2 * RAM allocation and memory access
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 2.1 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 "exec/page-vary.h"
22 #include "qapi/error.h"
24 #include "qemu/cutils.h"
25 #include "qemu/cacheflush.h"
26 #include "qemu/hbitmap.h"
27 #include "qemu/madvise.h"
30 #include "hw/core/tcg-cpu-ops.h"
31 #endif /* CONFIG_TCG */
33 #include "exec/exec-all.h"
34 #include "exec/target_page.h"
35 #include "hw/qdev-core.h"
36 #include "hw/qdev-properties.h"
37 #include "hw/boards.h"
38 #include "hw/xen/xen.h"
39 #include "sysemu/kvm.h"
40 #include "sysemu/tcg.h"
41 #include "sysemu/qtest.h"
42 #include "qemu/timer.h"
43 #include "qemu/config-file.h"
44 #include "qemu/error-report.h"
45 #include "qemu/qemu-print.h"
47 #include "qemu/memalign.h"
48 #include "exec/memory.h"
49 #include "exec/ioport.h"
50 #include "sysemu/dma.h"
51 #include "sysemu/hostmem.h"
52 #include "sysemu/hw_accel.h"
53 #include "sysemu/xen-mapcache.h"
54 #include "trace/trace-root.h"
56 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
57 #include <linux/falloc.h>
60 #include "qemu/rcu_queue.h"
61 #include "qemu/main-loop.h"
62 #include "exec/translate-all.h"
63 #include "sysemu/replay.h"
65 #include "exec/memory-internal.h"
66 #include "exec/ram_addr.h"
68 #include "qemu/pmem.h"
70 #include "migration/vmstate.h"
72 #include "qemu/range.h"
74 #include "qemu/mmap-alloc.h"
77 #include "monitor/monitor.h"
79 #ifdef CONFIG_LIBDAXCTL
80 #include <daxctl/libdaxctl.h>
83 //#define DEBUG_SUBPAGE
85 /* ram_list is read under rcu_read_lock()/rcu_read_unlock(). Writes
86 * are protected by the ramlist lock.
88 RAMList ram_list = { .blocks = QLIST_HEAD_INITIALIZER(ram_list.blocks) };
90 static MemoryRegion *system_memory;
91 static MemoryRegion *system_io;
93 AddressSpace address_space_io;
94 AddressSpace address_space_memory;
96 static MemoryRegion io_mem_unassigned;
98 typedef struct PhysPageEntry PhysPageEntry;
100 struct PhysPageEntry {
101 /* How many bits skip to next level (in units of L2_SIZE). 0 for a leaf. */
103 /* index into phys_sections (!skip) or phys_map_nodes (skip) */
107 #define PHYS_MAP_NODE_NIL (((uint32_t)~0) >> 6)
109 /* Size of the L2 (and L3, etc) page tables. */
110 #define ADDR_SPACE_BITS 64
113 #define P_L2_SIZE (1 << P_L2_BITS)
115 #define P_L2_LEVELS (((ADDR_SPACE_BITS - TARGET_PAGE_BITS - 1) / P_L2_BITS) + 1)
117 typedef PhysPageEntry Node[P_L2_SIZE];
119 typedef struct PhysPageMap {
122 unsigned sections_nb;
123 unsigned sections_nb_alloc;
125 unsigned nodes_nb_alloc;
127 MemoryRegionSection *sections;
130 struct AddressSpaceDispatch {
131 MemoryRegionSection *mru_section;
132 /* This is a multi-level map on the physical address space.
133 * The bottom level has pointers to MemoryRegionSections.
135 PhysPageEntry phys_map;
139 #define SUBPAGE_IDX(addr) ((addr) & ~TARGET_PAGE_MASK)
140 typedef struct subpage_t {
144 uint16_t sub_section[];
147 #define PHYS_SECTION_UNASSIGNED 0
149 static void io_mem_init(void);
150 static void memory_map_init(void);
151 static void tcg_log_global_after_sync(MemoryListener *listener);
152 static void tcg_commit(MemoryListener *listener);
155 * CPUAddressSpace: all the information a CPU needs about an AddressSpace
156 * @cpu: the CPU whose AddressSpace this is
157 * @as: the AddressSpace itself
158 * @memory_dispatch: its dispatch pointer (cached, RCU protected)
159 * @tcg_as_listener: listener for tracking changes to the AddressSpace
161 struct CPUAddressSpace {
164 struct AddressSpaceDispatch *memory_dispatch;
165 MemoryListener tcg_as_listener;
168 struct DirtyBitmapSnapshot {
171 unsigned long dirty[];
174 static void phys_map_node_reserve(PhysPageMap *map, unsigned nodes)
176 static unsigned alloc_hint = 16;
177 if (map->nodes_nb + nodes > map->nodes_nb_alloc) {
178 map->nodes_nb_alloc = MAX(alloc_hint, map->nodes_nb + nodes);
179 map->nodes = g_renew(Node, map->nodes, map->nodes_nb_alloc);
180 alloc_hint = map->nodes_nb_alloc;
184 static uint32_t phys_map_node_alloc(PhysPageMap *map, bool leaf)
191 ret = map->nodes_nb++;
193 assert(ret != PHYS_MAP_NODE_NIL);
194 assert(ret != map->nodes_nb_alloc);
196 e.skip = leaf ? 0 : 1;
197 e.ptr = leaf ? PHYS_SECTION_UNASSIGNED : PHYS_MAP_NODE_NIL;
198 for (i = 0; i < P_L2_SIZE; ++i) {
199 memcpy(&p[i], &e, sizeof(e));
204 static void phys_page_set_level(PhysPageMap *map, PhysPageEntry *lp,
205 hwaddr *index, uint64_t *nb, uint16_t leaf,
209 hwaddr step = (hwaddr)1 << (level * P_L2_BITS);
211 if (lp->skip && lp->ptr == PHYS_MAP_NODE_NIL) {
212 lp->ptr = phys_map_node_alloc(map, level == 0);
214 p = map->nodes[lp->ptr];
215 lp = &p[(*index >> (level * P_L2_BITS)) & (P_L2_SIZE - 1)];
217 while (*nb && lp < &p[P_L2_SIZE]) {
218 if ((*index & (step - 1)) == 0 && *nb >= step) {
224 phys_page_set_level(map, lp, index, nb, leaf, level - 1);
230 static void phys_page_set(AddressSpaceDispatch *d,
231 hwaddr index, uint64_t nb,
234 /* Wildly overreserve - it doesn't matter much. */
235 phys_map_node_reserve(&d->map, 3 * P_L2_LEVELS);
237 phys_page_set_level(&d->map, &d->phys_map, &index, &nb, leaf, P_L2_LEVELS - 1);
240 /* Compact a non leaf page entry. Simply detect that the entry has a single child,
241 * and update our entry so we can skip it and go directly to the destination.
243 static void phys_page_compact(PhysPageEntry *lp, Node *nodes)
245 unsigned valid_ptr = P_L2_SIZE;
250 if (lp->ptr == PHYS_MAP_NODE_NIL) {
255 for (i = 0; i < P_L2_SIZE; i++) {
256 if (p[i].ptr == PHYS_MAP_NODE_NIL) {
263 phys_page_compact(&p[i], nodes);
267 /* We can only compress if there's only one child. */
272 assert(valid_ptr < P_L2_SIZE);
274 /* Don't compress if it won't fit in the # of bits we have. */
275 if (P_L2_LEVELS >= (1 << 6) &&
276 lp->skip + p[valid_ptr].skip >= (1 << 6)) {
280 lp->ptr = p[valid_ptr].ptr;
281 if (!p[valid_ptr].skip) {
282 /* If our only child is a leaf, make this a leaf. */
283 /* By design, we should have made this node a leaf to begin with so we
284 * should never reach here.
285 * But since it's so simple to handle this, let's do it just in case we
290 lp->skip += p[valid_ptr].skip;
294 void address_space_dispatch_compact(AddressSpaceDispatch *d)
296 if (d->phys_map.skip) {
297 phys_page_compact(&d->phys_map, d->map.nodes);
301 static inline bool section_covers_addr(const MemoryRegionSection *section,
304 /* Memory topology clips a memory region to [0, 2^64); size.hi > 0 means
305 * the section must cover the entire address space.
