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/>.
20 #include "qemu/osdep.h"
21 #include "qemu-common.h"
22 #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 #include "hw/qdev-properties.h"
31 #if !defined(CONFIG_USER_ONLY)
32 #include "hw/boards.h"
33 #include "hw/xen/xen.h"
35 #include "sysemu/kvm.h"
36 #include "sysemu/sysemu.h"
37 #include "sysemu/tcg.h"
38 #include "sysemu/qtest.h"
39 #include "qemu/timer.h"
40 #include "qemu/config-file.h"
41 #include "qemu/error-report.h"
42 #include "qemu/qemu-print.h"
43 #if defined(CONFIG_USER_ONLY)
45 #else /* !CONFIG_USER_ONLY */
46 #include "exec/memory.h"
47 #include "exec/ioport.h"
48 #include "sysemu/dma.h"
49 #include "sysemu/hostmem.h"
50 #include "sysemu/hw_accel.h"
51 #include "exec/address-spaces.h"
52 #include "sysemu/xen-mapcache.h"
53 #include "trace/trace-root.h"
55 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
56 #include <linux/falloc.h>
60 #include "qemu/rcu_queue.h"
61 #include "qemu/main-loop.h"
62 #include "translate-all.h"
63 #include "sysemu/replay.h"
65 #include "exec/memory-internal.h"
66 #include "exec/ram_addr.h"
69 #include "qemu/pmem.h"
71 #include "migration/vmstate.h"
73 #include "qemu/range.h"
75 #include "qemu/mmap-alloc.h"
78 #include "monitor/monitor.h"
80 #ifdef CONFIG_LIBDAXCTL
81 #include <daxctl/libdaxctl.h>
84 //#define DEBUG_SUBPAGE
86 #if !defined(CONFIG_USER_ONLY)
87 /* ram_list is read under rcu_read_lock()/rcu_read_unlock(). Writes
88 * are protected by the ramlist lock.
90 RAMList ram_list = { .blocks = QLIST_HEAD_INITIALIZER(ram_list.blocks) };
92 static MemoryRegion *system_memory;
93 static MemoryRegion *system_io;
95 AddressSpace address_space_io;
96 AddressSpace address_space_memory;
98 static MemoryRegion io_mem_unassigned;
101 uintptr_t qemu_host_page_size;
102 intptr_t qemu_host_page_mask;
104 #if !defined(CONFIG_USER_ONLY)
105 /* 0 = Do not count executed instructions.
106 1 = Precise instruction counting.
107 2 = Adaptive rate instruction counting. */
110 typedef struct PhysPageEntry PhysPageEntry;
112 struct PhysPageEntry {
113 /* How many bits skip to next level (in units of L2_SIZE). 0 for a leaf. */
115 /* index into phys_sections (!skip) or phys_map_nodes (skip) */
119 #define PHYS_MAP_NODE_NIL (((uint32_t)~0) >> 6)
121 /* Size of the L2 (and L3, etc) page tables. */
122 #define ADDR_SPACE_BITS 64
125 #define P_L2_SIZE (1 << P_L2_BITS)
127 #define P_L2_LEVELS (((ADDR_SPACE_BITS - TARGET_PAGE_BITS - 1) / P_L2_BITS) + 1)
129 typedef PhysPageEntry Node[P_L2_SIZE];
131 typedef struct PhysPageMap {
134 unsigned sections_nb;
135 unsigned sections_nb_alloc;
137 unsigned nodes_nb_alloc;
139 MemoryRegionSection *sections;
142 struct AddressSpaceDispatch {
143 MemoryRegionSection *mru_section;
144 /* This is a multi-level map on the physical address space.
145 * The bottom level has pointers to MemoryRegionSections.
147 PhysPageEntry phys_map;
151 #define SUBPAGE_IDX(addr) ((addr) & ~TARGET_PAGE_MASK)
152 typedef struct subpage_t {
156 uint16_t sub_section[];
159 #define PHYS_SECTION_UNASSIGNED 0
161 static void io_mem_init(void);
162 static void memory_map_init(void);
163 static void tcg_log_global_after_sync(MemoryListener *listener);
164 static void tcg_commit(MemoryListener *listener);
167 * CPUAddressSpace: all the information a CPU needs about an AddressSpace
168 * @cpu: the CPU whose AddressSpace this is
169 * @as: the AddressSpace itself
170 * @memory_dispatch: its dispatch pointer (cached, RCU protected)
171 * @tcg_as_listener: listener for tracking changes to the AddressSpace
173 struct CPUAddressSpace {
176 struct AddressSpaceDispatch *memory_dispatch;
177 MemoryListener tcg_as_listener;
180 struct DirtyBitmapSnapshot {
183 unsigned long dirty[];
188 #if !defined(CONFIG_USER_ONLY)
190 static void phys_map_node_reserve(PhysPageMap *map, unsigned nodes)
192 static unsigned alloc_hint = 16;
193 if (map->nodes_nb + nodes > map->nodes_nb_alloc) {
194 map->nodes_nb_alloc = MAX(alloc_hint, map->nodes_nb + nodes);
195 map->nodes = g_renew(Node, map->nodes, map->nodes_nb_alloc);
196 alloc_hint = map->nodes_nb_alloc;
200 static uint32_t phys_map_node_alloc(PhysPageMap *map, bool leaf)
207 ret = map->nodes_nb++;
209 assert(ret != PHYS_MAP_NODE_NIL);
210 assert(ret != map->nodes_nb_alloc);
212 e.skip = leaf ? 0 : 1;
213 e.ptr = leaf ? PHYS_SECTION_UNASSIGNED : PHYS_MAP_NODE_NIL;
214 for (i = 0; i < P_L2_SIZE; ++i) {
215 memcpy(&p[i], &e, sizeof(e));
220 static void phys_page_set_level(PhysPageMap *map, PhysPageEntry *lp,
221 hwaddr *index, uint64_t *nb, uint16_t leaf,
225 hwaddr step = (hwaddr)1 << (level * P_L2_BITS);
227 if (lp->skip && lp->ptr == PHYS_MAP_NODE_NIL) {
228 lp->ptr = phys_map_node_alloc(map, level == 0);
230 p = map->nodes[lp->ptr];
231 lp = &p[(*index >> (level * P_L2_BITS)) & (P_L2_SIZE - 1)];
233 while (*nb && lp < &p[P_L2_SIZE]) {
234 if ((*index & (step - 1)) == 0 && *nb >= step) {
240 phys_page_set_level(map, lp, index, nb, leaf, level - 1);
246 static void phys_page_set(AddressSpaceDispatch *d,
247 hwaddr index, uint64_t nb,
250 /* Wildly overreserve - it doesn't matter much. */
251 phys_map_node_reserve(&d->map, 3 * P_L2_LEVELS);
253 phys_page_set_level(&d->map, &d->phys_map, &index, &nb, leaf, P_L2_LEVELS - 1);
256 /* Compact a non leaf page entry. Simply detect that the entry has a single child,
257 * and update our entry so we can skip it and go directly to the destination.
259 static void phys_page_compact(PhysPageEntry *lp, Node *nodes)
261 unsigned valid_ptr = P_L2_SIZE;
266 if (lp->ptr == PHYS_MAP_NODE_NIL) {
271 for (i = 0; i < P_L2_SIZE; i++) {
272 if (p[i].ptr == PHYS_MAP_NODE_NIL) {
279 phys_page_compact(&p[i], nodes);
283 /* We can only compress if there's only one child. */
288 assert(valid_ptr < P_L2_SIZE);
290 /* Don't compress if it won't fit in the # of bits we have. */
291 if (P_L2_LEVELS >= (1 << 6) &&
292 lp->skip + p[valid_ptr].skip >= (1 << 6)) {
296 lp->ptr = p[valid_ptr].ptr;
297 if (!p[valid_ptr].skip) {
298 /* If our only child is a leaf, make this a leaf. */
299 /* By design, we should have made this node a leaf to begin with so we
300 * should never reach here.
301 * But since it's so simple to handle this, let's do it just in case we
306 lp->skip += p[valid_ptr].skip;
310 void address_space_dispatch_compact(AddressSpaceDispatch *d)
312 if (d->phys_map.skip) {
313 phys_page_compact(&d->phys_map, d->map.nodes);
317 static inline bool section_covers_addr(const MemoryRegionSection *section,
320 /* Memory topology clips a memory region to [0, 2^64); size.hi > 0 means
321 * the section must cover the entire address space.
323 return int128_gethi(section->size) ||
324 range_covers_byte(section->offset_within_address_space,
325 int128_getlo(section->size), addr);
328 static MemoryRegionSection *phys_page_find(AddressSpaceDispatch *d, hwaddr addr)
330 PhysPageEntry lp = d->phys_map, *p;
331 Node *nodes = d->map.nodes;
332 MemoryRegionSection *sections = d->map.sections;
333 hwaddr index = addr >> TARGET_PAGE_BITS;
336 for (i = P_L2_LEVELS; lp.skip && (i -= lp.skip) >= 0;) {
337 if (lp.ptr == PHYS_MAP_NODE_NIL) {
338 return §ions[PHYS_SECTION_UNASSIGNED];
341 lp = p[(index >> (i * P_L2_BITS)) & (P_L2_SIZE - 1)];
344 if (section_covers_addr(§ions[lp.ptr], addr)) {
345 return §ions[lp.ptr];
347 return §ions[PHYS_SECTION_UNASSIGNED];
351 /* Called from RCU critical section */
352 static MemoryRegionSection *address_space_lookup_region(AddressSpaceDispatch *d,
354 bool resolve_subpage)
356 MemoryRegionSection *section = atomic_read(&d->mru_section);
359 if (!section || section == &d->map.sections[PHYS_SECTION_UNASSIGNED] ||
360 !section_covers_addr(section, addr)) {
361 section = phys_page_find(d, addr);
362 atomic_set(&d->mru_section, section);
364 if (resolve_subpage && section->mr->subpage) {
365 subpage = container_of(section->mr, subpage_t, iomem);
366 section = &d->map.sections[subpage->sub_section[SUBPAGE_IDX(addr)]];
371 /* Called from RCU critical section */
372 static MemoryRegionSection *
373 address_space_translate_internal(AddressSpaceDispatch *d, hwaddr addr, hwaddr *xlat,
374 hwaddr *plen, bool resolve_subpage)
376 MemoryRegionSection *section;
380 section = address_space_lookup_region(d, addr, resolve_subpage);
381 /* Compute offset within MemoryRegionSection */
382 addr -= section->offset_within_address_space;
384 /* Compute offset within MemoryRegion */
385 *xlat = addr + section->offset_within_region;
389 /* MMIO registers can be expected to perform full-width accesses based only
390 * on their address, without considering adjacent registers that could
391 * decode to completely different MemoryRegions. When such registers
392 * exist (e.g. I/O ports 0xcf8 and 0xcf9 on most PC chipsets), MMIO
393 * regions overlap wildly. For this reason we cannot clamp the accesses
396 * If the length is small (as is the case for address_space_ldl/stl),
397 * everything works fine. If the incoming length is large, however,
398 * the caller really has to do the clamping through memory_access_size.
400 if (memory_region_is_ram(mr)) {
401 diff = int128_sub(section->size, int128_make64(addr));
402 *plen = int128_get64(int128_min(diff, int128_make64(*plen)));
408 * address_space_translate_iommu - translate an address through an IOMMU
409 * memory region and then through the target address space.
411 * @iommu_mr: the IOMMU memory region that we start the translation from
412 * @addr: the address to be translated through the MMU
413 * @xlat: the translated address offset within the destination memory region.
414 * It cannot be %NULL.
415 * @plen_out: valid read/write length of the translated address. It
417 * @page_mask_out: page mask for the translated address. This
418 * should only be meaningful for IOMMU translated
419 * addresses, since there may be huge pages that this bit
420 * would tell. It can be %NULL if we don't care about it.
421 * @is_write: whether the translation operation is for write
422 * @is_mmio: whether this can be MMIO, set true if it can
423 * @target_as: the address space targeted by the IOMMU
424 * @attrs: transaction attributes
426 * This function is called from RCU critical section. It is the common
427 * part of flatview_do_translate and address_space_translate_cached.
429 static MemoryRegionSection address_space_translate_iommu(IOMMUMemoryRegion *iommu_mr,
432 hwaddr *page_mask_out,
435 AddressSpace **target_as,
438 MemoryRegionSection *section;
439 hwaddr page_mask = (hwaddr)-1;
443 IOMMUMemoryRegionClass *imrc = memory_region_get_iommu_class_nocheck(iommu_mr);
447 if (imrc->attrs_to_index) {
448 iommu_idx = imrc->attrs_to_index(iommu_mr, attrs);
451 iotlb = imrc->translate(iommu_mr, addr, is_write ?
452 IOMMU_WO : IOMMU_RO, iommu_idx);
454 if (!(iotlb.perm & (1 << is_write))) {
458 addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
459 | (addr & iotlb.addr_mask));
460 page_mask &= iotlb.addr_mask;
461 *plen_out = MIN(*plen_out, (addr | iotlb.addr_mask) - addr + 1);
462 *target_as = iotlb.target_as;
464 section = address_space_translate_internal(
465 address_space_to_dispatch(iotlb.target_as), addr, xlat,
468 iommu_mr = memory_region_get_iommu(section->mr);
469 } while (unlikely(iommu_mr));
472 *page_mask_out = page_mask;
477 return (MemoryRegionSection) { .mr = &io_mem_unassigned };
481 * flatview_do_translate - translate an address in FlatView
483 * @fv: the flat view that we want to translate on
484 * @addr: the address to be translated in above address space
485 * @xlat: the translated address offset within memory region. It
487 * @plen_out: valid read/write length of the translated address. It
488 * can be @NULL when we don't care about it.
489 * @page_mask_out: page mask for the translated address. This
490 * should only be meaningful for IOMMU translated
491 * addresses, since there may be huge pages that this bit
492 * would tell. It can be @NULL if we don't care about it.
