len, int is_write)
2747 page = addr & TARGET_PAGE_MASK;
2748 l = (page + TARGET_PAGE_SIZE) - addr;
2751 flags = page_get_flags(page);
2752 if (!(flags & PAGE_VALID))
2755 if (!(flags & PAGE_WRITE))
2757 /* XXX: this code should not depend on lock_user */
2758 if (!(p = lock_user(VERIFY_WRITE, addr, l, 0)))
2761 unlock_user(p, addr, l);
2763 if (!(flags & PAGE_READ))
2765 /* XXX: this code should not depend on lock_user */
2766 if (!(p = lock_user(VERIFY_READ, addr, l, 1)))
2769 unlock_user(p, addr, 0);
2780 static void invalidate_and_set_dirty(MemoryRegion *mr, hwaddr addr,
2783 uint8_t dirty_log_mask = memory_region_get_dirty_log_mask(mr);
2784 addr += memory_region_get_ram_addr(mr);
2786 /* No early return if dirty_log_mask is or becomes 0, because
2787 * cpu_physical_memory_set_dirty_range will still call
2788 * xen_modified_memory.
2790 if (dirty_log_mask) {
2792 cpu_physical_memory_range_includes_clean(addr, length, dirty_log_mask);
2794 if (dirty_log_mask & (1 << DIRTY_MEMORY_CODE)) {
2796 tb_invalidate_phys_range(addr, addr + length);
2798 dirty_log_mask &= ~(1 << DIRTY_MEMORY_CODE);
2800 cpu_physical_memory_set_dirty_range(addr, length, dirty_log_mask);
2803 static int memory_access_size(MemoryRegion *mr, unsigned l, hwaddr addr)
2805 unsigned access_size_max = mr->ops->valid.max_access_size;
2807 /* Regions are assumed to support 1-4 byte accesses unless
2808 otherwise specified. */
2809 if (access_size_max == 0) {
2810 access_size_max = 4;
2813 /* Bound the maximum access by the alignment of the address. */
2814 if (!mr->ops->impl.unaligned) {
2815 unsigned align_size_max = addr & -addr;
2816 if (align_size_max != 0 && align_size_max < access_size_max) {
2817 access_size_max = align_size_max;
2821 /* Don't attempt accesses larger than the maximum. */
2822 if (l > access_size_max) {
2823 l = access_size_max;
2830 static bool prepare_mmio_access(MemoryRegion *mr)
2832 bool unlocked = !qemu_mutex_iothread_locked();
2833 bool release_lock = false;
2835 if (unlocked && mr->global_locking) {
2836 qemu_mutex_lock_iothread();
2838 release_lock = true;
2840 if (mr->flush_coalesced_mmio) {
2842 qemu_mutex_lock_iothread();
2844 qemu_flush_coalesced_mmio_buffer();
2846 qemu_mutex_unlock_iothread();
2850 return release_lock;
2853 /* Called within RCU critical section. */
2854 static MemTxResult address_space_write_continue(AddressSpace *as, hwaddr addr,
2857 int len, hwaddr addr1,
2858 hwaddr l, MemoryRegion *mr)
2862 MemTxResult result = MEMTX_OK;
2863 bool release_lock = false;
2866 if (!memory_access_is_direct(mr, true)) {
2867 release_lock |= prepare_mmio_access(mr);
2868 l = memory_access_size(mr, l, addr1);
2869 /* XXX: could force current_cpu to NULL to avoid
2873 /* 64 bit write access */
2875 result |= memory_region_dispatch_write(mr, addr1, val, 8,
2879 /* 32 bit write access */
2880 val = (uint32_t)ldl_p(buf);
2881 result |= memory_region_dispatch_write(mr, addr1, val, 4,
2885 /* 16 bit write access */
2887 result |= memory_region_dispatch_write(mr, addr1, val, 2,
2891 /* 8 bit write access */
2893 result |= memory_region_dispatch_write(mr, addr1, val, 1,
2901 ptr = qemu_map_ram_ptr(mr->ram_block, addr1);
2902 memcpy(ptr, buf, l);
2903 invalidate_and_set_dirty(mr, addr1, l);