307 return int128_gethi(section->size) ||
308 range_covers_byte(section->offset_within_address_space,
309 int128_getlo(section->size), addr);
312 static MemoryRegionSection *phys_page_find(AddressSpaceDispatch *d, hwaddr addr)
314 PhysPageEntry lp = d->phys_map, *p;
315 Node *nodes = d->map.nodes;
316 MemoryRegionSection *sections = d->map.sections;
317 hwaddr index = addr >> TARGET_PAGE_BITS;
320 for (i = P_L2_LEVELS; lp.skip && (i -= lp.skip) >= 0;) {
321 if (lp.ptr == PHYS_MAP_NODE_NIL) {
322 return §ions[PHYS_SECTION_UNASSIGNED];
325 lp = p[(index >> (i * P_L2_BITS)) & (P_L2_SIZE - 1)];
328 if (section_covers_addr(§ions[lp.ptr], addr)) {
329 return §ions[lp.ptr];
331 return §ions[PHYS_SECTION_UNASSIGNED];
335 /* Called from RCU critical section */
336 static MemoryRegionSection *address_space_lookup_region(AddressSpaceDispatch *d,
338 bool resolve_subpage)
340 MemoryRegionSection *section = qatomic_read(&d->mru_section);
343 if (!section || section == &d->map.sections[PHYS_SECTION_UNASSIGNED] ||
344 !section_covers_addr(section, addr)) {
345 section = phys_page_find(d, addr);
346 qatomic_set(&d->mru_section, section);
348 if (resolve_subpage && section->mr->subpage) {
349 subpage = container_of(section->mr, subpage_t, iomem);
350 section = &d->map.sections[subpage->sub_section[SUBPAGE_IDX(addr)]];
355 /* Called from RCU critical section */
356 static MemoryRegionSection *
357 address_space_translate_internal(AddressSpaceDispatch *d, hwaddr addr, hwaddr *xlat,
358 hwaddr *plen, bool resolve_subpage)
360 MemoryRegionSection *section;
364 section = address_space_lookup_region(d, addr, resolve_subpage);
365 /* Compute offset within MemoryRegionSection */
366 addr -= section->offset_within_address_space;
368 /* Compute offset within MemoryRegion */
369 *xlat = addr + section->offset_within_region;
373 /* MMIO registers can be expected to perform full-width accesses based only
374 * on their address, without considering adjacent registers that could
375 * decode to completely different MemoryRegions. When such registers
376 * exist (e.g. I/O ports 0xcf8 and 0xcf9 on most PC chipsets), MMIO
377 * regions overlap wildly. For this reason we cannot clamp the accesses
380 * If the length is small (as is the case for address_space_ldl/stl),
381 * everything works fine. If the incoming length is large, however,
382 * the caller really has to do the clamping through memory_access_size.
384 if (memory_region_is_ram(mr)) {
385 diff = int128_sub(section->size, int128_make64(addr));
386 *plen = int128_get64(int128_min(diff, int128_make64(*plen)));
392 * address_space_translate_iommu - translate an address through an IOMMU
393 * memory region and then through the target address space.
395 * @iommu_mr: the IOMMU memory region that we start the translation from
396 * @addr: the address to be translated through the MMU
397 * @xlat: the translated address offset within the destination memory region.
398 * It cannot be %NULL.
399 * @plen_out: valid read/write length of the translated address. It
401 * @page_mask_out: page mask for the translated address. This
402 * should only be meaningful for IOMMU translated
403 * addresses, since there may be huge pages that this bit
404 * would tell. It can be %NULL if we don't care about it.
405 * @is_write: whether the translation operation is for write
406 * @is_mmio: whether this can be MMIO, set true if it can
407 * @target_as: the address space targeted by the IOMMU
408 * @attrs: transaction attributes
410 * This function is called from RCU critical section. It is the common
411 * part of flatview_do_translate and address_space_translate_cached.
413 static MemoryRegionSection address_space_translate_iommu(IOMMUMemoryRegion *iommu_mr,
416 hwaddr *page_mask_out,
419 AddressSpace **target_as,
422 MemoryRegionSection *section;
423 hwaddr page_mask = (hwaddr)-1;
427 IOMMUMemoryRegionClass *imrc = memory_region_get_iommu_class_nocheck(iommu_mr);
431 if (imrc->attrs_to_index) {
432 iommu_idx = imrc->attrs_to_index(iommu_mr, attrs);
435 iotlb = imrc->translate(iommu_mr, addr, is_write ?
436 IOMMU_WO : IOMMU_RO, iommu_idx);
438 if (!(iotlb.perm & (1 << is_write))) {
442 addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
443 | (addr & iotlb.addr_mask));
444 page_mask &= iotlb.addr_mask;
445 *plen_out = MIN(*plen_out, (addr | iotlb.addr_mask) - addr + 1);
446 *target_as = iotlb.target_as;
448 section = address_space_translate_internal(
449 address_space_to_dispatch(iotlb.target_as), addr, xlat,
452 iommu_mr = memory_region_get_iommu(section->mr);
453 } while (unlikely(iommu_mr));
456 *page_mask_out = page_mask;
461 return (MemoryRegionSection) { .mr = &io_mem_unassigned };
465 * flatview_do_translate - translate an address in FlatView
467 * @fv: the flat view that we want to translate on
468 * @addr: the address to be translated in above address space
469 * @xlat: the translated address offset within memory region. It
471 * @plen_out: valid read/write length of the translated address. It
472 * can be @NULL when we don't care about it.
473 * @page_mask_out: page mask for the translated address. This
474 * should only be meaningful for IOMMU translated
475 * addresses, since there may be huge pages that this bit
476 * would tell. It can be @NULL if we don't care about it.
477 * @is_write: whether the translation operation is for write
478 * @is_mmio: whether this can be MMIO, set true if it can
479 * @target_as: the address space targeted by the IOMMU
480 * @attrs: memory transaction attributes
482 * This function is called from RCU critical section
484 static MemoryRegionSection flatview_do_translate(FlatView *fv,
488 hwaddr *page_mask_out,
491 AddressSpace **target_as,
494 MemoryRegionSection *section;
495 IOMMUMemoryRegion *iommu_mr;
496 hwaddr plen = (hwaddr)(-1);
502 section = address_space_translate_internal(
503 flatview_to_dispatch(fv), addr, xlat,
506 iommu_mr = memory_region_get_iommu(section->mr);
507 if (unlikely(iommu_mr)) {
508 return address_space_translate_iommu(iommu_mr, xlat,
509 plen_out, page_mask_out,
514 /* Not behind an IOMMU, use default page size. */
515 *page_mask_out = ~TARGET_PAGE_MASK;
521 /* Called from RCU critical section */
522 IOMMUTLBEntry address_space_get_iotlb_entry(AddressSpace *as, hwaddr addr,
523 bool is_write, MemTxAttrs attrs)
525 MemoryRegionSection section;
526 hwaddr xlat, page_mask;
529 * This can never be MMIO, and we don't really care about plen,
532 section = flatview_do_translate(address_space_to_flatview(as), addr, &xlat,
533 NULL, &page_mask, is_write, false, &as,
536 /* Illegal translation */
537 if (section.mr == &io_mem_unassigned) {
541 /* Convert memory region offset into address space offset */
542 xlat += section.offset_within_address_space -
543 section.offset_within_region;
545 return (IOMMUTLBEntry) {
547 .iova = addr & ~page_mask,
548 .translated_addr = xlat & ~page_mask,
549 .addr_mask = page_mask,
550 /* IOTLBs are for DMAs, and DMA only allows on RAMs. */
555 return (IOMMUTLBEntry) {0};
558 /* Called from RCU critical section */
559 MemoryRegion *flatview_translate(FlatView *fv, hwaddr addr, hwaddr *xlat,
560 hwaddr *plen, bool is_write,
564 MemoryRegionSection section;
565 AddressSpace *as = NULL;
567 /* This can be MMIO, so setup MMIO bit. */
568 section = flatview_do_translate(fv, addr, xlat, plen, NULL,
569 is_write, true, &as, attrs);
572 if (xen_enabled() && memory_access_is_direct(mr, is_write)) {
573 hwaddr page = ((addr & TARGET_PAGE_MASK) + TARGET_PAGE_SIZE) - addr;
574 *plen = MIN(page, *plen);
580 typedef struct TCGIOMMUNotifier {
588 static void tcg_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
590 TCGIOMMUNotifier *notifier = container_of(n, TCGIOMMUNotifier, n);
592 if (!notifier->active) {
595 tlb_flush(notifier->cpu);
596 notifier->active = false;
597 /* We leave the notifier struct on the list to avoid reallocating it later.
598 * Generally the number of IOMMUs a CPU deals with will be small.
599 * In any case we can't unregister the iommu notifier from a notify
604 static void tcg_register_iommu_notifier(CPUState *cpu,
605 IOMMUMemoryRegion *iommu_mr,
608 /* Make sure this CPU has an IOMMU notifier registered for this
609 * IOMMU/IOMMU index combination, so that we can flush its TLB
610 * when the IOMMU tells us the mappings we've cached have changed.
612 MemoryRegion *mr = MEMORY_REGION(iommu_mr);
613 TCGIOMMUNotifier *notifier = NULL;
616 for (i = 0; i < cpu->iommu_notifiers->len; i++) {
617 notifier = g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier *, i);
618 if (notifier->mr == mr && notifier->iommu_idx == iommu_idx) {
622 if (i == cpu->iommu_notifiers->len) {
623 /* Not found, add a new entry at the end of the array */
624 cpu->iommu_notifiers = g_array_set_size(cpu->iommu_notifiers, i + 1);
625 notifier = g_new0(TCGIOMMUNotifier, 1);
626 g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier *, i) = notifier;
629 notifier->iommu_idx = iommu_idx;
631 /* Rather than trying to register interest in the specific part
632 * of the iommu's address space that we've accessed and then
633 * expand it later as subsequent accesses touch more of it, we
634 * just register interest in the whole thing, on the assumption
635 * that iommu reconfiguration will be rare.