493 * @is_write: whether the translation operation is for write
494 * @is_mmio: whether this can be MMIO, set true if it can
495 * @target_as: the address space targeted by the IOMMU
496 * @attrs: memory transaction attributes
498 * This function is called from RCU critical section
500 static MemoryRegionSection flatview_do_translate(FlatView *fv,
504 hwaddr *page_mask_out,
507 AddressSpace **target_as,
510 MemoryRegionSection *section;
511 IOMMUMemoryRegion *iommu_mr;
512 hwaddr plen = (hwaddr)(-1);
518 section = address_space_translate_internal(
519 flatview_to_dispatch(fv), addr, xlat,
522 iommu_mr = memory_region_get_iommu(section->mr);
523 if (unlikely(iommu_mr)) {
524 return address_space_translate_iommu(iommu_mr, xlat,
525 plen_out, page_mask_out,
530 /* Not behind an IOMMU, use default page size. */
531 *page_mask_out = ~TARGET_PAGE_MASK;
537 /* Called from RCU critical section */
538 IOMMUTLBEntry address_space_get_iotlb_entry(AddressSpace *as, hwaddr addr,
539 bool is_write, MemTxAttrs attrs)
541 MemoryRegionSection section;
542 hwaddr xlat, page_mask;
545 * This can never be MMIO, and we don't really care about plen,
548 section = flatview_do_translate(address_space_to_flatview(as), addr, &xlat,
549 NULL, &page_mask, is_write, false, &as,
552 /* Illegal translation */
553 if (section.mr == &io_mem_unassigned) {
557 /* Convert memory region offset into address space offset */
558 xlat += section.offset_within_address_space -
559 section.offset_within_region;
561 return (IOMMUTLBEntry) {
563 .iova = addr & ~page_mask,
564 .translated_addr = xlat & ~page_mask,
565 .addr_mask = page_mask,
566 /* IOTLBs are for DMAs, and DMA only allows on RAMs. */
571 return (IOMMUTLBEntry) {0};
574 /* Called from RCU critical section */
575 MemoryRegion *flatview_translate(FlatView *fv, hwaddr addr, hwaddr *xlat,
576 hwaddr *plen, bool is_write,
580 MemoryRegionSection section;
581 AddressSpace *as = NULL;
583 /* This can be MMIO, so setup MMIO bit. */
584 section = flatview_do_translate(fv, addr, xlat, plen, NULL,
585 is_write, true, &as, attrs);
588 if (xen_enabled() && memory_access_is_direct(mr, is_write)) {
589 hwaddr page = ((addr & TARGET_PAGE_MASK) + TARGET_PAGE_SIZE) - addr;
590 *plen = MIN(page, *plen);
596 typedef struct TCGIOMMUNotifier {
604 static void tcg_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
606 TCGIOMMUNotifier *notifier = container_of(n, TCGIOMMUNotifier, n);
608 if (!notifier->active) {
611 tlb_flush(notifier->cpu);
612 notifier->active = false;
613 /* We leave the notifier struct on the list to avoid reallocating it later.
614 * Generally the number of IOMMUs a CPU deals with will be small.
615 * In any case we can't unregister the iommu notifier from a notify
620 static void tcg_register_iommu_notifier(CPUState *cpu,
621 IOMMUMemoryRegion *iommu_mr,
624 /* Make sure this CPU has an IOMMU notifier registered for this
625 * IOMMU/IOMMU index combination, so that we can flush its TLB
626 * when the IOMMU tells us the mappings we've cached have changed.
628 MemoryRegion *mr = MEMORY_REGION(iommu_mr);
629 TCGIOMMUNotifier *notifier;
633 for (i = 0; i < cpu->iommu_notifiers->len; i++) {
634 notifier = g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier *, i);
635 if (notifier->mr == mr && notifier->iommu_idx == iommu_idx) {
639 if (i == cpu->iommu_notifiers->len) {
640 /* Not found, add a new entry at the end of the array */
641 cpu->iommu_notifiers = g_array_set_size(cpu->iommu_notifiers, i + 1);
642 notifier = g_new0(TCGIOMMUNotifier, 1);
643 g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier *, i) = notifier;
646 notifier->iommu_idx = iommu_idx;
648 /* Rather than trying to register interest in the specific part
649 * of the iommu's address space that we've accessed and then
650 * expand it later as subsequent accesses touch more of it, we
651 * just register interest in the whole thing, on the assumption
652 * that iommu reconfiguration will be rare.
654 iommu_notifier_init(¬ifier->n,
655 tcg_iommu_unmap_notify,
656 IOMMU_NOTIFIER_UNMAP,
660 ret = memory_region_register_iommu_notifier(notifier->mr, ¬ifier->n,
663 error_report_err(err);
668 if (!notifier->active) {
669 notifier->active = true;
673 static void tcg_iommu_free_notifier_list(CPUState *cpu)
675 /* Destroy the CPU's notifier list */
677 TCGIOMMUNotifier *notifier;
679 for (i = 0; i < cpu->iommu_notifiers->len; i++) {
680 notifier = g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier *, i);
681 memory_region_unregister_iommu_notifier(notifier->mr, ¬ifier->n);
684 g_array_free(cpu->iommu_notifiers, true);
687 /* Called from RCU critical section */
688 MemoryRegionSection *
689 address_space_translate_for_iotlb(CPUState *cpu, int asidx, hwaddr addr,
690 hwaddr *xlat, hwaddr *plen,
691 MemTxAttrs attrs, int *prot)
693 MemoryRegionSection *section;
694 IOMMUMemoryRegion *iommu_mr;
695 IOMMUMemoryRegionClass *imrc;
698 AddressSpaceDispatch *d = atomic_rcu_read(&cpu->cpu_ases[asidx].memory_dispatch);
701 section = address_space_translate_internal(d, addr, &addr, plen, false);
703 iommu_mr = memory_region_get_iommu(section->mr);
708 imrc = memory_region_get_iommu_class_nocheck(iommu_mr);
710 iommu_idx = imrc->attrs_to_index(iommu_mr, attrs);
711 tcg_register_iommu_notifier(cpu, iommu_mr, iommu_idx);
712 /* We need all the permissions, so pass IOMMU_NONE so the IOMMU
713 * doesn't short-cut its translation table walk.
715 iotlb = imrc->translate(iommu_mr, addr, IOMMU_NONE, iommu_idx);
716 addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
717 | (addr & iotlb.addr_mask));
718 /* Update the caller's prot bits to remove permissions the IOMMU
719 * is giving us a failure response for. If we get down to no
720 * permissions left at all we can give up now.
722 if (!(iotlb.perm & IOMMU_RO)) {
723 *prot &= ~(PAGE_READ | PAGE_EXEC);
725 if (!(iotlb.perm & IOMMU_WO)) {
726 *prot &= ~PAGE_WRITE;
733 d = flatview_to_dispatch(address_space_to_flatview(iotlb.target_as));
736 assert(!memory_region_is_iommu(section->mr));
741 return &d->map.sections[PHYS_SECTION_UNASSIGNED];
745 #if !defined(CONFIG_USER_ONLY)
747 static int cpu_common_post_load(void *opaque, int version_id)
749 CPUState *cpu = opaque;
751 /* 0x01 was CPU_INTERRUPT_EXIT. This line can be removed when the
752 version_id is increased. */
753 cpu->interrupt_request &= ~0x01;
756 /* loadvm has just updated the content of RAM, bypassing the
757 * usual mechanisms that ensure we flush TBs for writes to
758 * memory we've translated code from. So we must flush all TBs,
759 * which will now be stale.
766 static int cpu_common_pre_load(void *opaque)
768 CPUState *cpu = opaque;
770 cpu->exception_index = -1;
775 static bool cpu_common_exception_index_needed(void *opaque)
777 CPUState *cpu = opaque;
779 return tcg_enabled() && cpu->exception_index != -1;
782 static const VMStateDescription vmstate_cpu_common_exception_index = {
783 .name = "cpu_common/exception_index",
785 .minimum_version_id = 1,
786 .needed = cpu_common_exception_index_needed,
787 .fields = (VMStateField[]) {
788 VMSTATE_INT32(exception_index, CPUState),
789 VMSTATE_END_OF_LIST()
793 static bool cpu_common_crash_occurred_needed(void *opaque)
795 CPUState *cpu = opaque;
797 return cpu->crash_occurred;
800 static const VMStateDescription vmstate_cpu_common_crash_occurred = {
801 .name = "cpu_common/crash_occurred",
803 .minimum_version_id = 1,
804 .needed = cpu_common_crash_occurred_needed,
805 .fields = (VMStateField[]) {
806 VMSTATE_BOOL(crash_occurred, CPUState),
807 VMSTATE_END_OF_LIST()
811 const VMStateDescription vmstate_cpu_common = {
812 .name = "cpu_common",
814 .minimum_version_id = 1,
815 .pre_load = cpu_common_pre_load,
816 .post_load = cpu_common_post_load,
817 .fields = (VMStateField[]) {
818 VMSTATE_UINT32(halted, CPUState),
819 VMSTATE_UINT32(interrupt_request, CPUState),
820 VMSTATE_END_OF_LIST()
822 .subsections = (const VMStateDescription*[]) {
823 &vmstate_cpu_common_exception_index,
824 &vmstate_cpu_common_crash_occurred,
829 void cpu_address_space_init(CPUState *cpu, int asidx,
830 const char *prefix, MemoryRegion *mr)
832 CPUAddressSpace *newas;
833 AddressSpace *as = g_new0(AddressSpace, 1);
837 as_name = g_strdup_printf("%s-%d", prefix, cpu->cpu_index);
838 address_space_init(as, mr, as_name);
841 /* Target code should have set num_ases before calling us */
842 assert(asidx < cpu->num_ases);
845 /* address space 0 gets the convenience alias */
849 /* KVM cannot currently support multiple address spaces. */
850 assert(asidx == 0 || !kvm_enabled());
852 if (!cpu->cpu_ases) {
853 cpu->cpu_ases = g_new0(CPUAddressSpace, cpu->num_ases);
856 newas = &cpu->cpu_ases[asidx];
860 newas->tcg_as_listener.log_global_after_sync = tcg_log_global_after_sync;
861 newas->tcg_as_listener.commit = tcg_commit;
862 memory_listener_register(&newas->tcg_as_listener, as);
866 AddressSpace *cpu_get_address_space(CPUState *cpu, int asidx)
868 /* Return the AddressSpace corresponding to the specified index */
869 return cpu->cpu_ases[asidx].as;
873 void cpu_exec_unrealizefn(CPUState *cpu)
875 CPUClass *cc = CPU_GET_CLASS(cpu);
878 cpu_list_remove(cpu);
880 if (cc->vmsd != NULL) {
881 vmstate_unregister(NULL, cc->vmsd, cpu);
883 if (qdev_get_vmsd(DEVICE(cpu)) == NULL) {
884 vmstate_unregister(NULL, &vmstate_cpu_common, cpu);
886 #ifndef CONFIG_USER_ONLY
887 tcg_iommu_free_notifier_list(cpu);
891 Property cpu_common_props[] = {
892 #ifndef CONFIG_USER_ONLY
893 /* Create a memory property for softmmu CPU object,
894 * so users can wire up its memory. (This can't go in hw/core/cpu.c
895 * because that file is compiled only once for both user-mode
896 * and system builds.) The default if no link is set up is to use
897 * the system address space.
899 DEFINE_PROP_LINK("memory", CPUState, memory, TYPE_MEMORY_REGION,
902 DEFINE_PROP_BOOL("start-powered-off", CPUState, start_powered_off, false),
903 DEFINE_PROP_END_OF_LIST(),
906 void cpu_exec_initfn(CPUState *cpu)
911 #ifndef CONFIG_USER_ONLY
912 cpu->thread_id = qemu_get_thread_id();
913 cpu->memory = system_memory;
914 object_ref(OBJECT(cpu->memory));
918 void cpu_exec_realizefn(CPUState *cpu, Error **errp)
920 CPUClass *cc = CPU_GET_CLASS(cpu);
921 static bool tcg_target_initialized;
925 if (tcg_enabled() && !tcg_target_initialized) {
926 tcg_target_initialized = true;
927 cc->tcg_initialize();
931 qemu_plugin_vcpu_init_hook(cpu);
933 #ifdef CONFIG_USER_ONLY
934 assert(cc->vmsd == NULL);
935 #else /* !CONFIG_USER_ONLY */
936 if (qdev_get_vmsd(DEVICE(cpu)) == NULL) {
937 vmstate_register(NULL, cpu->cpu_index, &vmstate_cpu_common, cpu);
939 if (cc->vmsd != NULL) {
940 vmstate_register(NULL, cpu->cpu_index, cc->vmsd, cpu);
943 cpu->iommu_notifiers = g_array_new(false, true, sizeof(TCGIOMMUNotifier *));
947 const char *parse_cpu_option(const char *cpu_option)
951 gchar **model_pieces;
952 const char *cpu_type;
954 model_pieces = g_strsplit(cpu_option, ",", 2);
955 if (!model_pieces[0]) {
956 error_report("-cpu option cannot be empty");
960 oc = cpu_class_by_name(CPU_RESOLVING_TYPE, model_pieces[0]);
962 error_report("unable to find CPU model '%s'", model_pieces[0]);
963 g_strfreev(model_pieces);
967 cpu_type = object_class_get_name(oc);
969 cc->parse_features(cpu_type, model_pieces[1], &error_fatal);
970 g_strfreev(model_pieces);
974 #if defined(CONFIG_USER_ONLY)
975 void tb_invalidate_phys_addr(target_ulong addr)
978 tb_invalidate_phys_page_range(addr, addr + 1);
982 static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
984 tb_invalidate_phys_addr(pc);
987 void tb_invalidate_phys_addr(AddressSpace *as, hwaddr addr, MemTxAttrs attrs)
993 if (!tcg_enabled()) {
997 RCU_READ_LOCK_GUARD();
998 mr = address_space_translate(as, addr, &addr, &l, false, attrs);
999 if (!(memory_region_is_ram(mr)
1000 || memory_region_is_romd(mr))) {
1003 ram_addr = memory_region_get_ram_addr(mr) + addr;
1004 tb_invalidate_phys_page_range(ram_addr, ram_addr + 1);
1007 static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
1010 * There may not be a virtual to physical translation for the pc
1011 * right now, but there may exist cached TB for this pc.
1012 * Flush the whole TB cache to force re-translation of such TBs.
1013 * This is heavyweight, but we're debugging anyway.