2907 qemu_mutex_unlock_iothread();
2908 release_lock = false;
2920 mr = address_space_translate(as, addr, &addr1, &l, true);
2926 MemTxResult address_space_write(AddressSpace *as, hwaddr addr, MemTxAttrs attrs,
2927 const uint8_t *buf, int len)
2932 MemTxResult result = MEMTX_OK;
2937 mr = address_space_translate(as, addr, &addr1, &l, true);
2938 result = address_space_write_continue(as, addr, attrs, buf, len,
2946 /* Called within RCU critical section. */
2947 MemTxResult address_space_read_continue(AddressSpace *as, hwaddr addr,
2948 MemTxAttrs attrs, uint8_t *buf,
2949 int len, hwaddr addr1, hwaddr l,
2954 MemTxResult result = MEMTX_OK;
2955 bool release_lock = false;
2958 if (!memory_access_is_direct(mr, false)) {
2960 release_lock |= prepare_mmio_access(mr);
2961 l = memory_access_size(mr, l, addr1);
2964 /* 64 bit read access */
2965 result |= memory_region_dispatch_read(mr, addr1, &val, 8,
2970 /* 32 bit read access */
2971 result |= memory_region_dispatch_read(mr, addr1, &val, 4,
2976 /* 16 bit read access */
2977 result |= memory_region_dispatch_read(mr, addr1, &val, 2,
2982 /* 8 bit read access */
2983 result |= memory_region_dispatch_read(mr, addr1, &val, 1,
2992 ptr = qemu_map_ram_ptr(mr->ram_block, addr1);
2993 memcpy(buf, ptr, l);
2997 qemu_mutex_unlock_iothread();
2998 release_lock = false;
3010 mr = address_space_translate(as, addr, &addr1, &l, false);
3016 MemTxResult address_space_read_full(AddressSpace *as, hwaddr addr,
3017 MemTxAttrs attrs, uint8_t *buf, int len)
3022 MemTxResult result = MEMTX_OK;
3027 mr = address_space_translate(as, addr, &addr1, &l, false);
3028 result = address_space_read_continue(as, addr, attrs, buf, len,
3036 MemTxResult address_space_rw(AddressSpace *as, hwaddr addr, MemTxAttrs attrs,
3037 uint8_t *buf, int len, bool is_write)
3040 return address_space_write(as, addr, attrs, (uint8_t *)buf, len);
3042 return address_space_read(as, addr, attrs, (uint8_t *)buf, len);
3046 void cpu_physical_memory_rw(hwaddr addr, uint8_t *buf,
3047 int len, int is_write)
3049 address_space_rw(&address_space_memory, addr, MEMTXATTRS_UNSPECIFIED,
3050 buf, len, is_write);
3053 enum write_rom_type {
3058 static inline void cpu_physical_memory_write_rom_internal(AddressSpace *as,
3059 hwaddr addr, const uint8_t *buf, int len, enum write_rom_type type)
3069 mr = address_space_translate(as, addr, &addr1, &l, true);
3071 if (!(memory_region_is_ram(mr) ||
3072 memory_region_is_romd(mr))) {
3073 l = memory_access_size(mr, l, addr1);
3076 ptr = qemu_map_ram_ptr(mr->ram_block, addr1);
3079 memcpy(ptr, buf, l);
3080 invalidate_and_set_dirty(mr, addr1, l);
3083 flush_icache_range((uintptr_t)ptr, (uintptr_t)ptr + l);
3094 /* used for ROM loading : can write in RAM and ROM */
3095 void cpu_physical_memory_write_rom(AddressSpace *as, hwaddr addr,
3096 const uint8_t *buf, int len)
3098 cpu_physical_memory_write_rom_internal(as, addr, buf, len, WRITE_DATA);
3101 void cpu_flush_icache_range(hwaddr start, int len)
3104 * This function should do the same thing as an icache flush that was
3105 * triggered from within the guest. For TCG we are always cache coherent,
3106 * so there is no need to flush anything. For KVM / Xen we need to flush
3107 * the host's instruction cache at least.