637 iommu_notifier_init(¬ifier->n,
638 tcg_iommu_unmap_notify,
639 IOMMU_NOTIFIER_UNMAP,
643 memory_region_register_iommu_notifier(notifier->mr, ¬ifier->n,
647 if (!notifier->active) {
648 notifier->active = true;
652 void tcg_iommu_free_notifier_list(CPUState *cpu)
654 /* Destroy the CPU's notifier list */
656 TCGIOMMUNotifier *notifier;
658 for (i = 0; i < cpu->iommu_notifiers->len; i++) {
659 notifier = g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier *, i);
660 memory_region_unregister_iommu_notifier(notifier->mr, ¬ifier->n);
663 g_array_free(cpu->iommu_notifiers, true);
666 void tcg_iommu_init_notifier_list(CPUState *cpu)
668 cpu->iommu_notifiers = g_array_new(false, true, sizeof(TCGIOMMUNotifier *));
671 /* Called from RCU critical section */
672 MemoryRegionSection *
673 address_space_translate_for_iotlb(CPUState *cpu, int asidx, hwaddr orig_addr,
674 hwaddr *xlat, hwaddr *plen,
675 MemTxAttrs attrs, int *prot)
677 MemoryRegionSection *section;
678 IOMMUMemoryRegion *iommu_mr;
679 IOMMUMemoryRegionClass *imrc;
682 hwaddr addr = orig_addr;
683 AddressSpaceDispatch *d = cpu->cpu_ases[asidx].memory_dispatch;
686 section = address_space_translate_internal(d, addr, &addr, plen, false);
688 iommu_mr = memory_region_get_iommu(section->mr);
693 imrc = memory_region_get_iommu_class_nocheck(iommu_mr);
695 iommu_idx = imrc->attrs_to_index(iommu_mr, attrs);
696 tcg_register_iommu_notifier(cpu, iommu_mr, iommu_idx);
697 /* We need all the permissions, so pass IOMMU_NONE so the IOMMU
698 * doesn't short-cut its translation table walk.
700 iotlb = imrc->translate(iommu_mr, addr, IOMMU_NONE, iommu_idx);
701 addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
702 | (addr & iotlb.addr_mask));
703 /* Update the caller's prot bits to remove permissions the IOMMU
704 * is giving us a failure response for. If we get down to no
705 * permissions left at all we can give up now.
707 if (!(iotlb.perm & IOMMU_RO)) {
708 *prot &= ~(PAGE_READ | PAGE_EXEC);
710 if (!(iotlb.perm & IOMMU_WO)) {
711 *prot &= ~PAGE_WRITE;
718 d = flatview_to_dispatch(address_space_to_flatview(iotlb.target_as));
721 assert(!memory_region_is_iommu(section->mr));
727 * We should be given a page-aligned address -- certainly
728 * tlb_set_page_with_attrs() does so. The page offset of xlat
729 * is used to index sections[], and PHYS_SECTION_UNASSIGNED = 0.
730 * The page portion of xlat will be logged by memory_region_access_valid()
731 * when this memory access is rejected, so use the original untranslated
734 assert((orig_addr & ~TARGET_PAGE_MASK) == 0);
736 return &d->map.sections[PHYS_SECTION_UNASSIGNED];
739 void cpu_address_space_init(CPUState *cpu, int asidx,
740 const char *prefix, MemoryRegion *mr)
742 CPUAddressSpace *newas;
743 AddressSpace *as = g_new0(AddressSpace, 1);
747 as_name = g_strdup_printf("%s-%d", prefix, cpu->cpu_index);
748 address_space_init(as, mr, as_name);
751 /* Target code should have set num_ases before calling us */
752 assert(asidx < cpu->num_ases);
755 /* address space 0 gets the convenience alias */
759 /* KVM cannot currently support multiple address spaces. */
760 assert(asidx == 0 || !kvm_enabled());
762 if (!cpu->cpu_ases) {
763 cpu->cpu_ases = g_new0(CPUAddressSpace, cpu->num_ases);
766 newas = &cpu->cpu_ases[asidx];
770 newas->tcg_as_listener.log_global_after_sync = tcg_log_global_after_sync;
771 newas->tcg_as_listener.commit = tcg_commit;
772 newas->tcg_as_listener.name = "tcg";
773 memory_listener_register(&newas->tcg_as_listener, as);
777 AddressSpace *cpu_get_address_space(CPUState *cpu, int asidx)
779 /* Return the AddressSpace corresponding to the specified index */
780 return cpu->cpu_ases[asidx].as;
783 /* Called from RCU critical section */
784 static RAMBlock *qemu_get_ram_block(ram_addr_t addr)
788 block = qatomic_rcu_read(&ram_list.mru_block);
789 if (block && addr - block->offset < block->max_length) {
792 RAMBLOCK_FOREACH(block) {
793 if (addr - block->offset < block->max_length) {
798 fprintf(stderr, "Bad ram offset %" PRIx64 "\n", (uint64_t)addr);
802 /* It is safe to write mru_block outside the iothread lock. This
807 * xxx removed from list
811 * call_rcu(reclaim_ramblock, xxx);
814 * qatomic_rcu_set is not needed here. The block was already published
815 * when it was placed into the list. Here we're just making an extra
816 * copy of the pointer.
818 ram_list.mru_block = block;
822 static void tlb_reset_dirty_range_all(ram_addr_t start, ram_addr_t length)
829 assert(tcg_enabled());
830 end = TARGET_PAGE_ALIGN(start + length);
831 start &= TARGET_PAGE_MASK;
833 RCU_READ_LOCK_GUARD();
834 block = qemu_get_ram_block(start);
835 assert(block == qemu_get_ram_block(end - 1));
836 start1 = (uintptr_t)ramblock_ptr(block, start - block->offset);
838 tlb_reset_dirty(cpu, start1, length);
842 /* Note: start and end must be within the same ram block. */
843 bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start,
847 DirtyMemoryBlocks *blocks;
848 unsigned long end, page, start_page;
851 uint64_t mr_offset, mr_size;
857 end = TARGET_PAGE_ALIGN(start + length) >> TARGET_PAGE_BITS;
858 start_page = start >> TARGET_PAGE_BITS;
861 WITH_RCU_READ_LOCK_GUARD() {
862 blocks = qatomic_rcu_read(&ram_list.dirty_memory[client]);
863 ramblock = qemu_get_ram_block(start);
864 /* Range sanity check on the ramblock */
865 assert(start >= ramblock->offset &&
866 start + length <= ramblock->offset + ramblock->used_length);
869 unsigned long idx = page / DIRTY_MEMORY_BLOCK_SIZE;
870 unsigned long offset = page % DIRTY_MEMORY_BLOCK_SIZE;
871 unsigned long num = MIN(end - page,
872 DIRTY_MEMORY_BLOCK_SIZE - offset);
874 dirty |= bitmap_test_and_clear_atomic(blocks->blocks[idx],
879 mr_offset = (ram_addr_t)(start_page << TARGET_PAGE_BITS) - ramblock->offset;
880 mr_size = (end - start_page) << TARGET_PAGE_BITS;
881 memory_region_clear_dirty_bitmap(ramblock->mr, mr_offset, mr_size);
884 if (dirty && tcg_enabled()) {
885 tlb_reset_dirty_range_all(start, length);
891 DirtyBitmapSnapshot *cpu_physical_memory_snapshot_and_clear_dirty
892 (MemoryRegion *mr, hwaddr offset, hwaddr length, unsigned client)
894 DirtyMemoryBlocks *blocks;
895 ram_addr_t start = memory_region_get_ram_addr(mr) + offset;
896 unsigned long align = 1UL << (TARGET_PAGE_BITS + BITS_PER_LEVEL);
897 ram_addr_t first = QEMU_ALIGN_DOWN(start, align);
898 ram_addr_t last = QEMU_ALIGN_UP(start + length, align);
899 DirtyBitmapSnapshot *snap;
900 unsigned long page, end, dest;
902 snap = g_malloc0(sizeof(*snap) +
903 ((last - first) >> (TARGET_PAGE_BITS + 3)));
907 page = first >> TARGET_PAGE_BITS;
908 end = last >> TARGET_PAGE_BITS;
911 WITH_RCU_READ_LOCK_GUARD() {
912 blocks = qatomic_rcu_read(&ram_list.dirty_memory[client]);
915 unsigned long idx = page / DIRTY_MEMORY_BLOCK_SIZE;
916 unsigned long ofs = page % DIRTY_MEMORY_BLOCK_SIZE;
917 unsigned long num = MIN(end - page,
918 DIRTY_MEMORY_BLOCK_SIZE - ofs);
920 assert(QEMU_IS_ALIGNED(ofs, (1 << BITS_PER_LEVEL)));
921 assert(QEMU_IS_ALIGNED(num, (1 << BITS_PER_LEVEL)));
922 ofs >>= BITS_PER_LEVEL;
924 bitmap_copy_and_clear_atomic(snap->dirty + dest,
925 blocks->blocks[idx] + ofs,
928 dest += num >> BITS_PER_LEVEL;
933 tlb_reset_dirty_range_all(start, length);
936 memory_region_clear_dirty_bitmap(mr, offset, length);
941 bool cpu_physical_memory_snapshot_get_dirty(DirtyBitmapSnapshot *snap,
945 unsigned long page, end;
947 assert(start >= snap->start);
948 assert(start + length <= snap->end);
950 end = TARGET_PAGE_ALIGN(start + length - snap->start) >> TARGET_PAGE_BITS;
951 page = (start - snap->start) >> TARGET_PAGE_BITS;
954 if (test_bit(page, snap->dirty)) {
962 /* Called from RCU critical section */
963 hwaddr memory_region_section_get_iotlb(CPUState *cpu,
964 MemoryRegionSection *section)
966 AddressSpaceDispatch *d = flatview_to_dispatch(section->fv);
967 return section - d->map.sections;
970 static int subpage_register(subpage_t *mmio, uint32_t start, uint32_t end,
972 static subpage_t *subpage_init(FlatView *fv, hwaddr base);
974 static uint16_t phys_section_add(PhysPageMap *map,
975 MemoryRegionSection *section)
977 /* The physical section number is ORed with a page-aligned
978 * pointer to produce the iotlb entries. Thus it should
979 * never overflow into the page-aligned value.