1019 #ifndef CONFIG_USER_ONLY
1020 /* Add a watchpoint. */
1021 int cpu_watchpoint_insert(CPUState *cpu, vaddr addr, vaddr len,
1022 int flags, CPUWatchpoint **watchpoint)
1027 /* forbid ranges which are empty or run off the end of the address space */
1028 if (len == 0 || (addr + len - 1) < addr) {
1029 error_report("tried to set invalid watchpoint at %"
1030 VADDR_PRIx ", len=%" VADDR_PRIu, addr, len);
1033 wp = g_malloc(sizeof(*wp));
1039 /* keep all GDB-injected watchpoints in front */
1040 if (flags & BP_GDB) {
1041 QTAILQ_INSERT_HEAD(&cpu->watchpoints, wp, entry);
1043 QTAILQ_INSERT_TAIL(&cpu->watchpoints, wp, entry);
1046 in_page = -(addr | TARGET_PAGE_MASK);
1047 if (len <= in_page) {
1048 tlb_flush_page(cpu, addr);
1058 /* Remove a specific watchpoint. */
1059 int cpu_watchpoint_remove(CPUState *cpu, vaddr addr, vaddr len,
1064 QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
1065 if (addr == wp->vaddr && len == wp->len
1066 && flags == (wp->flags & ~BP_WATCHPOINT_HIT)) {
1067 cpu_watchpoint_remove_by_ref(cpu, wp);
1074 /* Remove a specific watchpoint by reference. */
1075 void cpu_watchpoint_remove_by_ref(CPUState *cpu, CPUWatchpoint *watchpoint)
1077 QTAILQ_REMOVE(&cpu->watchpoints, watchpoint, entry);
1079 tlb_flush_page(cpu, watchpoint->vaddr);
1084 /* Remove all matching watchpoints. */
1085 void cpu_watchpoint_remove_all(CPUState *cpu, int mask)
1087 CPUWatchpoint *wp, *next;
1089 QTAILQ_FOREACH_SAFE(wp, &cpu->watchpoints, entry, next) {
1090 if (wp->flags & mask) {
1091 cpu_watchpoint_remove_by_ref(cpu, wp);
1096 /* Return true if this watchpoint address matches the specified
1097 * access (ie the address range covered by the watchpoint overlaps
1098 * partially or completely with the address range covered by the
1101 static inline bool watchpoint_address_matches(CPUWatchpoint *wp,
1102 vaddr addr, vaddr len)
1104 /* We know the lengths are non-zero, but a little caution is
1105 * required to avoid errors in the case where the range ends
1106 * exactly at the top of the address space and so addr + len
1107 * wraps round to zero.
1109 vaddr wpend = wp->vaddr + wp->len - 1;
1110 vaddr addrend = addr + len - 1;
1112 return !(addr > wpend || wp->vaddr > addrend);
1115 /* Return flags for watchpoints that match addr + prot. */
1116 int cpu_watchpoint_address_matches(CPUState *cpu, vaddr addr, vaddr len)
1121 QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
1122 if (watchpoint_address_matches(wp, addr, len)) {
1128 #endif /* !CONFIG_USER_ONLY */
1130 /* Add a breakpoint. */
1131 int cpu_breakpoint_insert(CPUState *cpu, vaddr pc, int flags,
1132 CPUBreakpoint **breakpoint)
1136 bp = g_malloc(sizeof(*bp));
1141 /* keep all GDB-injected breakpoints in front */
1142 if (flags & BP_GDB) {
1143 QTAILQ_INSERT_HEAD(&cpu->breakpoints, bp, entry);
1145 QTAILQ_INSERT_TAIL(&cpu->breakpoints, bp, entry);
1148 breakpoint_invalidate(cpu, pc);
1156 /* Remove a specific breakpoint. */
1157 int cpu_breakpoint_remove(CPUState *cpu, vaddr pc, int flags)
1161 QTAILQ_FOREACH(bp, &cpu->breakpoints, entry) {
1162 if (bp->pc == pc && bp->flags == flags) {
1163 cpu_breakpoint_remove_by_ref(cpu, bp);
1170 /* Remove a specific breakpoint by reference. */
1171 void cpu_breakpoint_remove_by_ref(CPUState *cpu, CPUBreakpoint *breakpoint)
1173 QTAILQ_REMOVE(&cpu->breakpoints, breakpoint, entry);
1175 breakpoint_invalidate(cpu, breakpoint->pc);
1180 /* Remove all matching breakpoints. */
1181 void cpu_breakpoint_remove_all(CPUState *cpu, int mask)
1183 CPUBreakpoint *bp, *next;
1185 QTAILQ_FOREACH_SAFE(bp, &cpu->breakpoints, entry, next) {
1186 if (bp->flags & mask) {
1187 cpu_breakpoint_remove_by_ref(cpu, bp);
1192 /* enable or disable single step mode. EXCP_DEBUG is returned by the
1193 CPU loop after each instruction */
1194 void cpu_single_step(CPUState *cpu, int enabled)
1196 if (cpu->singlestep_enabled != enabled) {
1197 cpu->singlestep_enabled = enabled;
1198 if (kvm_enabled()) {
1199 kvm_update_guest_debug(cpu, 0);
1201 /* must flush all the translated code to avoid inconsistencies */
1202 /* XXX: only flush what is necessary */
1208 void cpu_abort(CPUState *cpu, const char *fmt, ...)
1215 fprintf(stderr, "qemu: fatal: ");
1216 vfprintf(stderr, fmt, ap);
1217 fprintf(stderr, "\n");
1218 cpu_dump_state(cpu, stderr, CPU_DUMP_FPU | CPU_DUMP_CCOP);
1219 if (qemu_log_separate()) {
1220 FILE *logfile = qemu_log_lock();
1221 qemu_log("qemu: fatal: ");
1222 qemu_log_vprintf(fmt, ap2);
1224 log_cpu_state(cpu, CPU_DUMP_FPU | CPU_DUMP_CCOP);
1226 qemu_log_unlock(logfile);
1232 #if defined(CONFIG_USER_ONLY)
1234 struct sigaction act;
1235 sigfillset(&act.sa_mask);
1236 act.sa_handler = SIG_DFL;
1238 sigaction(SIGABRT, &act, NULL);
1244 #if !defined(CONFIG_USER_ONLY)
1245 /* Called from RCU critical section */
1246 static RAMBlock *qemu_get_ram_block(ram_addr_t addr)
1250 block = atomic_rcu_read(&ram_list.mru_block);
1251 if (block && addr - block->offset < block->max_length) {
1254 RAMBLOCK_FOREACH(block) {
1255 if (addr - block->offset < block->max_length) {
1260 fprintf(stderr, "Bad ram offset %" PRIx64 "\n", (uint64_t)addr);
1264 /* It is safe to write mru_block outside the iothread lock. This
1269 * xxx removed from list
1273 * call_rcu(reclaim_ramblock, xxx);
1276 * atomic_rcu_set is not needed here. The block was already published
1277 * when it was placed into the list. Here we're just making an extra
1278 * copy of the pointer.
1280 ram_list.mru_block = block;
1284 static void tlb_reset_dirty_range_all(ram_addr_t start, ram_addr_t length)
1291 assert(tcg_enabled());
1292 end = TARGET_PAGE_ALIGN(start + length);
1293 start &= TARGET_PAGE_MASK;
1295 RCU_READ_LOCK_GUARD();
1296 block = qemu_get_ram_block(start);
1297 assert(block == qemu_get_ram_block(end - 1));
1298 start1 = (uintptr_t)ramblock_ptr(block, start - block->offset);
1300 tlb_reset_dirty(cpu, start1, length);
1304 /* Note: start and end must be within the same ram block. */
1305 bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start,
1309 DirtyMemoryBlocks *blocks;
1310 unsigned long end, page, start_page;
1313 uint64_t mr_offset, mr_size;
1319 end = TARGET_PAGE_ALIGN(start + length) >> TARGET_PAGE_BITS;
1320 start_page = start >> TARGET_PAGE_BITS;
1323 WITH_RCU_READ_LOCK_GUARD() {
1324 blocks = atomic_rcu_read(&ram_list.dirty_memory[client]);
1325 ramblock = qemu_get_ram_block(start);
1326 /* Range sanity check on the ramblock */
1327 assert(start >= ramblock->offset &&
1328 start + length <= ramblock->offset + ramblock->used_length);
1330 while (page < end) {
1331 unsigned long idx = page / DIRTY_MEMORY_BLOCK_SIZE;
1332 unsigned long offset = page % DIRTY_MEMORY_BLOCK_SIZE;
1333 unsigned long num = MIN(end - page,
1334 DIRTY_MEMORY_BLOCK_SIZE - offset);
1336 dirty |= bitmap_test_and_clear_atomic(blocks->blocks[idx],
1341 mr_offset = (ram_addr_t)(start_page << TARGET_PAGE_BITS) - ramblock->offset;
1342 mr_size = (end - start_page) << TARGET_PAGE_BITS;
1343 memory_region_clear_dirty_bitmap(ramblock->mr, mr_offset, mr_size);
1346 if (dirty && tcg_enabled()) {
1347 tlb_reset_dirty_range_all(start, length);
1353 DirtyBitmapSnapshot *cpu_physical_memory_snapshot_and_clear_dirty
1354 (MemoryRegion *mr, hwaddr offset, hwaddr length, unsigned client)
1356 DirtyMemoryBlocks *blocks;
1357 ram_addr_t start = memory_region_get_ram_addr(mr) + offset;
1358 unsigned long align = 1UL << (TARGET_PAGE_BITS + BITS_PER_LEVEL);
1359 ram_addr_t first = QEMU_ALIGN_DOWN(start, align);
1360 ram_addr_t last = QEMU_ALIGN_UP(start + length, align);
1361 DirtyBitmapSnapshot *snap;
1362 unsigned long page, end, dest;
1364 snap = g_malloc0(sizeof(*snap) +
1365 ((last - first) >> (TARGET_PAGE_BITS + 3)));
1366 snap->start = first;
1369 page = first >> TARGET_PAGE_BITS;
1370 end = last >> TARGET_PAGE_BITS;
1373 WITH_RCU_READ_LOCK_GUARD() {
1374 blocks = atomic_rcu_read(&ram_list.dirty_memory[client]);
1376 while (page < end) {
1377 unsigned long idx = page / DIRTY_MEMORY_BLOCK_SIZE;
1378 unsigned long offset = page % DIRTY_MEMORY_BLOCK_SIZE;
1379 unsigned long num = MIN(end - page,
1380 DIRTY_MEMORY_BLOCK_SIZE - offset);
1382 assert(QEMU_IS_ALIGNED(offset, (1 << BITS_PER_LEVEL)));
1383 assert(QEMU_IS_ALIGNED(num, (1 << BITS_PER_LEVEL)));
1384 offset >>= BITS_PER_LEVEL;
1386 bitmap_copy_and_clear_atomic(snap->dirty + dest,
1387 blocks->blocks[idx] + offset,
1390 dest += num >> BITS_PER_LEVEL;
1394 if (tcg_enabled()) {
1395 tlb_reset_dirty_range_all(start, length);
1398 memory_region_clear_dirty_bitmap(mr, offset, length);
1403 bool cpu_physical_memory_snapshot_get_dirty(DirtyBitmapSnapshot *snap,
1407 unsigned long page, end;
1409 assert(start >= snap->start);
1410 assert(start + length <= snap->end);
1412 end = TARGET_PAGE_ALIGN(start + length - snap->start) >> TARGET_PAGE_BITS;
1413 page = (start - snap->start) >> TARGET_PAGE_BITS;
1415 while (page < end) {
1416 if (test_bit(page, snap->dirty)) {
1424 /* Called from RCU critical section */
1425 hwaddr memory_region_section_get_iotlb(CPUState *cpu,
1426 MemoryRegionSection *section)
1428 AddressSpaceDispatch *d = flatview_to_dispatch(section->fv);
1429 return section - d->map.sections;
1431 #endif /* defined(CONFIG_USER_ONLY) */
1433 #if !defined(CONFIG_USER_ONLY)
1435 static int subpage_register(subpage_t *mmio, uint32_t start, uint32_t end,
1437 static subpage_t *subpage_init(FlatView *fv, hwaddr base);
1439 static void *(*phys_mem_alloc)(size_t size, uint64_t *align, bool shared) =
1440 qemu_anon_ram_alloc;
1443 * Set a custom physical guest memory alloator.
1444 * Accelerators with unusual needs may need this. Hopefully, we can
1445 * get rid of it eventually.
1447 void phys_mem_set_alloc(void *(*alloc)(size_t, uint64_t *align, bool shared))
1449 phys_mem_alloc = alloc;
1452 static uint16_t phys_section_add(PhysPageMap *map,
1453 MemoryRegionSection *section)
1455 /* The physical section number is ORed with a page-aligned
1456 * pointer to produce the iotlb entries. Thus it should
1457 * never overflow into the page-aligned value.
1459 assert(map->sections_nb < TARGET_PAGE_SIZE);
1461 if (map->sections_nb == map->sections_nb_alloc) {
1462 map->sections_nb_alloc = MAX(map->sections_nb_alloc * 2, 16);
1463 map->sections = g_renew(MemoryRegionSection, map->sections,
1464 map->sections_nb_alloc);
1466 map->sections[map->sections_nb] = *section;
1467 memory_region_ref(section->mr);
1468 return map->sections_nb++;
1471 static void phys_section_destroy(MemoryRegion *mr)
1473 bool have_sub_page = mr->subpage;
1475 memory_region_unref(mr);
1477 if (have_sub_page) {
1478 subpage_t *subpage = container_of(mr, subpage_t, iomem);
1479 object_unref(OBJECT(&subpage->iomem));
1484 static void phys_sections_free(PhysPageMap *map)
1486 while (map->sections_nb > 0) {
1487 MemoryRegionSection *section = &map->sections[--map->sections_nb];
1488 phys_section_destroy(section->mr);
1490 g_free(map->sections);
1494 static void register_subpage(FlatView *fv, MemoryRegionSection *section)
1496 AddressSpaceDispatch *d = flatview_to_dispatch(fv);
1498 hwaddr base = section->offset_within_address_space
1500 MemoryRegionSection *existing = phys_page_find(d, base);
1501 MemoryRegionSection subsection = {
1502 .offset_within_address_space = base,
1503 .size = int128_make64(TARGET_PAGE_SIZE),
1507 assert(existing->mr->subpage || existing->mr == &io_mem_unassigned);
1509 if (!(existing->mr->subpage)) {
1510 subpage = subpage_init(fv, base);
1512 subsection.mr = &subpage->iomem;
1513 phys_page_set(d, base >> TARGET_PAGE_BITS, 1,
1514 phys_section_add(&d->map, &subsection));
1516 subpage = container_of(existing->mr, subpage_t, iomem);
1518 start = section->offset_within_address_space & ~TARGET_PAGE_MASK;
1519 end = start + int128_get64(section->size) - 1;
1520 subpage_register(subpage, start, end,
1521 phys_section_add(&d->map, section));
1525 static void register_multipage(FlatView *fv,
1526 MemoryRegionSection *section)
1528 AddressSpaceDispatch *d = flatview_to_dispatch(fv);
1529 hwaddr start_addr = section->offset_within_address_space;
1530 uint16_t section_index = phys_section_add(&d->map, section);
1531 uint64_t num_pages = int128_get64(int128_rshift(section->size,
1535 phys_page_set(d, start_addr >> TARGET_PAGE_BITS, num_pages, section_index);
1539 * The range in *section* may look like this:
1543 * where s stands for subpage and P for page.