3109 if (tcg_enabled()) {
3113 cpu_physical_memory_write_rom_internal(&address_space_memory,
3114 start, NULL, len, FLUSH_CACHE);
3125 static BounceBuffer bounce;
3127 typedef struct MapClient {
3129 QLIST_ENTRY(MapClient) link;
3132 QemuMutex map_client_list_lock;
3133 static QLIST_HEAD(map_client_list, MapClient) map_client_list
3134 = QLIST_HEAD_INITIALIZER(map_client_list);
3136 static void cpu_unregister_map_client_do(MapClient *client)
3138 QLIST_REMOVE(client, link);
3142 static void cpu_notify_map_clients_locked(void)
3146 while (!QLIST_EMPTY(&map_client_list)) {
3147 client = QLIST_FIRST(&map_client_list);
3148 qemu_bh_schedule(client->bh);
3149 cpu_unregister_map_client_do(client);
3153 void cpu_register_map_client(QEMUBH *bh)
3155 MapClient *client = g_malloc(sizeof(*client));
3157 qemu_mutex_lock(&map_client_list_lock);
3159 QLIST_INSERT_HEAD(&map_client_list, client, link);
3160 if (!atomic_read(&bounce.in_use)) {
3161 cpu_notify_map_clients_locked();
3163 qemu_mutex_unlock(&map_client_list_lock);
3166 void cpu_exec_init_all(void)
3168 qemu_mutex_init(&ram_list.mutex);
3169 /* The data structures we set up here depend on knowing the page size,
3170 * so no more changes can be made after this point.
3171 * In an ideal world, nothing we did before we had finished the
3172 * machine setup would care about the target page size, and we could
3173 * do this much later, rather than requiring board models to state
3174 * up front what their requirements are.
3176 finalize_target_page_bits();
3179 qemu_mutex_init(&map_client_list_lock);
3182 void cpu_unregister_map_client(QEMUBH *bh)
3186 qemu_mutex_lock(&map_client_list_lock);
3187 QLIST_FOREACH(client, &map_client_list, link) {
3188 if (client->bh == bh) {
3189 cpu_unregister_map_client_do(client);
3193 qemu_mutex_unlock(&map_client_list_lock);
3196 static void cpu_notify_map_clients(void)
3198 qemu_mutex_lock(&map_client_list_lock);
3199 cpu_notify_map_clients_locked();
3200 qemu_mutex_unlock(&map_client_list_lock);
3203 bool address_space_access_valid(AddressSpace *as, hwaddr addr, int len, bool is_write)
3211 mr = address_space_translate(as, addr, &xlat, &l, is_write);
3212 if (!memory_access_is_direct(mr, is_write)) {
3213 l = memory_access_size(mr, l, addr);
3214 if (!memory_region_access_valid(mr, xlat, l, is_write)) {
3228 address_space_extend_translation(AddressSpace *as, hwaddr addr, hwaddr target_len,
3229 MemoryRegion *mr, hwaddr base, hwaddr len,
3234 MemoryRegion *this_mr;
3240 if (target_len == 0) {
3245 this_mr = address_space_translate(as, addr, &xlat, &len, is_write);
3246 if (this_mr != mr || xlat != base + done) {
3252 /* Map a physical memory region into a host virtual address.
3253 * May map a subset of the requested range, given by and returned in *plen.
3254 * May return NULL if resources needed to perform the mapping are exhausted.
3255 * Use only for reads OR writes - not for read-modify-write operations.
3256 * Use cpu_register_map_client() to know when retrying the map operation is
3257 * likely to succeed.