981 assert(map->sections_nb < TARGET_PAGE_SIZE);
983 if (map->sections_nb == map->sections_nb_alloc) {
984 map->sections_nb_alloc = MAX(map->sections_nb_alloc * 2, 16);
985 map->sections = g_renew(MemoryRegionSection, map->sections,
986 map->sections_nb_alloc);
988 map->sections[map->sections_nb] = *section;
989 memory_region_ref(section->mr);
990 return map->sections_nb++;
993 static void phys_section_destroy(MemoryRegion *mr)
995 bool have_sub_page = mr->subpage;
997 memory_region_unref(mr);
1000 subpage_t *subpage = container_of(mr, subpage_t, iomem);
1001 object_unref(OBJECT(&subpage->iomem));
1006 static void phys_sections_free(PhysPageMap *map)
1008 while (map->sections_nb > 0) {
1009 MemoryRegionSection *section = &map->sections[--map->sections_nb];
1010 phys_section_destroy(section->mr);
1012 g_free(map->sections);
1016 static void register_subpage(FlatView *fv, MemoryRegionSection *section)
1018 AddressSpaceDispatch *d = flatview_to_dispatch(fv);
1020 hwaddr base = section->offset_within_address_space
1022 MemoryRegionSection *existing = phys_page_find(d, base);
1023 MemoryRegionSection subsection = {
1024 .offset_within_address_space = base,
1025 .size = int128_make64(TARGET_PAGE_SIZE),
1029 assert(existing->mr->subpage || existing->mr == &io_mem_unassigned);
1031 if (!(existing->mr->subpage)) {
1032 subpage = subpage_init(fv, base);
1034 subsection.mr = &subpage->iomem;
1035 phys_page_set(d, base >> TARGET_PAGE_BITS, 1,
1036 phys_section_add(&d->map, &subsection));
1038 subpage = container_of(existing->mr, subpage_t, iomem);
1040 start = section->offset_within_address_space & ~TARGET_PAGE_MASK;
1041 end = start + int128_get64(section->size) - 1;
1042 subpage_register(subpage, start, end,
1043 phys_section_add(&d->map, section));
1047 static void register_multipage(FlatView *fv,
1048 MemoryRegionSection *section)
1050 AddressSpaceDispatch *d = flatview_to_dispatch(fv);
1051 hwaddr start_addr = section->offset_within_address_space;
1052 uint16_t section_index = phys_section_add(&d->map, section);
1053 uint64_t num_pages = int128_get64(int128_rshift(section->size,
1057 phys_page_set(d, start_addr >> TARGET_PAGE_BITS, num_pages, section_index);
1061 * The range in *section* may look like this:
1065 * where s stands for subpage and P for page.
1067 void flatview_add_to_dispatch(FlatView *fv, MemoryRegionSection *section)
1069 MemoryRegionSection remain = *section;
1070 Int128 page_size = int128_make64(TARGET_PAGE_SIZE);
1072 /* register first subpage */
1073 if (remain.offset_within_address_space & ~TARGET_PAGE_MASK) {
1074 uint64_t left = TARGET_PAGE_ALIGN(remain.offset_within_address_space)
1075 - remain.offset_within_address_space;
1077 MemoryRegionSection now = remain;
1078 now.size = int128_min(int128_make64(left), now.size);
1079 register_subpage(fv, &now);
1080 if (int128_eq(remain.size, now.size)) {
1083 remain.size = int128_sub(remain.size, now.size);
1084 remain.offset_within_address_space += int128_get64(now.size);
1085 remain.offset_within_region += int128_get64(now.size);
1088 /* register whole pages */
1089 if (int128_ge(remain.size, page_size)) {
1090 MemoryRegionSection now = remain;
1091 now.size = int128_and(now.size, int128_neg(page_size));
1092 register_multipage(fv, &now);
1093 if (int128_eq(remain.size, now.size)) {
1096 remain.size = int128_sub(remain.size, now.size);
1097 remain.offset_within_address_space += int128_get64(now.size);
1098 remain.offset_within_region += int128_get64(now.size);
1101 /* register last subpage */
1102 register_subpage(fv, &remain);
1105 void qemu_flush_coalesced_mmio_buffer(void)
1108 kvm_flush_coalesced_mmio_buffer();
1111 void qemu_mutex_lock_ramlist(void)
1113 qemu_mutex_lock(&ram_list.mutex);
1116 void qemu_mutex_unlock_ramlist(void)
1118 qemu_mutex_unlock(&ram_list.mutex);
1121 GString *ram_block_format(void)
1125 GString *buf = g_string_new("");
1127 RCU_READ_LOCK_GUARD();
1128 g_string_append_printf(buf, "%24s %8s %18s %18s %18s %18s %3s\n",
1129 "Block Name", "PSize", "Offset", "Used", "Total",
1132 RAMBLOCK_FOREACH(block) {
1133 psize = size_to_str(block->page_size);
1134 g_string_append_printf(buf, "%24s %8s 0x%016" PRIx64 " 0x%016" PRIx64
1135 " 0x%016" PRIx64 " 0x%016" PRIx64 " %3s\n",
1136 block->idstr, psize,
1137 (uint64_t)block->offset,
1138 (uint64_t)block->used_length,
1139 (uint64_t)block->max_length,
1140 (uint64_t)(uintptr_t)block->host,
1141 block->mr->readonly ? "ro" : "rw");
1149 static int find_min_backend_pagesize(Object *obj, void *opaque)
1151 long *hpsize_min = opaque;
1153 if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
1154 HostMemoryBackend *backend = MEMORY_BACKEND(obj);
1155 long hpsize = host_memory_backend_pagesize(backend);
1157 if (host_memory_backend_is_mapped(backend) && (hpsize < *hpsize_min)) {
1158 *hpsize_min = hpsize;
1165 static int find_max_backend_pagesize(Object *obj, void *opaque)
1167 long *hpsize_max = opaque;
1169 if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
1170 HostMemoryBackend *backend = MEMORY_BACKEND(obj);
1171 long hpsize = host_memory_backend_pagesize(backend);
1173 if (host_memory_backend_is_mapped(backend) && (hpsize > *hpsize_max)) {
1174 *hpsize_max = hpsize;
1182 * TODO: We assume right now that all mapped host memory backends are
1183 * used as RAM, however some might be used for different purposes.