1545 void flatview_add_to_dispatch(FlatView *fv, MemoryRegionSection *section)
1547 MemoryRegionSection remain = *section;
1548 Int128 page_size = int128_make64(TARGET_PAGE_SIZE);
1550 /* register first subpage */
1551 if (remain.offset_within_address_space & ~TARGET_PAGE_MASK) {
1552 uint64_t left = TARGET_PAGE_ALIGN(remain.offset_within_address_space)
1553 - remain.offset_within_address_space;
1555 MemoryRegionSection now = remain;
1556 now.size = int128_min(int128_make64(left), now.size);
1557 register_subpage(fv, &now);
1558 if (int128_eq(remain.size, now.size)) {
1561 remain.size = int128_sub(remain.size, now.size);
1562 remain.offset_within_address_space += int128_get64(now.size);
1563 remain.offset_within_region += int128_get64(now.size);
1566 /* register whole pages */
1567 if (int128_ge(remain.size, page_size)) {
1568 MemoryRegionSection now = remain;
1569 now.size = int128_and(now.size, int128_neg(page_size));
1570 register_multipage(fv, &now);
1571 if (int128_eq(remain.size, now.size)) {
1574 remain.size = int128_sub(remain.size, now.size);
1575 remain.offset_within_address_space += int128_get64(now.size);
1576 remain.offset_within_region += int128_get64(now.size);
1579 /* register last subpage */
1580 register_subpage(fv, &remain);
1583 void qemu_flush_coalesced_mmio_buffer(void)
1586 kvm_flush_coalesced_mmio_buffer();
1589 void qemu_mutex_lock_ramlist(void)
1591 qemu_mutex_lock(&ram_list.mutex);
1594 void qemu_mutex_unlock_ramlist(void)
1596 qemu_mutex_unlock(&ram_list.mutex);
1599 void ram_block_dump(Monitor *mon)
1604 RCU_READ_LOCK_GUARD();
1605 monitor_printf(mon, "%24s %8s %18s %18s %18s\n",
1606 "Block Name", "PSize", "Offset", "Used", "Total");
1607 RAMBLOCK_FOREACH(block) {
1608 psize = size_to_str(block->page_size);
1609 monitor_printf(mon, "%24s %8s 0x%016" PRIx64 " 0x%016" PRIx64
1610 " 0x%016" PRIx64 "\n", block->idstr, psize,
1611 (uint64_t)block->offset,
1612 (uint64_t)block->used_length,
1613 (uint64_t)block->max_length);
1620 * FIXME TOCTTOU: this iterates over memory backends' mem-path, which
1621 * may or may not name the same files / on the same filesystem now as
1622 * when we actually open and map them. Iterate over the file
1623 * descriptors instead, and use qemu_fd_getpagesize().
1625 static int find_min_backend_pagesize(Object *obj, void *opaque)
1627 long *hpsize_min = opaque;
1629 if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
1630 HostMemoryBackend *backend = MEMORY_BACKEND(obj);
1631 long hpsize = host_memory_backend_pagesize(backend);
1633 if (host_memory_backend_is_mapped(backend) && (hpsize < *hpsize_min)) {
1634 *hpsize_min = hpsize;
1641 static int find_max_backend_pagesize(Object *obj, void *opaque)
1643 long *hpsize_max = opaque;
1645 if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
1646 HostMemoryBackend *backend = MEMORY_BACKEND(obj);
1647 long hpsize = host_memory_backend_pagesize(backend);
1649 if (host_memory_backend_is_mapped(backend) && (hpsize > *hpsize_max)) {
1650 *hpsize_max = hpsize;
1658 * TODO: We assume right now that all mapped host memory backends are
1659 * used as RAM, however some might be used for different purposes.
1661 long qemu_minrampagesize(void)
1663 long hpsize = LONG_MAX;
1664 Object *memdev_root = object_resolve_path("/objects", NULL);
1666 object_child_foreach(memdev_root, find_min_backend_pagesize, &hpsize);
1670 long qemu_maxrampagesize(void)
1673 Object *memdev_root = object_resolve_path("/objects", NULL);
1675 object_child_foreach(memdev_root, find_max_backend_pagesize, &pagesize);
1679 long qemu_minrampagesize(void)
1681 return qemu_real_host_page_size;
1683 long qemu_maxrampagesize(void)
1685 return qemu_real_host_page_size;
1690 static int64_t get_file_size(int fd)
1693 #if defined(__linux__)
1696 if (fstat(fd, &st) < 0) {
1700 /* Special handling for devdax character devices */
1701 if (S_ISCHR(st.st_mode)) {
1702 g_autofree char *subsystem_path = NULL;
1703 g_autofree char *subsystem = NULL;
1705 subsystem_path = g_strdup_printf("/sys/dev/char/%d:%d/subsystem",
1706 major(st.st_rdev), minor(st.st_rdev));
1707 subsystem = g_file_read_link(subsystem_path, NULL);
1709 if (subsystem && g_str_has_suffix(subsystem, "/dax")) {
1710 g_autofree char *size_path = NULL;
1711 g_autofree char *size_str = NULL;
1713 size_path = g_strdup_printf("/sys/dev/char/%d:%d/size",
1714 major(st.st_rdev), minor(st.st_rdev));
1716 if (g_file_get_contents(size_path, &size_str, NULL, NULL)) {
1717 return g_ascii_strtoll(size_str, NULL, 0);
1721 #endif /* defined(__linux__) */
1723 /* st.st_size may be zero for special files yet lseek(2) works */
1724 size = lseek(fd, 0, SEEK_END);
1731 static int64_t get_file_align(int fd)
1734 #if defined(__linux__) && defined(CONFIG_LIBDAXCTL)
1737 if (fstat(fd, &st) < 0) {
1741 /* Special handling for devdax character devices */
1742 if (S_ISCHR(st.st_mode)) {
1743 g_autofree char *path = NULL;
1744 g_autofree char *rpath = NULL;
1745 struct daxctl_ctx *ctx;
1746 struct daxctl_region *region;
1749 path = g_strdup_printf("/sys/dev/char/%d:%d",
1750 major(st.st_rdev), minor(st.st_rdev));
1751 rpath = realpath(path, NULL);
1753 rc = daxctl_new(&ctx);
1758 daxctl_region_foreach(ctx, region) {
1759 if (strstr(rpath, daxctl_region_get_path(region))) {
1760 align = daxctl_region_get_align(region);
1766 #endif /* defined(__linux__) && defined(CONFIG_LIBDAXCTL) */
1771 static int file_ram_open(const char *path,
1772 const char *region_name,
1777 char *sanitized_name;
1783 fd = open(path, O_RDWR);
1785 /* @path names an existing file, use it */
1788 if (errno == ENOENT) {
1789 /* @path names a file that doesn't exist, create it */
1790 fd = open(path, O_RDWR | O_CREAT | O_EXCL, 0644);
1795 } else if (errno == EISDIR) {
1796 /* @path names a directory, create a file there */
1797 /* Make name safe to use with mkstemp by replacing '/' with '_'. */
1798 sanitized_name = g_strdup(region_name);
1799 for (c = sanitized_name; *c != '\0'; c++) {
1805 filename = g_strdup_printf("%s/qemu_back_mem.%s.XXXXXX", path,
1807 g_free(sanitized_name);
1809 fd = mkstemp(filename);
1817 if (errno != EEXIST && errno != EINTR) {
1818 error_setg_errno(errp, errno,
1819 "can't open backing store %s for guest RAM",
1824 * Try again on EINTR and EEXIST. The latter happens when
1825 * something else creates the file between our two open().
1832 static void *file_ram_alloc(RAMBlock *block,
1840 block->page_size = qemu_fd_getpagesize(fd);
1841 if (block->mr->align % block->page_size) {
1842 error_setg(errp, "alignment 0x%" PRIx64
1843 " must be multiples of page size 0x%zx",
1844 block->mr->align, block->page_size);
1846 } else if (block->mr->align && !is_power_of_2(block->mr->align)) {
1847 error_setg(errp, "alignment 0x%" PRIx64
1848 " must be a power of two", block->mr->align);
1851 block->mr->align = MAX(block->page_size, block->mr->align);
1852 #if defined(__s390x__)
1853 if (kvm_enabled()) {
1854 block->mr->align = MAX(block->mr->align, QEMU_VMALLOC_ALIGN);
1858 if (memory < block->page_size) {
1859 error_setg(errp, "memory size 0x" RAM_ADDR_FMT " must be equal to "
1860 "or larger than page size 0x%zx",
1861 memory, block->page_size);
1865 memory = ROUND_UP(memory, block->page_size);
1868 * ftruncate is not supported by hugetlbfs in older
1869 * hosts, so don't bother bailing out on errors.
1870 * If anything goes wrong with it under other filesystems,
1873 * Do not truncate the non-empty backend file to avoid corrupting
1874 * the existing data in the file. Disabling shrinking is not
1875 * enough. For example, the current vNVDIMM implementation stores
1876 * the guest NVDIMM labels at the end of the backend file. If the
1877 * backend file is later extended, QEMU will not be able to find
1878 * those labels. Therefore, extending the non-empty backend file
1879 * is disabled as well.
1881 if (truncate && ftruncate(fd, memory)) {
1882 perror("ftruncate");
1885 area = qemu_ram_mmap(fd, memory, block->mr->align,
1886 block->flags & RAM_SHARED, block->flags & RAM_PMEM);
1887 if (area == MAP_FAILED) {
1888 error_setg_errno(errp, errno,
1889 "unable to map backing store for guest RAM");
1898 /* Allocate space within the ram_addr_t space that governs the
1900 * Called with the ramlist lock held.
1902 static ram_addr_t find_ram_offset(ram_addr_t size)
1904 RAMBlock *block, *next_block;
1905 ram_addr_t offset = RAM_ADDR_MAX, mingap = RAM_ADDR_MAX;
1907 assert(size != 0); /* it would hand out same offset multiple times */
1909 if (QLIST_EMPTY_RCU(&ram_list.blocks)) {
1913 RAMBLOCK_FOREACH(block) {
1914 ram_addr_t candidate, next = RAM_ADDR_MAX;
1916 /* Align blocks to start on a 'long' in the bitmap
1917 * which makes the bitmap sync'ing take the fast path.
1919 candidate = block->offset + block->max_length;
1920 candidate = ROUND_UP(candidate, BITS_PER_LONG << TARGET_PAGE_BITS);
1922 /* Search for the closest following block
1925 RAMBLOCK_FOREACH(next_block) {
1926 if (next_block->offset >= candidate) {
1927 next = MIN(next, next_block->offset);
1931 /* If it fits remember our place and remember the size
1932 * of gap, but keep going so that we might find a smaller
1933 * gap to fill so avoiding fragmentation.
1935 if (next - candidate >= size && next - candidate < mingap) {
1937 mingap = next - candidate;
1940 trace_find_ram_offset_loop(size, candidate, offset, next, mingap);
1943 if (offset == RAM_ADDR_MAX) {
1944 fprintf(stderr, "Failed to find gap of requested size: %" PRIu64 "\n",
1949 trace_find_ram_offset(size, offset);
1954 static unsigned long last_ram_page(void)
1957 ram_addr_t last = 0;
1959 RCU_READ_LOCK_GUARD();
1960 RAMBLOCK_FOREACH(block) {
1961 last = MAX(last, block->offset + block->max_length);
1963 return last >> TARGET_PAGE_BITS;
1966 static void qemu_ram_setup_dump(void *addr, ram_addr_t size)
1970 /* Use MADV_DONTDUMP, if user doesn't want the guest memory in the core */
1971 if (!machine_dump_guest_core(current_machine)) {
1972 ret = qemu_madvise(addr, size, QEMU_MADV_DONTDUMP);
1974 perror("qemu_madvise");
1975 fprintf(stderr, "madvise doesn't support MADV_DONTDUMP, "
1976 "but dump_guest_core=off specified\n");
1981 const char *qemu_ram_get_idstr(RAMBlock *rb)
1986 void *qemu_ram_get_host_addr(RAMBlock *rb)
1991 ram_addr_t qemu_ram_get_offset(RAMBlock *rb)
1996 ram_addr_t qemu_ram_get_used_length(RAMBlock *rb)
1998 return rb->used_length;
2001 bool qemu_ram_is_shared(RAMBlock *rb)
2003 return rb->flags & RAM_SHARED;
2006 /* Note: Only set at the start of postcopy */
2007 bool qemu_ram_is_uf_zeroable(RAMBlock *rb)
2009 return rb->flags & RAM_UF_ZEROPAGE;
2012 void qemu_ram_set_uf_zeroable(RAMBlock *rb)
2014 rb->flags |= RAM_UF_ZEROPAGE;
2017 bool qemu_ram_is_migratable(RAMBlock *rb)
2019 return rb->flags & RAM_MIGRATABLE;
2022 void qemu_ram_set_migratable(RAMBlock *rb)
2024 rb->flags |= RAM_MIGRATABLE;
2027 void qemu_ram_unset_migratable(RAMBlock *rb)
2029 rb->flags &= ~RAM_MIGRATABLE;
2032 /* Called with iothread lock held. */
2033 void qemu_ram_set_idstr(RAMBlock *new_block, const char *name, DeviceState *dev)
2038 assert(!new_block->idstr[0]);
2041 char *id = qdev_get_dev_path(dev);
2043 snprintf(new_block->idstr, sizeof(new_block->idstr), "%s/", id);
2047 pstrcat(new_block->idstr, sizeof(new_block->idstr), name);
2049 RCU_READ_LOCK_GUARD();
2050 RAMBLOCK_FOREACH(block) {
2051 if (block != new_block &&
2052 !strcmp(block->idstr, new_block->idstr)) {
2053 fprintf(stderr, "RAMBlock \"%s\" already registered, abort!\n",
2060 /* Called with iothread lock held. */
2061 void qemu_ram_unset_idstr(RAMBlock *block)
2063 /* FIXME: arch_init.c assumes that this is not called throughout
2064 * migration. Ignore the problem since hot-unplug during migration
2065 * does not work anyway.