3259 void *address_space_map(AddressSpace *as,
3275 mr = address_space_translate(as, addr, &xlat, &l, is_write);
3277 if (!memory_access_is_direct(mr, is_write)) {
3278 if (atomic_xchg(&bounce.in_use, true)) {
3282 /* Avoid unbounded allocations */
3283 l = MIN(l, TARGET_PAGE_SIZE);
3284 bounce.buffer = qemu_memalign(TARGET_PAGE_SIZE, l);
3288 memory_region_ref(mr);
3291 address_space_read(as, addr, MEMTXATTRS_UNSPECIFIED,
3297 return bounce.buffer;
3301 memory_region_ref(mr);
3302 *plen = address_space_extend_translation(as, addr, len, mr, xlat, l, is_write);
3303 ptr = qemu_ram_ptr_length(mr->ram_block, xlat, plen);
3309 /* Unmaps a memory region previously mapped by address_space_map().
3310 * Will also mark the memory as dirty if is_write == 1. access_len gives
3311 * the amount of memory that was actually read or written by the caller.
3313 void address_space_unmap(AddressSpace *as, void *buffer, hwaddr len,
3314 int is_write, hwaddr access_len)
3316 if (buffer != bounce.buffer) {
3320 mr = memory_region_from_host(buffer, &addr1);
3323 invalidate_and_set_dirty(mr, addr1, access_len);
3325 if (xen_enabled()) {
3326 xen_invalidate_map_cache_entry(buffer);
3328 memory_region_unref(mr);
3332 address_space_write(as, bounce.addr, MEMTXATTRS_UNSPECIFIED,
3333 bounce.buffer, access_len);
3335 qemu_vfree(bounce.buffer);
3336 bounce.buffer = NULL;
3337 memory_region_unref(bounce.mr);
3338 atomic_mb_set(&bounce.in_use, false);
3339 cpu_notify_map_clients();
3342 void *cpu_physical_memory_map(hwaddr addr,
3346 return address_space_map(&address_space_memory, addr, plen, is_write);
3349 void cpu_physical_memory_unmap(void *buffer, hwaddr len,
3350 int is_write, hwaddr access_len)
3352 return address_space_unmap(&address_space_memory, buffer, len, is_write, access_len);
3355 #define ARG1_DECL AddressSpace *as
3358 #define TRANSLATE(...) address_space_translate(as, __VA_ARGS__)
3359 #define IS_DIRECT(mr, is_write) memory_access_is_direct(mr, is_write)
3360 #define MAP_RAM(mr, ofs) qemu_map_ram_ptr((mr)->ram_block, ofs)
3361 #define INVALIDATE(mr, ofs, len) invalidate_and_set_dirty(mr, ofs, len)
3362 #define RCU_READ_LOCK(...) rcu_read_lock()
3363 #define RCU_READ_UNLOCK(...) rcu_read_unlock()
3364 #include "memory_ldst.inc.c"
3366 int64_t address_space_cache_init(MemoryRegionCache *cache,
3378 void address_space_cache_invalidate(MemoryRegionCache *cache,
3384 void address_space_cache_destroy(MemoryRegionCache *cache)
3389 #define ARG1_DECL MemoryRegionCache *cache
3391 #define SUFFIX _cached
3392 #define TRANSLATE(addr, ...) \
3393 address_space_translate(cache->as, cache->xlat + (addr), __VA_ARGS__)
3394 #define IS_DIRECT(mr, is_write) true
3395 #define MAP_RAM(mr, ofs) qemu_map_ram_ptr((mr)->ram_block, ofs)
3396 #define INVALIDATE(mr, ofs, len) invalidate_and_set_dirty(mr, ofs, len)
3397 #define RCU_READ_LOCK() rcu_read_lock()
3398 #define RCU_READ_UNLOCK() rcu_read_unlock()
3399 #include "memory_ldst.inc.c"
3401 /* virtual memory access for debug (includes writing to ROM) */
3402 int cpu_memory_rw_debug(CPUState *cpu, target_ulong addr,
3403 uint8_t *buf, int len, int is_write)
3409 cpu_synchronize_state(cpu);
3414 page = addr & TARGET_PAGE_MASK;
3415 phys_addr = cpu_get_phys_page_attrs_debug(cpu, page, &attrs);
3416 asidx = cpu_asidx_from_attrs(cpu, attrs);
3417 /* if no physical page mapped, return an error */
3418 if (phys_addr == -1)
3420 l = (page + TARGET_PAGE_SIZE) - addr;
3423 phys_addr += (addr & ~TARGET_PAGE_MASK);
3425 cpu_physical_memory_write_rom(cpu->cpu_ases[asidx].as,
3428 address_space_rw(cpu->cpu_ases[asidx].as, phys_addr,
3429 MEMTXATTRS_UNSPECIFIED,
3440 * Allows code that needs to deal with migration bitmaps etc to still be built
3441 * target independent.