1185 long qemu_minrampagesize(void)
1187 long hpsize = LONG_MAX;
1188 Object *memdev_root = object_resolve_path("/objects", NULL);
1190 object_child_foreach(memdev_root, find_min_backend_pagesize, &hpsize);
1194 long qemu_maxrampagesize(void)
1197 Object *memdev_root = object_resolve_path("/objects", NULL);
1199 object_child_foreach(memdev_root, find_max_backend_pagesize, &pagesize);
1204 static int64_t get_file_size(int fd)
1207 #if defined(__linux__)
1210 if (fstat(fd, &st) < 0) {
1214 /* Special handling for devdax character devices */
1215 if (S_ISCHR(st.st_mode)) {
1216 g_autofree char *subsystem_path = NULL;
1217 g_autofree char *subsystem = NULL;
1219 subsystem_path = g_strdup_printf("/sys/dev/char/%d:%d/subsystem",
1220 major(st.st_rdev), minor(st.st_rdev));
1221 subsystem = g_file_read_link(subsystem_path, NULL);
1223 if (subsystem && g_str_has_suffix(subsystem, "/dax")) {
1224 g_autofree char *size_path = NULL;
1225 g_autofree char *size_str = NULL;
1227 size_path = g_strdup_printf("/sys/dev/char/%d:%d/size",
1228 major(st.st_rdev), minor(st.st_rdev));
1230 if (g_file_get_contents(size_path, &size_str, NULL, NULL)) {
1231 return g_ascii_strtoll(size_str, NULL, 0);
1235 #endif /* defined(__linux__) */
1237 /* st.st_size may be zero for special files yet lseek(2) works */
1238 size = lseek(fd, 0, SEEK_END);
1245 static int64_t get_file_align(int fd)
1248 #if defined(__linux__) && defined(CONFIG_LIBDAXCTL)
1251 if (fstat(fd, &st) < 0) {
1255 /* Special handling for devdax character devices */
1256 if (S_ISCHR(st.st_mode)) {
1257 g_autofree char *path = NULL;
1258 g_autofree char *rpath = NULL;
1259 struct daxctl_ctx *ctx;
1260 struct daxctl_region *region;
1263 path = g_strdup_printf("/sys/dev/char/%d:%d",
1264 major(st.st_rdev), minor(st.st_rdev));
1265 rpath = realpath(path, NULL);
1270 rc = daxctl_new(&ctx);
1275 daxctl_region_foreach(ctx, region) {
1276 if (strstr(rpath, daxctl_region_get_path(region))) {
1277 align = daxctl_region_get_align(region);
1283 #endif /* defined(__linux__) && defined(CONFIG_LIBDAXCTL) */
1288 static int file_ram_open(const char *path,
1289 const char *region_name,
1294 char *sanitized_name;
1300 fd = open(path, readonly ? O_RDONLY : O_RDWR);
1303 * open(O_RDONLY) won't fail with EISDIR. Check manually if we
1304 * opened a directory and fail similarly to how we fail ENOENT
1305 * in readonly mode. Note that mkstemp() would imply O_RDWR.
1308 struct stat file_stat;
1310 if (fstat(fd, &file_stat)) {
1312 if (errno == EINTR) {
1316 } else if (S_ISDIR(file_stat.st_mode)) {
1321 /* @path names an existing file, use it */
1324 if (errno == ENOENT) {
1326 /* Refuse to create new, readonly files. */
1329 /* @path names a file that doesn't exist, create it */
1330 fd = open(path, O_RDWR | O_CREAT | O_EXCL, 0644);
1335 } else if (errno == EISDIR) {
1336 /* @path names a directory, create a file there */
1337 /* Make name safe to use with mkstemp by replacing '/' with '_'. */
1338 sanitized_name = g_strdup(region_name);
1339 for (c = sanitized_name; *c != '\0'; c++) {
1345 filename = g_strdup_printf("%s/qemu_back_mem.%s.XXXXXX", path,
1347 g_free(sanitized_name);
1349 fd = mkstemp(filename);
1357 if (errno != EEXIST && errno != EINTR) {
1361 * Try again on EINTR and EEXIST. The latter happens when
1362 * something else creates the file between our two open().
1369 static void *file_ram_alloc(RAMBlock *block,
1376 uint32_t qemu_map_flags;
1379 block->page_size = qemu_fd_getpagesize(fd);
1380 if (block->mr->align % block->page_size) {
1381 error_setg(errp, "alignment 0x%" PRIx64
1382 " must be multiples of page size 0x%zx",
1383 block->mr->align, block->page_size);
1385 } else if (block->mr->align && !is_power_of_2(block->mr->align)) {
1386 error_setg(errp, "alignment 0x%" PRIx64
1387 " must be a power of two", block->mr->align);
1389 } else if (offset % block->page_size) {
1390 error_setg(errp, "offset 0x%" PRIx64
1391 " must be multiples of page size 0x%zx",
1392 offset, block->page_size);
1395 block->mr->align = MAX(block->page_size, block->mr->align);
1396 #if defined(__s390x__)
1397 if (kvm_enabled()) {
1398 block->mr->align = MAX(block->mr->align, QEMU_VMALLOC_ALIGN);
1402 if (memory < block->page_size) {
1403 error_setg(errp, "memory size 0x" RAM_ADDR_FMT " must be equal to "
1404 "or larger than page size 0x%zx",
1405 memory, block->page_size);
1409 memory = ROUND_UP(memory, block->page_size);
1412 * ftruncate is not supported by hugetlbfs in older
1413 * hosts, so don't bother bailing out on errors.
1414 * If anything goes wrong with it under other filesystems,
1417 * Do not truncate the non-empty backend file to avoid corrupting
1418 * the existing data in the file. Disabling shrinking is not
1419 * enough. For example, the current vNVDIMM implementation stores
1420 * the guest NVDIMM labels at the end of the backend file. If the
1421 * backend file is later extended, QEMU will not be able to find
1422 * those labels. Therefore, extending the non-empty backend file
1423 * is disabled as well.
1425 if (truncate && ftruncate(fd, offset + memory)) {
1426 perror("ftruncate");
1429 qemu_map_flags = (block->flags & RAM_READONLY) ? QEMU_MAP_READONLY : 0;
1430 qemu_map_flags |= (block->flags & RAM_SHARED) ? QEMU_MAP_SHARED : 0;
1431 qemu_map_flags |= (block->flags & RAM_PMEM) ? QEMU_MAP_SYNC : 0;
1432 qemu_map_flags |= (block->flags & RAM_NORESERVE) ? QEMU_MAP_NORESERVE : 0;
1433 area = qemu_ram_mmap(fd, memory, block->mr->align, qemu_map_flags, offset);
1434 if (area == MAP_FAILED) {
1435 error_setg_errno(errp, errno,
1436 "unable to map backing store for guest RAM");
1441 block->fd_offset = offset;
1446 /* Allocate space within the ram_addr_t space that governs the
1448 * Called with the ramlist lock held.
1450 static ram_addr_t find_ram_offset(ram_addr_t size)
1452 RAMBlock *block, *next_block;
1453 ram_addr_t offset = RAM_ADDR_MAX, mingap = RAM_ADDR_MAX;
1455 assert(size != 0); /* it would hand out same offset multiple times */
1457 if (QLIST_EMPTY_RCU(&ram_list.blocks)) {
1461 RAMBLOCK_FOREACH(block) {
1462 ram_addr_t candidate, next = RAM_ADDR_MAX;
1464 /* Align blocks to start on a 'long' in the bitmap
1465 * which makes the bitmap sync'ing take the fast path.
1467 candidate = block->offset + block->max_length;
1468 candidate = ROUND_UP(candidate, BITS_PER_LONG << TARGET_PAGE_BITS);
1470 /* Search for the closest following block
1473 RAMBLOCK_FOREACH(next_block) {
1474 if (next_block->offset >= candidate) {
1475 next = MIN(next, next_block->offset);
1479 /* If it fits remember our place and remember the size
1480 * of gap, but keep going so that we might find a smaller
1481 * gap to fill so avoiding fragmentation.