2068 memset(block->idstr, 0, sizeof(block->idstr));
2072 size_t qemu_ram_pagesize(RAMBlock *rb)
2074 return rb->page_size;
2077 /* Returns the largest size of page in use */
2078 size_t qemu_ram_pagesize_largest(void)
2083 RAMBLOCK_FOREACH(block) {
2084 largest = MAX(largest, qemu_ram_pagesize(block));
2090 static int memory_try_enable_merging(void *addr, size_t len)
2092 if (!machine_mem_merge(current_machine)) {
2093 /* disabled by the user */
2097 return qemu_madvise(addr, len, QEMU_MADV_MERGEABLE);
2100 /* Only legal before guest might have detected the memory size: e.g. on
2101 * incoming migration, or right after reset.
2103 * As memory core doesn't know how is memory accessed, it is up to
2104 * resize callback to update device state and/or add assertions to detect
2105 * misuse, if necessary.
2107 int qemu_ram_resize(RAMBlock *block, ram_addr_t newsize, Error **errp)
2109 const ram_addr_t unaligned_size = newsize;
2113 newsize = HOST_PAGE_ALIGN(newsize);
2115 if (block->used_length == newsize) {
2117 * We don't have to resize the ram block (which only knows aligned
2118 * sizes), however, we have to notify if the unaligned size changed.
2120 if (unaligned_size != memory_region_size(block->mr)) {
2121 memory_region_set_size(block->mr, unaligned_size);
2122 if (block->resized) {
2123 block->resized(block->idstr, unaligned_size, block->host);
2129 if (!(block->flags & RAM_RESIZEABLE)) {
2130 error_setg_errno(errp, EINVAL,
2131 "Length mismatch: %s: 0x" RAM_ADDR_FMT
2132 " in != 0x" RAM_ADDR_FMT, block->idstr,
2133 newsize, block->used_length);
2137 if (block->max_length < newsize) {
2138 error_setg_errno(errp, EINVAL,
2139 "Length too large: %s: 0x" RAM_ADDR_FMT
2140 " > 0x" RAM_ADDR_FMT, block->idstr,
2141 newsize, block->max_length);
2145 cpu_physical_memory_clear_dirty_range(block->offset, block->used_length);
2146 block->used_length = newsize;
2147 cpu_physical_memory_set_dirty_range(block->offset, block->used_length,
2149 memory_region_set_size(block->mr, unaligned_size);
2150 if (block->resized) {
2151 block->resized(block->idstr, unaligned_size, block->host);
2157 * Trigger sync on the given ram block for range [start, start + length]
2158 * with the backing store if one is available.
2160 * @Note: this is supposed to be a synchronous op.
2162 void qemu_ram_msync(RAMBlock *block, ram_addr_t start, ram_addr_t length)
2164 /* The requested range should fit in within the block range */
2165 g_assert((start + length) <= block->used_length);
2167 #ifdef CONFIG_LIBPMEM
2168 /* The lack of support for pmem should not block the sync */
2169 if (ramblock_is_pmem(block)) {
2170 void *addr = ramblock_ptr(block, start);
2171 pmem_persist(addr, length);
2175 if (block->fd >= 0) {
2177 * Case there is no support for PMEM or the memory has not been
2178 * specified as persistent (or is not one) - use the msync.
2179 * Less optimal but still achieves the same goal
2181 void *addr = ramblock_ptr(block, start);
2182 if (qemu_msync(addr, length, block->fd)) {
2183 warn_report("%s: failed to sync memory range: start: "
2184 RAM_ADDR_FMT " length: " RAM_ADDR_FMT,
2185 __func__, start, length);
2190 /* Called with ram_list.mutex held */
2191 static void dirty_memory_extend(ram_addr_t old_ram_size,
2192 ram_addr_t new_ram_size)
2194 ram_addr_t old_num_blocks = DIV_ROUND_UP(old_ram_size,
2195 DIRTY_MEMORY_BLOCK_SIZE);
2196 ram_addr_t new_num_blocks = DIV_ROUND_UP(new_ram_size,
2197 DIRTY_MEMORY_BLOCK_SIZE);
2200 /* Only need to extend if block count increased */
2201 if (new_num_blocks <= old_num_blocks) {
2205 for (i = 0; i < DIRTY_MEMORY_NUM; i++) {
2206 DirtyMemoryBlocks *old_blocks;
2207 DirtyMemoryBlocks *new_blocks;
2210 old_blocks = atomic_rcu_read(&ram_list.dirty_memory[i]);
2211 new_blocks = g_malloc(sizeof(*new_blocks) +
2212 sizeof(new_blocks->blocks[0]) * new_num_blocks);
2214 if (old_num_blocks) {
2215 memcpy(new_blocks->blocks, old_blocks->blocks,
2216 old_num_blocks * sizeof(old_blocks->blocks[0]));
2219 for (j = old_num_blocks; j < new_num_blocks; j++) {
2220 new_blocks->blocks[j] = bitmap_new(DIRTY_MEMORY_BLOCK_SIZE);
2223 atomic_rcu_set(&ram_list.dirty_memory[i], new_blocks);
2226 g_free_rcu(old_blocks, rcu);
2231 static void ram_block_add(RAMBlock *new_block, Error **errp, bool shared)
2234 RAMBlock *last_block = NULL;
2235 ram_addr_t old_ram_size, new_ram_size;
2238 old_ram_size = last_ram_page();
2240 qemu_mutex_lock_ramlist();
2241 new_block->offset = find_ram_offset(new_block->max_length);
2243 if (!new_block->host) {
2244 if (xen_enabled()) {
2245 xen_ram_alloc(new_block->offset, new_block->max_length,
2246 new_block->mr, &err);
2248 error_propagate(errp, err);
2249 qemu_mutex_unlock_ramlist();
2253 new_block->host = phys_mem_alloc(new_block->max_length,
2254 &new_block->mr->align, shared);
2255 if (!new_block->host) {
2256 error_setg_errno(errp, errno,
2257 "cannot set up guest memory '%s'",
2258 memory_region_name(new_block->mr));
2259 qemu_mutex_unlock_ramlist();
2262 memory_try_enable_merging(new_block->host, new_block->max_length);
2266 new_ram_size = MAX(old_ram_size,
2267 (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS);
2268 if (new_ram_size > old_ram_size) {
2269 dirty_memory_extend(old_ram_size, new_ram_size);
2271 /* Keep the list sorted from biggest to smallest block. Unlike QTAILQ,
2272 * QLIST (which has an RCU-friendly variant) does not have insertion at
2273 * tail, so save the last element in last_block.
2275 RAMBLOCK_FOREACH(block) {
2277 if (block->max_length < new_block->max_length) {
2282 QLIST_INSERT_BEFORE_RCU(block, new_block, next);
2283 } else if (last_block) {
2284 QLIST_INSERT_AFTER_RCU(last_block, new_block, next);
2285 } else { /* list is empty */
2286 QLIST_INSERT_HEAD_RCU(&ram_list.blocks, new_block, next);
2288 ram_list.mru_block = NULL;
2290 /* Write list before version */
2293 qemu_mutex_unlock_ramlist();
2295 cpu_physical_memory_set_dirty_range(new_block->offset,
2296 new_block->used_length,
2299 if (new_block->host) {
2300 qemu_ram_setup_dump(new_block->host, new_block->max_length);
2301 qemu_madvise(new_block->host, new_block->max_length, QEMU_MADV_HUGEPAGE);
2303 * MADV_DONTFORK is also needed by KVM in absence of synchronous MMU
2304 * Configure it unless the machine is a qtest server, in which case
2305 * KVM is not used and it may be forked (eg for fuzzing purposes).
2307 if (!qtest_enabled()) {
2308 qemu_madvise(new_block->host, new_block->max_length,
2309 QEMU_MADV_DONTFORK);
2311 ram_block_notify_add(new_block->host, new_block->max_length);
2316 RAMBlock *qemu_ram_alloc_from_fd(ram_addr_t size, MemoryRegion *mr,
2317 uint32_t ram_flags, int fd,
2320 RAMBlock *new_block;
2321 Error *local_err = NULL;
2322 int64_t file_size, file_align;
2324 /* Just support these ram flags by now. */
2325 assert((ram_flags & ~(RAM_SHARED | RAM_PMEM)) == 0);
2327 if (xen_enabled()) {
2328 error_setg(errp, "-mem-path not supported with Xen");
2332 if (kvm_enabled() && !kvm_has_sync_mmu()) {
2334 "host lacks kvm mmu notifiers, -mem-path unsupported");
2338 if (phys_mem_alloc != qemu_anon_ram_alloc) {
2340 * file_ram_alloc() needs to allocate just like
2341 * phys_mem_alloc, but we haven't bothered to provide
2345 "-mem-path not supported with this accelerator");
2349 size = HOST_PAGE_ALIGN(size);
2350 file_size = get_file_size(fd);
2351 if (file_size > 0 && file_size < size) {
2352 error_setg(errp, "backing store size 0x%" PRIx64
2353 " does not match 'size' option 0x" RAM_ADDR_FMT,
2358 file_align = get_file_align(fd);
2359 if (file_align > 0 && mr && file_align > mr->align) {
2360 error_setg(errp, "backing store align 0x%" PRIx64
2361 " is larger than 'align' option 0x%" PRIx64,
2362 file_align, mr->align);
2366 new_block = g_malloc0(sizeof(*new_block));
2368 new_block->used_length = size;
2369 new_block->max_length = size;
2370 new_block->flags = ram_flags;
2371 new_block->host = file_ram_alloc(new_block, size, fd, !file_size, errp);
2372 if (!new_block->host) {
2377 ram_block_add(new_block, &local_err, ram_flags & RAM_SHARED);
2380 error_propagate(errp, local_err);
2388 RAMBlock *qemu_ram_alloc_from_file(ram_addr_t size, MemoryRegion *mr,
2389 uint32_t ram_flags, const char *mem_path,
2396 fd = file_ram_open(mem_path, memory_region_name(mr), &created, errp);
2401 block = qemu_ram_alloc_from_fd(size, mr, ram_flags, fd, errp);
2415 RAMBlock *qemu_ram_alloc_internal(ram_addr_t size, ram_addr_t max_size,
2416 void (*resized)(const char*,
2419 void *host, bool resizeable, bool share,
2420 MemoryRegion *mr, Error **errp)
2422 RAMBlock *new_block;
2423 Error *local_err = NULL;
2425 size = HOST_PAGE_ALIGN(size);
2426 max_size = HOST_PAGE_ALIGN(max_size);
2427 new_block = g_malloc0(sizeof(*new_block));
2429 new_block->resized = resized;
2430 new_block->used_length = size;
2431 new_block->max_length = max_size;
2432 assert(max_size >= size);
2434 new_block->page_size = qemu_real_host_page_size;
2435 new_block->host = host;
2437 new_block->flags |= RAM_PREALLOC;
2440 new_block->flags |= RAM_RESIZEABLE;
2442 ram_block_add(new_block, &local_err, share);
2445 error_propagate(errp, local_err);
2451 RAMBlock *qemu_ram_alloc_from_ptr(ram_addr_t size, void *host,
2452 MemoryRegion *mr, Error **errp)
2454 return qemu_ram_alloc_internal(size, size, NULL, host, false,
2458 RAMBlock *qemu_ram_alloc(ram_addr_t size, bool share,
2459 MemoryRegion *mr, Error **errp)
2461 return qemu_ram_alloc_internal(size, size, NULL, NULL, false,
2465 RAMBlock *qemu_ram_alloc_resizeable(ram_addr_t size, ram_addr_t maxsz,
2466 void (*resized)(const char*,
2469 MemoryRegion *mr, Error **errp)
2471 return qemu_ram_alloc_internal(size, maxsz, resized, NULL, true,
2475 static void reclaim_ramblock(RAMBlock *block)
2477 if (block->flags & RAM_PREALLOC) {
2479 } else if (xen_enabled()) {
2480 xen_invalidate_map_cache_entry(block->host);
2482 } else if (block->fd >= 0) {
2483 qemu_ram_munmap(block->fd, block->host, block->max_length);
2487 qemu_anon_ram_free(block->host, block->max_length);
2492 void qemu_ram_free(RAMBlock *block)
2499 ram_block_notify_remove(block->host, block->max_length);
2502 qemu_mutex_lock_ramlist();
2503 QLIST_REMOVE_RCU(block, next);
2504 ram_list.mru_block = NULL;
2505 /* Write list before version */
2508 call_rcu(block, reclaim_ramblock, rcu);
2509 qemu_mutex_unlock_ramlist();
2513 void qemu_ram_remap(ram_addr_t addr, ram_addr_t length)
2520 RAMBLOCK_FOREACH(block) {
2521 offset = addr - block->offset;
2522 if (offset < block->max_length) {
2523 vaddr = ramblock_ptr(block, offset);
2524 if (block->flags & RAM_PREALLOC) {
2526 } else if (xen_enabled()) {
2530 if (block->fd >= 0) {
2531 flags |= (block->flags & RAM_SHARED ?
2532 MAP_SHARED : MAP_PRIVATE);
2533 area = mmap(vaddr, length, PROT_READ | PROT_WRITE,
2534 flags, block->fd, offset);
2537 * Remap needs to match alloc. Accelerators that
2538 * set phys_mem_alloc never remap. If they did,
2539 * we'd need a remap hook here.