3443 size_t qemu_target_page_size(void)
3445 return TARGET_PAGE_SIZE;
3448 int qemu_target_page_bits(void)
3450 return TARGET_PAGE_BITS;
3453 int qemu_target_page_bits_min(void)
3455 return TARGET_PAGE_BITS_MIN;
3460 * A helper function for the _utterly broken_ virtio device model to find out if
3461 * it's running on a big endian machine. Don't do this at home kids!
3463 bool target_words_bigendian(void);
3464 bool target_words_bigendian(void)
3466 #if defined(TARGET_WORDS_BIGENDIAN)
3473 #ifndef CONFIG_USER_ONLY
3474 bool cpu_physical_memory_is_io(hwaddr phys_addr)
3481 mr = address_space_translate(&address_space_memory,
3482 phys_addr, &phys_addr, &l, false);
3484 res = !(memory_region_is_ram(mr) || memory_region_is_romd(mr));
3489 int qemu_ram_foreach_block(RAMBlockIterFunc func, void *opaque)
3495 RAMBLOCK_FOREACH(block) {
3496 ret = func(block->idstr, block->host, block->offset,
3497 block->used_length, opaque);
3507 * Unmap pages of memory from start to start+length such that
3508 * they a) read as 0, b) Trigger whatever fault mechanism
3509 * the OS provides for postcopy.
3510 * The pages must be unmapped by the end of the function.
3511 * Returns: 0 on success, none-0 on failure
3514 int ram_block_discard_range(RAMBlock *rb, uint64_t start, size_t length)
3518 uint8_t *host_startaddr = rb->host + start;
3520 if ((uintptr_t)host_startaddr & (rb->page_size - 1)) {
3521 error_report("ram_block_discard_range: Unaligned start address: %p",
3526 if ((start + length) <= rb->used_length) {
3527 uint8_t *host_endaddr = host_startaddr + length;
3528 if ((uintptr_t)host_endaddr & (rb->page_size - 1)) {
3529 error_report("ram_block_discard_range: Unaligned end address: %p",
3534 errno = ENOTSUP; /* If we are missing MADVISE etc */
3536 if (rb->page_size == qemu_host_page_size) {
3537 #if defined(CONFIG_MADVISE)
3538 /* Note: We need the madvise MADV_DONTNEED behaviour of definitely
3541 ret = madvise(host_startaddr, length, MADV_DONTNEED);
3544 /* Huge page case - unfortunately it can't do DONTNEED, but
3545 * it can do the equivalent by FALLOC_FL_PUNCH_HOLE in the
3548 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
3549 ret = fallocate(rb->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
3555 error_report("ram_block_discard_range: Failed to discard range "
3556 "%s:%" PRIx64 " +%zx (%d)",
3557 rb->idstr, start, length, ret);
3560 error_report("ram_block_discard_range: Overrun block '%s' (%" PRIu64
3561 "/%zx/" RAM_ADDR_FMT")",
3562 rb->idstr, start, length, rb->used_length);