1483 if (next - candidate >= size && next - candidate < mingap) {
1485 mingap = next - candidate;
1488 trace_find_ram_offset_loop(size, candidate, offset, next, mingap);
1491 if (offset == RAM_ADDR_MAX) {
1492 fprintf(stderr, "Failed to find gap of requested size: %" PRIu64 "\n",
1497 trace_find_ram_offset(size, offset);
1502 static unsigned long last_ram_page(void)
1505 ram_addr_t last = 0;
1507 RCU_READ_LOCK_GUARD();
1508 RAMBLOCK_FOREACH(block) {
1509 last = MAX(last, block->offset + block->max_length);
1511 return last >> TARGET_PAGE_BITS;
1514 static void qemu_ram_setup_dump(void *addr, ram_addr_t size)
1518 /* Use MADV_DONTDUMP, if user doesn't want the guest memory in the core */
1519 if (!machine_dump_guest_core(current_machine)) {
1520 ret = qemu_madvise(addr, size, QEMU_MADV_DONTDUMP);
1522 perror("qemu_madvise");
1523 fprintf(stderr, "madvise doesn't support MADV_DONTDUMP, "
1524 "but dump_guest_core=off specified\n");
1529 const char *qemu_ram_get_idstr(RAMBlock *rb)
1534 void *qemu_ram_get_host_addr(RAMBlock *rb)
1539 ram_addr_t qemu_ram_get_offset(RAMBlock *rb)
1544 ram_addr_t qemu_ram_get_used_length(RAMBlock *rb)
1546 return rb->used_length;
1549 ram_addr_t qemu_ram_get_max_length(RAMBlock *rb)
1551 return rb->max_length;
1554 bool qemu_ram_is_shared(RAMBlock *rb)
1556 return rb->flags & RAM_SHARED;
1559 bool qemu_ram_is_noreserve(RAMBlock *rb)
1561 return rb->flags & RAM_NORESERVE;
1564 /* Note: Only set at the start of postcopy */
1565 bool qemu_ram_is_uf_zeroable(RAMBlock *rb)
1567 return rb->flags & RAM_UF_ZEROPAGE;
1570 void qemu_ram_set_uf_zeroable(RAMBlock *rb)
1572 rb->flags |= RAM_UF_ZEROPAGE;
1575 bool qemu_ram_is_migratable(RAMBlock *rb)
1577 return rb->flags & RAM_MIGRATABLE;
1580 void qemu_ram_set_migratable(RAMBlock *rb)
1582 rb->flags |= RAM_MIGRATABLE;
1585 void qemu_ram_unset_migratable(RAMBlock *rb)
1587 rb->flags &= ~RAM_MIGRATABLE;
1590 bool qemu_ram_is_named_file(RAMBlock *rb)
1592 return rb->flags & RAM_NAMED_FILE;
1595 int qemu_ram_get_fd(RAMBlock *rb)
1600 /* Called with iothread lock held. */
1601 void qemu_ram_set_idstr(RAMBlock *new_block, const char *name, DeviceState *dev)
1606 assert(!new_block->idstr[0]);
1609 char *id = qdev_get_dev_path(dev);
1611 snprintf(new_block->idstr, sizeof(new_block->idstr), "%s/", id);
1615 pstrcat(new_block->idstr, sizeof(new_block->idstr), name);
1617 RCU_READ_LOCK_GUARD();
1618 RAMBLOCK_FOREACH(block) {
1619 if (block != new_block &&
1620 !strcmp(block->idstr, new_block->idstr)) {
1621 fprintf(stderr, "RAMBlock \"%s\" already registered, abort!\n",
1628 /* Called with iothread lock held. */
1629 void qemu_ram_unset_idstr(RAMBlock *block)
1631 /* FIXME: arch_init.c assumes that this is not called throughout
1632 * migration. Ignore the problem since hot-unplug during migration
1633 * does not work anyway.
1636 memset(block->idstr, 0, sizeof(block->idstr));
1640 size_t qemu_ram_pagesize(RAMBlock *rb)
1642 return rb->page_size;
1645 /* Returns the largest size of page in use */
1646 size_t qemu_ram_pagesize_largest(void)
1651 RAMBLOCK_FOREACH(block) {
1652 largest = MAX(largest, qemu_ram_pagesize(block));
1658 static int memory_try_enable_merging(void *addr, size_t len)
1660 if (!machine_mem_merge(current_machine)) {
1661 /* disabled by the user */
1665 return qemu_madvise(addr, len, QEMU_MADV_MERGEABLE);
1669 * Resizing RAM while migrating can result in the migration being canceled.
1670 * Care has to be taken if the guest might have already detected the memory.
1672 * As memory core doesn't know how is memory accessed, it is up to
1673 * resize callback to update device state and/or add assertions to detect
1674 * misuse, if necessary.
1676 int qemu_ram_resize(RAMBlock *block, ram_addr_t newsize, Error **errp)
1678 const ram_addr_t oldsize = block->used_length;
1679 const ram_addr_t unaligned_size = newsize;
1683 newsize = HOST_PAGE_ALIGN(newsize);
1685 if (block->used_length == newsize) {
1687 * We don't have to resize the ram block (which only knows aligned
1688 * sizes), however, we have to notify if the unaligned size changed.
1690 if (unaligned_size != memory_region_size(block->mr)) {
1691 memory_region_set_size(block->mr, unaligned_size);
1692 if (block->resized) {
1693 block->resized(block->idstr, unaligned_size, block->host);
1699 if (!(block->flags & RAM_RESIZEABLE)) {
1700 error_setg_errno(errp, EINVAL,
1701 "Size mismatch: %s: 0x" RAM_ADDR_FMT
1702 " != 0x" RAM_ADDR_FMT, block->idstr,
1703 newsize, block->used_length);
1707 if (block->max_length < newsize) {
1708 error_setg_errno(errp, EINVAL,
1709 "Size too large: %s: 0x" RAM_ADDR_FMT
1710 " > 0x" RAM_ADDR_FMT, block->idstr,
1711 newsize, block->max_length);
1715 /* Notify before modifying the ram block and touching the bitmaps. */
1717 ram_block_notify_resize(block->host, oldsize, newsize);
1720 cpu_physical_memory_clear_dirty_range(block->offset, block->used_length);
1721 block->used_length = newsize;
1722 cpu_physical_memory_set_dirty_range(block->offset, block->used_length,
1724 memory_region_set_size(block->mr, unaligned_size);
1725 if (block->resized) {
1726 block->resized(block->idstr, unaligned_size, block->host);
1732 * Trigger sync on the given ram block for range [start, start + length]
1733 * with the backing store if one is available.
1735 * @Note: this is supposed to be a synchronous op.
1737 void qemu_ram_msync(RAMBlock *block, ram_addr_t start, ram_addr_t length)
1739 /* The requested range should fit in within the block range */
1740 g_assert((start + length) <= block->used_length);
1742 #ifdef CONFIG_LIBPMEM
1743 /* The lack of support for pmem should not block the sync */
1744 if (ramblock_is_pmem(block)) {
1745 void *addr = ramblock_ptr(block, start);
1746 pmem_persist(addr, length);
1750 if (block->fd >= 0) {
1752 * Case there is no support for PMEM or the memory has not been
1753 * specified as persistent (or is not one) - use the msync.
1754 * Less optimal but still achieves the same goal
1756 void *addr = ramblock_ptr(block, start);
1757 if (qemu_msync(addr, length, block->fd)) {
1758 warn_report("%s: failed to sync memory range: start: "
1759 RAM_ADDR_FMT " length: " RAM_ADDR_FMT,
1760 __func__, start, length);
1765 /* Called with ram_list.mutex held */
1766 static void dirty_memory_extend(ram_addr_t old_ram_size,
1767 ram_addr_t new_ram_size)
1769 ram_addr_t old_num_blocks = DIV_ROUND_UP(old_ram_size,
1770 DIRTY_MEMORY_BLOCK_SIZE);
1771 ram_addr_t new_num_blocks = DIV_ROUND_UP(new_ram_size,
1772 DIRTY_MEMORY_BLOCK_SIZE);
1775 /* Only need to extend if block count increased */
1776 if (new_num_blocks <= old_num_blocks) {
1780 for (i = 0; i < DIRTY_MEMORY_NUM; i++) {
1781 DirtyMemoryBlocks *old_blocks;
1782 DirtyMemoryBlocks *new_blocks;
1785 old_blocks = qatomic_rcu_read(&ram_list.dirty_memory[i]);
1786 new_blocks = g_malloc(sizeof(*new_blocks) +
1787 sizeof(new_blocks->blocks[0]) * new_num_blocks);
1789 if (old_num_blocks) {
1790 memcpy(new_blocks->blocks, old_blocks->blocks,
1791 old_num_blocks * sizeof(old_blocks->blocks[0]));
1794 for (j = old_num_blocks; j < new_num_blocks; j++) {
1795 new_blocks->blocks[j] = bitmap_new(DIRTY_MEMORY_BLOCK_SIZE);
1798 qatomic_rcu_set(&ram_list.dirty_memory[i], new_blocks);
1801 g_free_rcu(old_blocks, rcu);
1806 static void ram_block_add(RAMBlock *new_block, Error **errp)
1808 const bool noreserve = qemu_ram_is_noreserve(new_block);
1809 const bool shared = qemu_ram_is_shared(new_block);
1811 RAMBlock *last_block = NULL;
1812 ram_addr_t old_ram_size, new_ram_size;
1815 old_ram_size = last_ram_page();
1817 qemu_mutex_lock_ramlist();
1818 new_block->offset = find_ram_offset(new_block->max_length);
1820 if (!new_block->host) {
1821 if (xen_enabled()) {
1822 xen_ram_alloc(new_block->offset, new_block->max_length,
1823 new_block->mr, &err);
1825 error_propagate(errp, err);
1826 qemu_mutex_unlock_ramlist();
1830 new_block->host = qemu_anon_ram_alloc(new_block->max_length,
1831 &new_block->mr->align,
1833 if (!new_block->host) {
1834 error_setg_errno(errp, errno,
1835 "cannot set up guest memory '%s'",
1836 memory_region_name(new_block->mr));
1837 qemu_mutex_unlock_ramlist();
1840 memory_try_enable_merging(new_block->host, new_block->max_length);
1844 new_ram_size = MAX(old_ram_size,
1845 (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS);
1846 if (new_ram_size > old_ram_size) {
1847 dirty_memory_extend(old_ram_size, new_ram_size);
1849 /* Keep the list sorted from biggest to smallest block. Unlike QTAILQ,
1850 * QLIST (which has an RCU-friendly variant) does not have insertion at
1851 * tail, so save the last element in last_block.