2541 assert(phys_mem_alloc == qemu_anon_ram_alloc);
2543 flags |= MAP_PRIVATE | MAP_ANONYMOUS;
2544 area = mmap(vaddr, length, PROT_READ | PROT_WRITE,
2547 if (area != vaddr) {
2548 error_report("Could not remap addr: "
2549 RAM_ADDR_FMT "@" RAM_ADDR_FMT "",
2553 memory_try_enable_merging(vaddr, length);
2554 qemu_ram_setup_dump(vaddr, length);
2559 #endif /* !_WIN32 */
2561 /* Return a host pointer to ram allocated with qemu_ram_alloc.
2562 * This should not be used for general purpose DMA. Use address_space_map
2563 * or address_space_rw instead. For local memory (e.g. video ram) that the
2564 * device owns, use memory_region_get_ram_ptr.
2566 * Called within RCU critical section.
2568 void *qemu_map_ram_ptr(RAMBlock *ram_block, ram_addr_t addr)
2570 RAMBlock *block = ram_block;
2572 if (block == NULL) {
2573 block = qemu_get_ram_block(addr);
2574 addr -= block->offset;
2577 if (xen_enabled() && block->host == NULL) {
2578 /* We need to check if the requested address is in the RAM
2579 * because we don't want to map the entire memory in QEMU.
2580 * In that case just map until the end of the page.
2582 if (block->offset == 0) {
2583 return xen_map_cache(addr, 0, 0, false);
2586 block->host = xen_map_cache(block->offset, block->max_length, 1, false);
2588 return ramblock_ptr(block, addr);
2591 /* Return a host pointer to guest's ram. Similar to qemu_map_ram_ptr
2592 * but takes a size argument.
2594 * Called within RCU critical section.
2596 static void *qemu_ram_ptr_length(RAMBlock *ram_block, ram_addr_t addr,
2597 hwaddr *size, bool lock)
2599 RAMBlock *block = ram_block;
2604 if (block == NULL) {
2605 block = qemu_get_ram_block(addr);
2606 addr -= block->offset;
2608 *size = MIN(*size, block->max_length - addr);
2610 if (xen_enabled() && block->host == NULL) {
2611 /* We need to check if the requested address is in the RAM
2612 * because we don't want to map the entire memory in QEMU.
2613 * In that case just map the requested area.
2615 if (block->offset == 0) {
2616 return xen_map_cache(addr, *size, lock, lock);
2619 block->host = xen_map_cache(block->offset, block->max_length, 1, lock);
2622 return ramblock_ptr(block, addr);
2625 /* Return the offset of a hostpointer within a ramblock */
2626 ram_addr_t qemu_ram_block_host_offset(RAMBlock *rb, void *host)
2628 ram_addr_t res = (uint8_t *)host - (uint8_t *)rb->host;
2629 assert((uintptr_t)host >= (uintptr_t)rb->host);
2630 assert(res < rb->max_length);
2636 * Translates a host ptr back to a RAMBlock, a ram_addr and an offset
2639 * ptr: Host pointer to look up
2640 * round_offset: If true round the result offset down to a page boundary
2641 * *ram_addr: set to result ram_addr
2642 * *offset: set to result offset within the RAMBlock
2644 * Returns: RAMBlock (or NULL if not found)
2646 * By the time this function returns, the returned pointer is not protected
2647 * by RCU anymore. If the caller is not within an RCU critical section and
2648 * does not hold the iothread lock, it must have other means of protecting the
2649 * pointer, such as a reference to the region that includes the incoming
2652 RAMBlock *qemu_ram_block_from_host(void *ptr, bool round_offset,
2656 uint8_t *host = ptr;
2658 if (xen_enabled()) {
2659 ram_addr_t ram_addr;
2660 RCU_READ_LOCK_GUARD();
2661 ram_addr = xen_ram_addr_from_mapcache(ptr);
2662 block = qemu_get_ram_block(ram_addr);
2664 *offset = ram_addr - block->offset;
2669 RCU_READ_LOCK_GUARD();
2670 block = atomic_rcu_read(&ram_list.mru_block);
2671 if (block && block->host && host - block->host < block->max_length) {
2675 RAMBLOCK_FOREACH(block) {
2676 /* This case append when the block is not mapped. */
2677 if (block->host == NULL) {
2680 if (host - block->host < block->max_length) {
2688 *offset = (host - block->host);
2690 *offset &= TARGET_PAGE_MASK;
2696 * Finds the named RAMBlock
2698 * name: The name of RAMBlock to find
2700 * Returns: RAMBlock (or NULL if not found)
2702 RAMBlock *qemu_ram_block_by_name(const char *name)
2706 RAMBLOCK_FOREACH(block) {
2707 if (!strcmp(name, block->idstr)) {
2715 /* Some of the softmmu routines need to translate from a host pointer
2716 (typically a TLB entry) back to a ram offset. */
2717 ram_addr_t qemu_ram_addr_from_host(void *ptr)
2722 block = qemu_ram_block_from_host(ptr, false, &offset);
2724 return RAM_ADDR_INVALID;
2727 return block->offset + offset;
2730 /* Generate a debug exception if a watchpoint has been hit. */
2731 void cpu_check_watchpoint(CPUState *cpu, vaddr addr, vaddr len,
2732 MemTxAttrs attrs, int flags, uintptr_t ra)
2734 CPUClass *cc = CPU_GET_CLASS(cpu);
2737 assert(tcg_enabled());
2738 if (cpu->watchpoint_hit) {
2740 * We re-entered the check after replacing the TB.
2741 * Now raise the debug interrupt so that it will
2742 * trigger after the current instruction.
2744 qemu_mutex_lock_iothread();
2745 cpu_interrupt(cpu, CPU_INTERRUPT_DEBUG);
2746 qemu_mutex_unlock_iothread();
2750 addr = cc->adjust_watchpoint_address(cpu, addr, len);
2751 QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
2752 if (watchpoint_address_matches(wp, addr, len)
2753 && (wp->flags & flags)) {
2754 if (flags == BP_MEM_READ) {
2755 wp->flags |= BP_WATCHPOINT_HIT_READ;
2757 wp->flags |= BP_WATCHPOINT_HIT_WRITE;
2759 wp->hitaddr = MAX(addr, wp->vaddr);
2760 wp->hitattrs = attrs;
2761 if (!cpu->watchpoint_hit) {
2762 if (wp->flags & BP_CPU &&
2763 !cc->debug_check_watchpoint(cpu, wp)) {
2764 wp->flags &= ~BP_WATCHPOINT_HIT;
2767 cpu->watchpoint_hit = wp;
2770 tb_check_watchpoint(cpu, ra);
2771 if (wp->flags & BP_STOP_BEFORE_ACCESS) {
2772 cpu->exception_index = EXCP_DEBUG;
2774 cpu_loop_exit_restore(cpu, ra);
2776 /* Force execution of one insn next time. */
2777 cpu->cflags_next_tb = 1 | curr_cflags();
2780 cpu_restore_state(cpu, ra, true);
2782 cpu_loop_exit_noexc(cpu);
2786 wp->flags &= ~BP_WATCHPOINT_HIT;
2791 static MemTxResult flatview_read(FlatView *fv, hwaddr addr,
2792 MemTxAttrs attrs, void *buf, hwaddr len);
2793 static MemTxResult flatview_write(FlatView *fv, hwaddr addr, MemTxAttrs attrs,
2794 const void *buf, hwaddr len);
2795 static bool flatview_access_valid(FlatView *fv, hwaddr addr, hwaddr len,
2796 bool is_write, MemTxAttrs attrs);
2798 static MemTxResult subpage_read(void *opaque, hwaddr addr, uint64_t *data,
2799 unsigned len, MemTxAttrs attrs)
2801 subpage_t *subpage = opaque;
2805 #if defined(DEBUG_SUBPAGE)
2806 printf("%s: subpage %p len %u addr " TARGET_FMT_plx "\n", __func__,
2807 subpage, len, addr);
2809 res = flatview_read(subpage->fv, addr + subpage->base, attrs, buf, len);
2813 *data = ldn_p(buf, len);
2817 static MemTxResult subpage_write(void *opaque, hwaddr addr,
2818 uint64_t value, unsigned len, MemTxAttrs attrs)
2820 subpage_t *subpage = opaque;
2823 #if defined(DEBUG_SUBPAGE)
2824 printf("%s: subpage %p len %u addr " TARGET_FMT_plx
2825 " value %"PRIx64"\n",
2826 __func__, subpage, len, addr, value);
2828 stn_p(buf, len, value);
2829 return flatview_write(subpage->fv, addr + subpage->base, attrs, buf, len);
2832 static bool subpage_accepts(void *opaque, hwaddr addr,
2833 unsigned len, bool is_write,
2836 subpage_t *subpage = opaque;
2837 #if defined(DEBUG_SUBPAGE)
2838 printf("%s: subpage %p %c len %u addr " TARGET_FMT_plx "\n",
2839 __func__, subpage, is_write ? 'w' : 'r', len, addr);
2842 return flatview_access_valid(subpage->fv, addr + subpage->base,
2843 len, is_write, attrs);
2846 static const MemoryRegionOps subpage_ops = {
2847 .read_with_attrs = subpage_read,
2848 .write_with_attrs = subpage_write,
2849 .impl.min_access_size = 1,
2850 .impl.max_access_size = 8,
2851 .valid.min_access_size = 1,
2852 .valid.max_access_size = 8,
2853 .valid.accepts = subpage_accepts,
2854 .endianness = DEVICE_NATIVE_ENDIAN,
2857 static int subpage_register(subpage_t *mmio, uint32_t start, uint32_t end,
2862 if (start >= TARGET_PAGE_SIZE || end >= TARGET_PAGE_SIZE)
2864 idx = SUBPAGE_IDX(start);
2865 eidx = SUBPAGE_IDX(end);
2866 #if defined(DEBUG_SUBPAGE)
2867 printf("%s: %p start %08x end %08x idx %08x eidx %08x section %d\n",
2868 __func__, mmio, start, end, idx, eidx, section);
2870 for (; idx <= eidx; idx++) {
2871 mmio->sub_section[idx] = section;
2877 static subpage_t *subpage_init(FlatView *fv, hwaddr base)
2881 /* mmio->sub_section is set to PHYS_SECTION_UNASSIGNED with g_malloc0 */
2882 mmio = g_malloc0(sizeof(subpage_t) + TARGET_PAGE_SIZE * sizeof(uint16_t));
2885 memory_region_init_io(&mmio->iomem, NULL, &subpage_ops, mmio,
2886 NULL, TARGET_PAGE_SIZE);
2887 mmio->iomem.subpage = true;
2888 #if defined(DEBUG_SUBPAGE)
2889 printf("%s: %p base " TARGET_FMT_plx " len %08x\n", __func__,
2890 mmio, base, TARGET_PAGE_SIZE);
2896 static uint16_t dummy_section(PhysPageMap *map, FlatView *fv, MemoryRegion *mr)
2899 MemoryRegionSection section = {
2902 .offset_within_address_space = 0,
2903 .offset_within_region = 0,
2904 .size = int128_2_64(),
2907 return phys_section_add(map, §ion);
2910 MemoryRegionSection *iotlb_to_section(CPUState *cpu,
2911 hwaddr index, MemTxAttrs attrs)
2913 int asidx = cpu_asidx_from_attrs(cpu, attrs);
2914 CPUAddressSpace *cpuas = &cpu->cpu_ases[asidx];
2915 AddressSpaceDispatch *d = atomic_rcu_read(&cpuas->memory_dispatch);
2916 MemoryRegionSection *sections = d->map.sections;
2918 return §ions[index & ~TARGET_PAGE_MASK];
2921 static void io_mem_init(void)
2923 memory_region_init_io(&io_mem_unassigned, NULL, &unassigned_mem_ops, NULL,
2927 AddressSpaceDispatch *address_space_dispatch_new(FlatView *fv)
2929 AddressSpaceDispatch *d = g_new0(AddressSpaceDispatch, 1);
2932 n = dummy_section(&d->map, fv, &io_mem_unassigned);
2933 assert(n == PHYS_SECTION_UNASSIGNED);
2935 d->phys_map = (PhysPageEntry) { .ptr = PHYS_MAP_NODE_NIL, .skip = 1 };
2940 void address_space_dispatch_free(AddressSpaceDispatch *d)
2942 phys_sections_free(&d->map);
2946 static void do_nothing(CPUState *cpu, run_on_cpu_data d)
2950 static void tcg_log_global_after_sync(MemoryListener *listener)
2952 CPUAddressSpace *cpuas;
2954 /* Wait for the CPU to end the current TB. This avoids the following
2958 * ---------------------- -------------------------
2959 * TLB check -> slow path
2960 * notdirty_mem_write
2964 * TLB check -> fast path
2968 * by pushing the migration thread's memory read after the vCPU thread has
2969 * written the memory.
2971 if (replay_mode == REPLAY_MODE_NONE) {
2973 * VGA can make calls to this function while updating the screen.
2974 * In record/replay mode this causes a deadlock, because
2975 * run_on_cpu waits for rr mutex. Therefore no races are possible
2976 * in this case and no need for making run_on_cpu when
2977 * record/replay is not enabled.
2979 cpuas = container_of(listener, CPUAddressSpace, tcg_as_listener);
2980 run_on_cpu(cpuas->cpu, do_nothing, RUN_ON_CPU_NULL);
2984 static void tcg_commit(MemoryListener *listener)
2986 CPUAddressSpace *cpuas;
2987 AddressSpaceDispatch *d;
2989 assert(tcg_enabled());
2990 /* since each CPU stores ram addresses in its TLB cache, we must
2991 reset the modified entries */
2992 cpuas = container_of(listener, CPUAddressSpace, tcg_as_listener);
2993 cpu_reloading_memory_map();
2994 /* The CPU and TLB are protected by the iothread lock.
2995 * We reload the dispatch pointer now because cpu_reloading_memory_map()
2996 * may have split the RCU critical section.
2998 d = address_space_to_dispatch(cpuas->as);
2999 atomic_rcu_set(&cpuas->memory_dispatch, d);
3000 tlb_flush(cpuas->cpu);
3003 static void memory_map_init(void)
3005 system_memory = g_malloc(sizeof(*system_memory));
3007 memory_region_init(system_memory, NULL, "system", UINT64_MAX);
3008 address_space_init(&address_space_memory, system_memory, "memory");
3010 system_io = g_malloc(sizeof(*system_io));
3011 memory_region_init_io(system_io, NULL, &unassigned_io_ops, NULL, "io",
3013 address_space_init(&address_space_io, system_io, "I/O");
3016 MemoryRegion *get_system_memory(void)
3018 return system_memory;
3021 MemoryRegion *get_system_io(void)
3026 #endif /* !defined(CONFIG_USER_ONLY) */
3028 /* physical memory access (slow version, mainly for debug) */
3029 #if defined(CONFIG_USER_ONLY)
3030 int cpu_memory_rw_debug(CPUState *cpu, target_ulong addr,
3031 void *ptr, target_ulong len, bool is_write)
3034 target_ulong l, page;
3039 page = addr & TARGET_PAGE_MASK;
3040 l = (page + TARGET_PAGE_SIZE) - addr;
3043 flags = page_get_flags(page);
3044 if (!(flags & PAGE_VALID))
3047 if (!(flags & PAGE_WRITE))
3049 /* XXX: this code should not depend on lock_user */
3050 if (!(p = lock_user(VERIFY_WRITE, addr, l, 0)))
3053 unlock_user(p, addr, l);
3055 if (!(flags & PAGE_READ))
3057 /* XXX: this code should not depend on lock_user */
3058 if (!(p = lock_user(VERIFY_READ, addr, l, 1)))
3061 unlock_user(p, addr, 0);
3072 static void invalidate_and_set_dirty(MemoryRegion *mr, hwaddr addr,
3075 uint8_t dirty_log_mask = memory_region_get_dirty_log_mask(mr);
3076 addr += memory_region_get_ram_addr(mr);
3078 /* No early return if dirty_log_mask is or becomes 0, because
3079 * cpu_physical_memory_set_dirty_range will still call
3080 * xen_modified_memory.
3082 if (dirty_log_mask) {
3084 cpu_physical_memory_range_includes_clean(addr, length, dirty_log_mask);
3086 if (dirty_log_mask & (1 << DIRTY_MEMORY_CODE)) {
3087 assert(tcg_enabled());
3088 tb_invalidate_phys_range(addr, addr + length);
3089 dirty_log_mask &= ~(1 << DIRTY_MEMORY_CODE);
3091 cpu_physical_memory_set_dirty_range(addr, length, dirty_log_mask);
3094 void memory_region_flush_rom_device(MemoryRegion *mr, hwaddr addr, hwaddr size)
3097 * In principle this function would work on other memory region types too,
3098 * but the ROM device use case is the only one where this operation is
3099 * necessary. Other memory regions should use the
3100 * address_space_read/write() APIs.
3102 assert(memory_region_is_romd(mr));
3104 invalidate_and_set_dirty(mr, addr, size);
3107 static int memory_access_size(MemoryRegion *mr, unsigned l, hwaddr addr)
3109 unsigned access_size_max = mr->ops->valid.max_access_size;
3111 /* Regions are assumed to support 1-4 byte accesses unless
3112 otherwise specified. */
3113 if (access_size_max == 0) {
3114 access_size_max = 4;
3117 /* Bound the maximum access by the alignment of the address. */
3118 if (!mr->ops->impl.unaligned) {
3119 unsigned align_size_max = addr & -addr;
3120 if (align_size_max != 0 && align_size_max < access_size_max) {
3121 access_size_max = align_size_max;
3125 /* Don't attempt accesses larger than the maximum. */
3126 if (l > access_size_max) {
3127 l = access_size_max;
3134 static bool prepare_mmio_access(MemoryRegion *mr)
3136 bool unlocked = !qemu_mutex_iothread_locked();
3137 bool release_lock = false;
3139 if (unlocked && mr->global_locking) {
3140 qemu_mutex_lock_iothread();
3142 release_lock = true;
3144 if (mr->flush_coalesced_mmio) {
3146 qemu_mutex_lock_iothread();
3148 qemu_flush_coalesced_mmio_buffer();
3150 qemu_mutex_unlock_iothread();
3154 return release_lock;
3157 /* Called within RCU critical section. */
3158 static MemTxResult flatview_write_continue(FlatView *fv, hwaddr addr,
3161 hwaddr len, hwaddr addr1,
3162 hwaddr l, MemoryRegion *mr)
3166 MemTxResult result = MEMTX_OK;
3167 bool release_lock = false;
3168 const uint8_t *buf = ptr;
3171 if (!memory_access_is_direct(mr, true)) {
3172 release_lock |= prepare_mmio_access(mr);
3173 l = memory_access_size(mr, l, addr1);
3174 /* XXX: could force current_cpu to NULL to avoid
3176 val = ldn_he_p(buf, l);
3177 result |= memory_region_dispatch_write(mr, addr1, val,
3178 size_memop(l), attrs);
3181 ram_ptr = qemu_ram_ptr_length(mr->ram_block, addr1, &l, false);
3182 memcpy(ram_ptr, buf, l);
3183 invalidate_and_set_dirty(mr, addr1, l);
3187 qemu_mutex_unlock_iothread();
3188 release_lock = false;
3200 mr = flatview_translate(fv, addr, &addr1, &l, true, attrs);
3206 /* Called from RCU critical section. */
3207 static MemTxResult flatview_write(FlatView *fv, hwaddr addr, MemTxAttrs attrs,
3208 const void *buf, hwaddr len)
3213 MemTxResult result = MEMTX_OK;
3216 mr = flatview_translate(fv, addr, &addr1, &l, true, attrs);
3217 result = flatview_write_continue(fv, addr, attrs, buf, len,
3223 /* Called within RCU critical section. */
3224 MemTxResult flatview_read_continue(FlatView *fv, hwaddr addr,
3225 MemTxAttrs attrs, void *ptr,
3226 hwaddr len, hwaddr addr1, hwaddr l,
3231 MemTxResult result = MEMTX_OK;
3232 bool release_lock = false;
3236 if (!memory_access_is_direct(mr, false)) {
3238 release_lock |= prepare_mmio_access(mr);
3239 l = memory_access_size(mr, l, addr1);
3240 result |= memory_region_dispatch_read(mr, addr1, &val,
3241 size_memop(l), attrs);
3242 stn_he_p(buf, l, val);
3245 ram_ptr = qemu_ram_ptr_length(mr->ram_block, addr1, &l, false);
3246 memcpy(buf, ram_ptr, l);
3250 qemu_mutex_unlock_iothread();
3251 release_lock = false;
3263 mr = flatview_translate(fv, addr, &addr1, &l, false, attrs);
3269 /* Called from RCU critical section. */
3270 static MemTxResult flatview_read(FlatView *fv, hwaddr addr,
3271 MemTxAttrs attrs, void *buf, hwaddr len)
3278 mr = flatview_translate(fv, addr, &addr1, &l, false, attrs);
3279 return flatview_read_continue(fv, addr, attrs, buf, len,
3283 MemTxResult address_space_read_full(AddressSpace *as, hwaddr addr,
3284 MemTxAttrs attrs, void *buf, hwaddr len)
3286 MemTxResult result = MEMTX_OK;
3290 RCU_READ_LOCK_GUARD();
3291 fv = address_space_to_flatview(as);
3292 result = flatview_read(fv, addr, attrs, buf, len);
3298 MemTxResult address_space_write(AddressSpace *as, hwaddr addr,
3300 const void *buf, hwaddr len)
3302 MemTxResult result = MEMTX_OK;
3306 RCU_READ_LOCK_GUARD();
3307 fv = address_space_to_flatview(as);
3308 result = flatview_write(fv, addr, attrs, buf, len);
3314 MemTxResult address_space_rw(AddressSpace *as, hwaddr addr, MemTxAttrs attrs,
3315 void *buf, hwaddr len, bool is_write)
3318 return address_space_write(as, addr, attrs, buf, len);
3320 return address_space_read_full(as, addr, attrs, buf, len);
3324 void cpu_physical_memory_rw(hwaddr addr, void *buf,
3325 hwaddr len, bool is_write)
3327 address_space_rw(&address_space_memory, addr, MEMTXATTRS_UNSPECIFIED,
3328 buf, len, is_write);
3331 enum write_rom_type {
3336 static inline MemTxResult address_space_write_rom_internal(AddressSpace *as,
3341 enum write_rom_type type)
3347 const uint8_t *buf = ptr;
3349 RCU_READ_LOCK_GUARD();
3352 mr = address_space_translate(as, addr, &addr1, &l, true, attrs);
3354 if (!(memory_region_is_ram(mr) ||
3355 memory_region_is_romd(mr))) {
3356 l = memory_access_size(mr, l, addr1);
3359 ram_ptr = qemu_map_ram_ptr(mr->ram_block, addr1);
3362 memcpy(ram_ptr, buf, l);
3363 invalidate_and_set_dirty(mr, addr1, l);
3366 flush_icache_range((uintptr_t)ram_ptr, (uintptr_t)ram_ptr + l);
3377 /* used for ROM loading : can write in RAM and ROM */
3378 MemTxResult address_space_write_rom(AddressSpace *as, hwaddr addr,
3380 const void *buf, hwaddr len)
3382 return address_space_write_rom_internal(as, addr, attrs,
3383 buf, len, WRITE_DATA);
3386 void cpu_flush_icache_range(hwaddr start, hwaddr len)
3389 * This function should do the same thing as an icache flush that was
3390 * triggered from within the guest. For TCG we are always cache coherent,
3391 * so there is no need to flush anything. For KVM / Xen we need to flush
3392 * the host's instruction cache at least.
3394 if (tcg_enabled()) {
3398 address_space_write_rom_internal(&address_space_memory,
3399 start, MEMTXATTRS_UNSPECIFIED,
3400 NULL, len, FLUSH_CACHE);
3411 static BounceBuffer bounce;
3413 typedef struct MapClient {
3415 QLIST_ENTRY(MapClient) link;
3418 QemuMutex map_client_list_lock;
3419 static QLIST_HEAD(, MapClient) map_client_list
3420 = QLIST_HEAD_INITIALIZER(map_client_list);
3422 static void cpu_unregister_map_client_do(MapClient *client)
3424 QLIST_REMOVE(client, link);
3428 static void cpu_notify_map_clients_locked(void)
3432 while (!QLIST_EMPTY(&map_client_list)) {
3433 client = QLIST_FIRST(&map_client_list);
3434 qemu_bh_schedule(client->bh);
3435 cpu_unregister_map_client_do(client);
3439 void cpu_register_map_client(QEMUBH *bh)
3441 MapClient *client = g_malloc(sizeof(*client));
3443 qemu_mutex_lock(&map_client_list_lock);
3445 QLIST_INSERT_HEAD(&map_client_list, client, link);
3446 if (!atomic_read(&bounce.in_use)) {
3447 cpu_notify_map_clients_locked();
3449 qemu_mutex_unlock(&map_client_list_lock);
3452 void cpu_exec_init_all(void)
3454 qemu_mutex_init(&ram_list.mutex);
3455 /* The data structures we set up here depend on knowing the page size,
3456 * so no more changes can be made after this point.
3457 * In an ideal world, nothing we did before we had finished the
3458 * machine setup would care about the target page size, and we could
3459 * do this much later, rather than requiring board models to state
3460 * up front what their requirements are.
3462 finalize_target_page_bits();
3465 qemu_mutex_init(&map_client_list_lock);
3468 void cpu_unregister_map_client(QEMUBH *bh)
3472 qemu_mutex_lock(&map_client_list_lock);
3473 QLIST_FOREACH(client, &map_client_list, link) {
3474 if (client->bh == bh) {
3475 cpu_unregister_map_client_do(client);
3479 qemu_mutex_unlock(&map_client_list_lock);
3482 static void cpu_notify_map_clients(void)
3484 qemu_mutex_lock(&map_client_list_lock);
3485 cpu_notify_map_clients_locked();
3486 qemu_mutex_unlock(&map_client_list_lock);
3489 static bool flatview_access_valid(FlatView *fv, hwaddr addr, hwaddr len,
3490 bool is_write, MemTxAttrs attrs)
3497 mr = flatview_translate(fv, addr, &xlat, &l, is_write, attrs);
3498 if (!memory_access_is_direct(mr, is_write)) {
3499 l = memory_access_size(mr, l, addr);
3500 if (!memory_region_access_valid(mr, xlat, l, is_write, attrs)) {
3511 bool address_space_access_valid(AddressSpace *as, hwaddr addr,
3512 hwaddr len, bool is_write,
3518 RCU_READ_LOCK_GUARD();
3519 fv = address_space_to_flatview(as);
3520 result = flatview_access_valid(fv, addr, len, is_write, attrs);
3525 flatview_extend_translation(FlatView *fv, hwaddr addr,
3527 MemoryRegion *mr, hwaddr base, hwaddr len,
3528 bool is_write, MemTxAttrs attrs)
3532 MemoryRegion *this_mr;
3538 if (target_len == 0) {
3543 this_mr = flatview_translate(fv, addr, &xlat,
3544 &len, is_write, attrs);
3545 if (this_mr != mr || xlat != base + done) {
3551 /* Map a physical memory region into a host virtual address.
3552 * May map a subset of the requested range, given by and returned in *plen.
3553 * May return NULL if resources needed to perform the mapping are exhausted.
3554 * Use only for reads OR writes - not for read-modify-write operations.
3555 * Use cpu_register_map_client() to know when retrying the map operation is
3556 * likely to succeed.
3558 void *address_space_map(AddressSpace *as,
3575 RCU_READ_LOCK_GUARD();
3576 fv = address_space_to_flatview(as);
3577 mr = flatview_translate(fv, addr, &xlat, &l, is_write, attrs);
3579 if (!memory_access_is_direct(mr, is_write)) {
3580 if (atomic_xchg(&bounce.in_use, true)) {
3584 /* Avoid unbounded allocations */
3585 l = MIN(l, TARGET_PAGE_SIZE);
3586 bounce.buffer = qemu_memalign(TARGET_PAGE_SIZE, l);
3590 memory_region_ref(mr);
3593 flatview_read(fv, addr, MEMTXATTRS_UNSPECIFIED,
3598 return bounce.buffer;
3602 memory_region_ref(mr);
3603 *plen = flatview_extend_translation(fv, addr, len, mr, xlat,
3604 l, is_write, attrs);
3605 ptr = qemu_ram_ptr_length(mr->ram_block, xlat, plen, true);
3610 /* Unmaps a memory region previously mapped by address_space_map().
3611 * Will also mark the memory as dirty if is_write is true. access_len gives
3612 * the amount of memory that was actually read or written by the caller.