1853 RAMBLOCK_FOREACH(block) {
1855 if (block->max_length < new_block->max_length) {
1860 QLIST_INSERT_BEFORE_RCU(block, new_block, next);
1861 } else if (last_block) {
1862 QLIST_INSERT_AFTER_RCU(last_block, new_block, next);
1863 } else { /* list is empty */
1864 QLIST_INSERT_HEAD_RCU(&ram_list.blocks, new_block, next);
1866 ram_list.mru_block = NULL;
1868 /* Write list before version */
1871 qemu_mutex_unlock_ramlist();
1873 cpu_physical_memory_set_dirty_range(new_block->offset,
1874 new_block->used_length,
1877 if (new_block->host) {
1878 qemu_ram_setup_dump(new_block->host, new_block->max_length);
1879 qemu_madvise(new_block->host, new_block->max_length, QEMU_MADV_HUGEPAGE);
1881 * MADV_DONTFORK is also needed by KVM in absence of synchronous MMU
1882 * Configure it unless the machine is a qtest server, in which case
1883 * KVM is not used and it may be forked (eg for fuzzing purposes).
1885 if (!qtest_enabled()) {
1886 qemu_madvise(new_block->host, new_block->max_length,
1887 QEMU_MADV_DONTFORK);
1889 ram_block_notify_add(new_block->host, new_block->used_length,
1890 new_block->max_length);
1895 RAMBlock *qemu_ram_alloc_from_fd(ram_addr_t size, MemoryRegion *mr,
1896 uint32_t ram_flags, int fd, off_t offset,
1899 RAMBlock *new_block;
1900 Error *local_err = NULL;
1901 int64_t file_size, file_align;
1903 /* Just support these ram flags by now. */
1904 assert((ram_flags & ~(RAM_SHARED | RAM_PMEM | RAM_NORESERVE |
1905 RAM_PROTECTED | RAM_NAMED_FILE | RAM_READONLY |
1906 RAM_READONLY_FD)) == 0);
1908 if (xen_enabled()) {
1909 error_setg(errp, "-mem-path not supported with Xen");
1913 if (kvm_enabled() && !kvm_has_sync_mmu()) {
1915 "host lacks kvm mmu notifiers, -mem-path unsupported");
1919 size = HOST_PAGE_ALIGN(size);
1920 file_size = get_file_size(fd);
1921 if (file_size > offset && file_size < (offset + size)) {
1922 error_setg(errp, "backing store size 0x%" PRIx64
1923 " does not match 'size' option 0x" RAM_ADDR_FMT,
1928 file_align = get_file_align(fd);
1929 if (file_align > 0 && file_align > mr->align) {
1930 error_setg(errp, "backing store align 0x%" PRIx64
1931 " is larger than 'align' option 0x%" PRIx64,
1932 file_align, mr->align);
1936 new_block = g_malloc0(sizeof(*new_block));
1938 new_block->used_length = size;
1939 new_block->max_length = size;
1940 new_block->flags = ram_flags;
1941 new_block->host = file_ram_alloc(new_block, size, fd, !file_size, offset,
1943 if (!new_block->host) {
1948 ram_block_add(new_block, &local_err);
1951 error_propagate(errp, local_err);
1959 RAMBlock *qemu_ram_alloc_from_file(ram_addr_t size, MemoryRegion *mr,
1960 uint32_t ram_flags, const char *mem_path,
1961 off_t offset, Error **errp)
1967 fd = file_ram_open(mem_path, memory_region_name(mr),
1968 !!(ram_flags & RAM_READONLY_FD), &created);
1970 error_setg_errno(errp, -fd, "can't open backing store %s for guest RAM",
1972 if (!(ram_flags & RAM_READONLY_FD) && !(ram_flags & RAM_SHARED) &&
1975 * If we can open the file R/O (note: will never create a new file)
1976 * and we are dealing with a private mapping, there are still ways
1977 * to consume such files and get RAM instead of ROM.
1979 fd = file_ram_open(mem_path, memory_region_name(mr), true,
1986 error_append_hint(errp, "Consider opening the backing store"
1987 " read-only but still creating writable RAM using"
1988 " '-object memory-backend-file,readonly=on,rom=off...'"
1989 " (see \"VM templating\" documentation)\n");
1994 block = qemu_ram_alloc_from_fd(size, mr, ram_flags, fd, offset, errp);
2008 RAMBlock *qemu_ram_alloc_internal(ram_addr_t size, ram_addr_t max_size,
2009 void (*resized)(const char*,
2012 void *host, uint32_t ram_flags,
2013 MemoryRegion *mr, Error **errp)
2015 RAMBlock *new_block;
2016 Error *local_err = NULL;
2018 assert((ram_flags & ~(RAM_SHARED | RAM_RESIZEABLE | RAM_PREALLOC |
2019 RAM_NORESERVE)) == 0);
2020 assert(!host ^ (ram_flags & RAM_PREALLOC));
2022 size = HOST_PAGE_ALIGN(size);
2023 max_size = HOST_PAGE_ALIGN(max_size);
2024 new_block = g_malloc0(sizeof(*new_block));
2026 new_block->resized = resized;
2027 new_block->used_length = size;
2028 new_block->max_length = max_size;
2029 assert(max_size >= size);
2031 new_block->page_size = qemu_real_host_page_size();
2032 new_block->host = host;
2033 new_block->flags = ram_flags;
2034 ram_block_add(new_block, &local_err);
2037 error_propagate(errp, local_err);
2043 RAMBlock *qemu_ram_alloc_from_ptr(ram_addr_t size, void *host,
2044 MemoryRegion *mr, Error **errp)
2046 return qemu_ram_alloc_internal(size, size, NULL, host, RAM_PREALLOC, mr,
2050 RAMBlock *qemu_ram_alloc(ram_addr_t size, uint32_t ram_flags,
2051 MemoryRegion *mr, Error **errp)
2053 assert((ram_flags & ~(RAM_SHARED | RAM_NORESERVE)) == 0);
2054 return qemu_ram_alloc_internal(size, size, NULL, NULL, ram_flags, mr, errp);
2057 RAMBlock *qemu_ram_alloc_resizeable(ram_addr_t size, ram_addr_t maxsz,
2058 void (*resized)(const char*,
2061 MemoryRegion *mr, Error **errp)
2063 return qemu_ram_alloc_internal(size, maxsz, resized, NULL,
2064 RAM_RESIZEABLE, mr, errp);
2067 static void reclaim_ramblock(RAMBlock *block)
2069 if (block->flags & RAM_PREALLOC) {
2071 } else if (xen_enabled()) {
2072 xen_invalidate_map_cache_entry(block->host);
2074 } else if (block->fd >= 0) {
2075 qemu_ram_munmap(block->fd, block->host, block->max_length);
2079 qemu_anon_ram_free(block->host, block->max_length);
2084 void qemu_ram_free(RAMBlock *block)
2091 ram_block_notify_remove(block->host, block->used_length,
2095 qemu_mutex_lock_ramlist();
2096 QLIST_REMOVE_RCU(block, next);
2097 ram_list.mru_block = NULL;
2098 /* Write list before version */
2101 call_rcu(block, reclaim_ramblock, rcu);
2102 qemu_mutex_unlock_ramlist();
2106 void qemu_ram_remap(ram_addr_t addr, ram_addr_t length)
2114 RAMBLOCK_FOREACH(block) {
2115 offset = addr - block->offset;
2116 if (offset < block->max_length) {
2117 vaddr = ramblock_ptr(block, offset);
2118 if (block->flags & RAM_PREALLOC) {
2120 } else if (xen_enabled()) {
2124 flags |= block->flags & RAM_SHARED ?
2125 MAP_SHARED : MAP_PRIVATE;
2126 flags |= block->flags & RAM_NORESERVE ? MAP_NORESERVE : 0;
2128 prot |= block->flags & RAM_READONLY ? 0 : PROT_WRITE;
2129 if (block->fd >= 0) {
2130 area = mmap(vaddr, length, prot, flags, block->fd,
2131 offset + block->fd_offset);
2133 flags |= MAP_ANONYMOUS;
2134 area = mmap(vaddr, length, prot, flags, -1, 0);
2136 if (area != vaddr) {
2137 error_report("Could not remap addr: "
2138 RAM_ADDR_FMT "@" RAM_ADDR_FMT "",
2142 memory_try_enable_merging(vaddr, length);
2143 qemu_ram_setup_dump(vaddr, length);
2148 #endif /* !_WIN32 */
2150 /* Return a host pointer to ram allocated with qemu_ram_alloc.
2151 * This should not be used for general purpose DMA. Use address_space_map
2152 * or address_space_rw instead. For local memory (e.g. video ram) that the
2153 * device owns, use memory_region_get_ram_ptr.
2155 * Called within RCU critical section.