3614 void address_space_unmap(AddressSpace *as, void *buffer, hwaddr len,
3615 bool is_write, hwaddr access_len)
3617 if (buffer != bounce.buffer) {
3621 mr = memory_region_from_host(buffer, &addr1);
3624 invalidate_and_set_dirty(mr, addr1, access_len);
3626 if (xen_enabled()) {
3627 xen_invalidate_map_cache_entry(buffer);
3629 memory_region_unref(mr);
3633 address_space_write(as, bounce.addr, MEMTXATTRS_UNSPECIFIED,
3634 bounce.buffer, access_len);
3636 qemu_vfree(bounce.buffer);
3637 bounce.buffer = NULL;
3638 memory_region_unref(bounce.mr);
3639 atomic_mb_set(&bounce.in_use, false);
3640 cpu_notify_map_clients();
3643 void *cpu_physical_memory_map(hwaddr addr,
3647 return address_space_map(&address_space_memory, addr, plen, is_write,
3648 MEMTXATTRS_UNSPECIFIED);
3651 void cpu_physical_memory_unmap(void *buffer, hwaddr len,
3652 bool is_write, hwaddr access_len)
3654 return address_space_unmap(&address_space_memory, buffer, len, is_write, access_len);
3657 #define ARG1_DECL AddressSpace *as
3660 #define TRANSLATE(...) address_space_translate(as, __VA_ARGS__)
3661 #define RCU_READ_LOCK(...) rcu_read_lock()
3662 #define RCU_READ_UNLOCK(...) rcu_read_unlock()
3663 #include "memory_ldst.c.inc"
3665 int64_t address_space_cache_init(MemoryRegionCache *cache,
3671 AddressSpaceDispatch *d;
3678 cache->fv = address_space_get_flatview(as);
3679 d = flatview_to_dispatch(cache->fv);
3680 cache->mrs = *address_space_translate_internal(d, addr, &cache->xlat, &l, true);
3683 memory_region_ref(mr);
3684 if (memory_access_is_direct(mr, is_write)) {
3685 /* We don't care about the memory attributes here as we're only
3686 * doing this if we found actual RAM, which behaves the same
3687 * regardless of attributes; so UNSPECIFIED is fine.
3689 l = flatview_extend_translation(cache->fv, addr, len, mr,
3690 cache->xlat, l, is_write,
3691 MEMTXATTRS_UNSPECIFIED);
3692 cache->ptr = qemu_ram_ptr_length(mr->ram_block, cache->xlat, &l, true);
3698 cache->is_write = is_write;
3702 void address_space_cache_invalidate(MemoryRegionCache *cache,
3706 assert(cache->is_write);
3707 if (likely(cache->ptr)) {
3708 invalidate_and_set_dirty(cache->mrs.mr, addr + cache->xlat, access_len);
3712 void address_space_cache_destroy(MemoryRegionCache *cache)
3714 if (!cache->mrs.mr) {
3718 if (xen_enabled()) {
3719 xen_invalidate_map_cache_entry(cache->ptr);
3721 memory_region_unref(cache->mrs.mr);
3722 flatview_unref(cache->fv);
3723 cache->mrs.mr = NULL;
3727 /* Called from RCU critical section. This function has the same
3728 * semantics as address_space_translate, but it only works on a
3729 * predefined range of a MemoryRegion that was mapped with
3730 * address_space_cache_init.
3732 static inline MemoryRegion *address_space_translate_cached(
3733 MemoryRegionCache *cache, hwaddr addr, hwaddr *xlat,
3734 hwaddr *plen, bool is_write, MemTxAttrs attrs)
3736 MemoryRegionSection section;
3738 IOMMUMemoryRegion *iommu_mr;
3739 AddressSpace *target_as;
3741 assert(!cache->ptr);
3742 *xlat = addr + cache->xlat;
3745 iommu_mr = memory_region_get_iommu(mr);
3751 section = address_space_translate_iommu(iommu_mr, xlat, plen,
3752 NULL, is_write, true,
3757 /* Called from RCU critical section. address_space_read_cached uses this
3758 * out of line function when the target is an MMIO or IOMMU region.
3761 address_space_read_cached_slow(MemoryRegionCache *cache, hwaddr addr,
3762 void *buf, hwaddr len)
3768 mr = address_space_translate_cached(cache, addr, &addr1, &l, false,
3769 MEMTXATTRS_UNSPECIFIED);
3770 return flatview_read_continue(cache->fv,
3771 addr, MEMTXATTRS_UNSPECIFIED, buf, len,
3775 /* Called from RCU critical section. address_space_write_cached uses this
3776 * out of line function when the target is an MMIO or IOMMU region.
3779 address_space_write_cached_slow(MemoryRegionCache *cache, hwaddr addr,
3780 const void *buf, hwaddr len)
3786 mr = address_space_translate_cached(cache, addr, &addr1, &l, true,
3787 MEMTXATTRS_UNSPECIFIED);
3788 return flatview_write_continue(cache->fv,
3789 addr, MEMTXATTRS_UNSPECIFIED, buf, len,
3793 #define ARG1_DECL MemoryRegionCache *cache
3795 #define SUFFIX _cached_slow
3796 #define TRANSLATE(...) address_space_translate_cached(cache, __VA_ARGS__)
3797 #define RCU_READ_LOCK() ((void)0)
3798 #define RCU_READ_UNLOCK() ((void)0)
3799 #include "memory_ldst.c.inc"
3801 /* virtual memory access for debug (includes writing to ROM) */
3802 int cpu_memory_rw_debug(CPUState *cpu, target_ulong addr,
3803 void *ptr, target_ulong len, bool is_write)
3806 target_ulong l, page;
3809 cpu_synchronize_state(cpu);
3815 page = addr & TARGET_PAGE_MASK;
3816 phys_addr = cpu_get_phys_page_attrs_debug(cpu, page, &attrs);
3817 asidx = cpu_asidx_from_attrs(cpu, attrs);
3818 /* if no physical page mapped, return an error */
3819 if (phys_addr == -1)
3821 l = (page + TARGET_PAGE_SIZE) - addr;
3824 phys_addr += (addr & ~TARGET_PAGE_MASK);
3826 res = address_space_write_rom(cpu->cpu_ases[asidx].as, phys_addr,
3829 res = address_space_read(cpu->cpu_ases[asidx].as, phys_addr,
3832 if (res != MEMTX_OK) {
3843 * Allows code that needs to deal with migration bitmaps etc to still be built
3844 * target independent.
3846 size_t qemu_target_page_size(void)
3848 return TARGET_PAGE_SIZE;
3851 int qemu_target_page_bits(void)
3853 return TARGET_PAGE_BITS;
3856 int qemu_target_page_bits_min(void)
3858 return TARGET_PAGE_BITS_MIN;
3862 bool target_words_bigendian(void)
3864 #if defined(TARGET_WORDS_BIGENDIAN)
3871 #ifndef CONFIG_USER_ONLY
3872 bool cpu_physical_memory_is_io(hwaddr phys_addr)
3878 RCU_READ_LOCK_GUARD();
3879 mr = address_space_translate(&address_space_memory,
3880 phys_addr, &phys_addr, &l, false,
3881 MEMTXATTRS_UNSPECIFIED);
3883 res = !(memory_region_is_ram(mr) || memory_region_is_romd(mr));
3887 int qemu_ram_foreach_block(RAMBlockIterFunc func, void *opaque)
3892 RCU_READ_LOCK_GUARD();
3893 RAMBLOCK_FOREACH(block) {
3894 ret = func(block, opaque);
3903 * Unmap pages of memory from start to start+length such that
3904 * they a) read as 0, b) Trigger whatever fault mechanism
3905 * the OS provides for postcopy.
3906 * The pages must be unmapped by the end of the function.
3907 * Returns: 0 on success, none-0 on failure
3910 int ram_block_discard_range(RAMBlock *rb, uint64_t start, size_t length)
3914 uint8_t *host_startaddr = rb->host + start;
3916 if (!QEMU_PTR_IS_ALIGNED(host_startaddr, rb->page_size)) {
3917 error_report("ram_block_discard_range: Unaligned start address: %p",
3922 if ((start + length) <= rb->used_length) {
3923 bool need_madvise, need_fallocate;
3924 if (!QEMU_IS_ALIGNED(length, rb->page_size)) {
3925 error_report("ram_block_discard_range: Unaligned length: %zx",
3930 errno = ENOTSUP; /* If we are missing MADVISE etc */
3932 /* The logic here is messy;
3933 * madvise DONTNEED fails for hugepages
3934 * fallocate works on hugepages and shmem
3936 need_madvise = (rb->page_size == qemu_host_page_size);
3937 need_fallocate = rb->fd != -1;
3938 if (need_fallocate) {
3939 /* For a file, this causes the area of the file to be zero'd
3940 * if read, and for hugetlbfs also causes it to be unmapped
3941 * so a userfault will trigger.
3943 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
3944 ret = fallocate(rb->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
3948 error_report("ram_block_discard_range: Failed to fallocate "
3949 "%s:%" PRIx64 " +%zx (%d)",
3950 rb->idstr, start, length, ret);
3955 error_report("ram_block_discard_range: fallocate not available/file"
3956 "%s:%" PRIx64 " +%zx (%d)",
3957 rb->idstr, start, length, ret);
3962 /* For normal RAM this causes it to be unmapped,
3963 * for shared memory it causes the local mapping to disappear
3964 * and to fall back on the file contents (which we just
3965 * fallocate'd away).
3967 #if defined(CONFIG_MADVISE)
3968 ret = madvise(host_startaddr, length, MADV_DONTNEED);
3971 error_report("ram_block_discard_range: Failed to discard range "
3972 "%s:%" PRIx64 " +%zx (%d)",
3973 rb->idstr, start, length, ret);
3978 error_report("ram_block_discard_range: MADVISE not available"
3979 "%s:%" PRIx64 " +%zx (%d)",
3980 rb->idstr, start, length, ret);
3984 trace_ram_block_discard_range(rb->idstr, host_startaddr, length,
3985 need_madvise, need_fallocate, ret);
3987 error_report("ram_block_discard_range: Overrun block '%s' (%" PRIu64
3988 "/%zx/" RAM_ADDR_FMT")",
3989 rb->idstr, start, length, rb->used_length);
3996 bool ramblock_is_pmem(RAMBlock *rb)
3998 return rb->flags & RAM_PMEM;
4003 void page_size_init(void)
4005 /* NOTE: we can always suppose that qemu_host_page_size >=
4007 if (qemu_host_page_size == 0) {
4008 qemu_host_page_size = qemu_real_host_page_size;
4010 if (qemu_host_page_size < TARGET_PAGE_SIZE) {
4011 qemu_host_page_size = TARGET_PAGE_SIZE;
4013 qemu_host_page_mask = -(intptr_t)qemu_host_page_size;
4016 #if !defined(CONFIG_USER_ONLY)
4018 static void mtree_print_phys_entries(int start, int end, int skip, int ptr)
4020 if (start == end - 1) {
4021 qemu_printf("\t%3d ", start);
4023 qemu_printf("\t%3d..%-3d ", start, end - 1);
4025 qemu_printf(" skip=%d ", skip);
4026 if (ptr == PHYS_MAP_NODE_NIL) {
4027 qemu_printf(" ptr=NIL");
4029 qemu_printf(" ptr=#%d", ptr);
4031 qemu_printf(" ptr=[%d]", ptr);
4036 #define MR_SIZE(size) (int128_nz(size) ? (hwaddr)int128_get64( \
4037 int128_sub((size), int128_one())) : 0)
4039 void mtree_print_dispatch(AddressSpaceDispatch *d, MemoryRegion *root)
4043 qemu_printf(" Dispatch\n");
4044 qemu_printf(" Physical sections\n");
4046 for (i = 0; i < d->map.sections_nb; ++i) {
4047 MemoryRegionSection *s = d->map.sections + i;
4048 const char *names[] = { " [unassigned]", " [not dirty]",
4049 " [ROM]", " [watch]" };
4051 qemu_printf(" #%d @" TARGET_FMT_plx ".." TARGET_FMT_plx
4054 s->offset_within_address_space,
4055 s->offset_within_address_space + MR_SIZE(s->mr->size),
4056 s->mr->name ? s->mr->name : "(noname)",
4057 i < ARRAY_SIZE(names) ? names[i] : "",
4058 s->mr == root ? " [ROOT]" : "",
4059 s == d->mru_section ? " [MRU]" : "",
4060 s->mr->is_iommu ? " [iommu]" : "");
4063 qemu_printf(" alias=%s", s->mr->alias->name ?
4064 s->mr->alias->name : "noname");
4069 qemu_printf(" Nodes (%d bits per level, %d levels) ptr=[%d] skip=%d\n",
4070 P_L2_BITS, P_L2_LEVELS, d->phys_map.ptr, d->phys_map.skip);
4071 for (i = 0; i < d->map.nodes_nb; ++i) {
4074 Node *n = d->map.nodes + i;
4076 qemu_printf(" [%d]\n", i);
4078 for (j = 0, jprev = 0, prev = *n[0]; j < ARRAY_SIZE(*n); ++j) {
4079 PhysPageEntry *pe = *n + j;
4081 if (pe->ptr == prev.ptr && pe->skip == prev.skip) {
4085 mtree_print_phys_entries(jprev, j, prev.skip, prev.ptr);
4091 if (jprev != ARRAY_SIZE(*n)) {
4092 mtree_print_phys_entries(jprev, j, prev.skip, prev.ptr);
4098 * If positive, discarding RAM is disabled. If negative, discarding RAM is
4099 * required to work and cannot be disabled.
4101 static int ram_block_discard_disabled;
4103 int ram_block_discard_disable(bool state)
4108 atomic_dec(&ram_block_discard_disabled);
4113 old = atomic_read(&ram_block_discard_disabled);
4117 } while (atomic_cmpxchg(&ram_block_discard_disabled, old, old + 1) != old);
4121 int ram_block_discard_require(bool state)
4126 atomic_inc(&ram_block_discard_disabled);
4131 old = atomic_read(&ram_block_discard_disabled);
4135 } while (atomic_cmpxchg(&ram_block_discard_disabled, old, old - 1) != old);
4139 bool ram_block_discard_is_disabled(void)
4141 return atomic_read(&ram_block_discard_disabled) > 0;
4144 bool ram_block_discard_is_required(void)
4146 return atomic_read(&ram_block_discard_disabled) < 0;