2157 void *qemu_map_ram_ptr(RAMBlock *ram_block, ram_addr_t addr)
2159 RAMBlock *block = ram_block;
2161 if (block == NULL) {
2162 block = qemu_get_ram_block(addr);
2163 addr -= block->offset;
2166 if (xen_enabled() && block->host == NULL) {
2167 /* We need to check if the requested address is in the RAM
2168 * because we don't want to map the entire memory in QEMU.
2169 * In that case just map until the end of the page.
2171 if (block->offset == 0) {
2172 return xen_map_cache(addr, 0, 0, false);
2175 block->host = xen_map_cache(block->offset, block->max_length, 1, false);
2177 return ramblock_ptr(block, addr);
2180 /* Return a host pointer to guest's ram. Similar to qemu_map_ram_ptr
2181 * but takes a size argument.
2183 * Called within RCU critical section.
2185 static void *qemu_ram_ptr_length(RAMBlock *ram_block, ram_addr_t addr,
2186 hwaddr *size, bool lock)
2188 RAMBlock *block = ram_block;
2193 if (block == NULL) {
2194 block = qemu_get_ram_block(addr);
2195 addr -= block->offset;
2197 *size = MIN(*size, block->max_length - addr);
2199 if (xen_enabled() && block->host == NULL) {
2200 /* We need to check if the requested address is in the RAM
2201 * because we don't want to map the entire memory in QEMU.
2202 * In that case just map the requested area.
2204 if (block->offset == 0) {
2205 return xen_map_cache(addr, *size, lock, lock);
2208 block->host = xen_map_cache(block->offset, block->max_length, 1, lock);
2211 return ramblock_ptr(block, addr);
2214 /* Return the offset of a hostpointer within a ramblock */
2215 ram_addr_t qemu_ram_block_host_offset(RAMBlock *rb, void *host)
2217 ram_addr_t res = (uint8_t *)host - (uint8_t *)rb->host;
2218 assert((uintptr_t)host >= (uintptr_t)rb->host);
2219 assert(res < rb->max_length);
2224 RAMBlock *qemu_ram_block_from_host(void *ptr, bool round_offset,
2228 uint8_t *host = ptr;
2230 if (xen_enabled()) {
2231 ram_addr_t ram_addr;
2232 RCU_READ_LOCK_GUARD();
2233 ram_addr = xen_ram_addr_from_mapcache(ptr);
2234 block = qemu_get_ram_block(ram_addr);
2236 *offset = ram_addr - block->offset;
2241 RCU_READ_LOCK_GUARD();
2242 block = qatomic_rcu_read(&ram_list.mru_block);
2243 if (block && block->host && host - block->host < block->max_length) {
2247 RAMBLOCK_FOREACH(block) {
2248 /* This case append when the block is not mapped. */
2249 if (block->host == NULL) {
2252 if (host - block->host < block->max_length) {
2260 *offset = (host - block->host);
2262 *offset &= TARGET_PAGE_MASK;
2268 * Finds the named RAMBlock
2270 * name: The name of RAMBlock to find
2272 * Returns: RAMBlock (or NULL if not found)
2274 RAMBlock *qemu_ram_block_by_name(const char *name)
2278 RAMBLOCK_FOREACH(block) {
2279 if (!strcmp(name, block->idstr)) {
2288 * Some of the system routines need to translate from a host pointer
2289 * (typically a TLB entry) back to a ram offset.
2291 ram_addr_t qemu_ram_addr_from_host(void *ptr)
2296 block = qemu_ram_block_from_host(ptr, false, &offset);
2298 return RAM_ADDR_INVALID;
2301 return block->offset + offset;
2304 ram_addr_t qemu_ram_addr_from_host_nofail(void *ptr)
2306 ram_addr_t ram_addr;
2308 ram_addr = qemu_ram_addr_from_host(ptr);
2309 if (ram_addr == RAM_ADDR_INVALID) {
2310 error_report("Bad ram pointer %p", ptr);
2316 static MemTxResult flatview_read(FlatView *fv, hwaddr addr,
2317 MemTxAttrs attrs, void *buf, hwaddr len);
2318 static MemTxResult flatview_write(FlatView *fv, hwaddr addr, MemTxAttrs attrs,
2319 const void *buf, hwaddr len);
2320 static bool flatview_access_valid(FlatView *fv, hwaddr addr, hwaddr len,
2321 bool is_write, MemTxAttrs attrs);
2323 static MemTxResult subpage_read(void *opaque, hwaddr addr, uint64_t *data,
2324 unsigned len, MemTxAttrs attrs)
2326 subpage_t *subpage = opaque;
2330 #if defined(DEBUG_SUBPAGE)
2331 printf("%s: subpage %p len %u addr " HWADDR_FMT_plx "\n", __func__,
2332 subpage, len, addr);
2334 res = flatview_read(subpage->fv, addr + subpage->base, attrs, buf, len);
2338 *data = ldn_p(buf, len);
2342 static MemTxResult subpage_write(void *opaque, hwaddr addr,
2343 uint64_t value, unsigned len, MemTxAttrs attrs)
2345 subpage_t *subpage = opaque;
2348 #if defined(DEBUG_SUBPAGE)
2349 printf("%s: subpage %p len %u addr " HWADDR_FMT_plx
2350 " value %"PRIx64"\n",
2351 __func__, subpage, len, addr, value);
2353 stn_p(buf, len, value);
2354 return flatview_write(subpage->fv, addr + subpage->base, attrs, buf, len);
2357 static bool subpage_accepts(void *opaque, hwaddr addr,
2358 unsigned len, bool is_write,
2361 subpage_t *subpage = opaque;
2362 #if defined(DEBUG_SUBPAGE)
2363 printf("%s: subpage %p %c len %u addr " HWADDR_FMT_plx "\n",
2364 __func__, subpage, is_write ? 'w' : 'r', len, addr);
2367 return flatview_access_valid(subpage->fv, addr + subpage->base,
2368 len, is_write, attrs);
2371 static const MemoryRegionOps subpage_ops = {
2372 .read_with_attrs = subpage_read,
2373 .write_with_attrs = subpage_write,
2374 .impl.min_access_size = 1,
2375 .impl.max_access_size = 8,
2376 .valid.min_access_size = 1,
2377 .valid.max_access_size = 8,
2378 .valid.accepts = subpage_accepts,
2379 .endianness = DEVICE_NATIVE_ENDIAN,
2382 static int subpage_register(subpage_t *mmio, uint32_t start, uint32_t end,
2387 if (start >= TARGET_PAGE_SIZE || end >= TARGET_PAGE_SIZE)
2389 idx = SUBPAGE_IDX(start);
2390 eidx = SUBPAGE_IDX(end);
2391 #if defined(DEBUG_SUBPAGE)
2392 printf("%s: %p start %08x end %08x idx %08x eidx %08x section %d\n",
2393 __func__, mmio, start, end, idx, eidx, section);
2395 for (; idx <= eidx; idx++) {
2396 mmio->sub_section[idx] = section;
2402 static subpage_t *subpage_init(FlatView *fv, hwaddr base)
2406 /* mmio->sub_section is set to PHYS_SECTION_UNASSIGNED with g_malloc0 */
2407 mmio = g_malloc0(sizeof(subpage_t) + TARGET_PAGE_SIZE * sizeof(uint16_t));
2410 memory_region_init_io(&mmio->iomem, NULL, &subpage_ops, mmio,
2411 NULL, TARGET_PAGE_SIZE);
2412 mmio->iomem.subpage = true;
2413 #if defined(DEBUG_SUBPAGE)
2414 printf("%s: %p base " HWADDR_FMT_plx " len %08x\n", __func__,
2415 mmio, base, TARGET_PAGE_SIZE);
2421 static uint16_t dummy_section(PhysPageMap *map, FlatView *fv, MemoryRegion *mr)
2424 MemoryRegionSection section = {
2427 .offset_within_address_space = 0,
2428 .offset_within_region = 0,
2429 .size = int128_2_64(),
2432 return phys_section_add(map, §ion);
2435 MemoryRegionSection *iotlb_to_section(CPUState *cpu,
2436 hwaddr index, MemTxAttrs attrs)
2438 int asidx = cpu_asidx_from_attrs(cpu, attrs);
2439 CPUAddressSpace *cpuas = &cpu->cpu_ases[asidx];
2440 AddressSpaceDispatch *d = cpuas->memory_dispatch;
2441 int section_index = index & ~TARGET_PAGE_MASK;
2442 MemoryRegionSection *ret;
2444 assert(section_index < d->map.sections_nb);
2445 ret = d->map.sections + section_index;
2447 assert(ret->mr->ops);
2452 static void io_mem_init(void)
2454 memory_region_init_io(&io_mem_unassigned, NULL, &unassigned_mem_ops, NULL,
2458 AddressSpaceDispatch *address_space_dispatch_new(FlatView *fv)
2460 AddressSpaceDispatch *d = g_new0(AddressSpaceDispatch, 1);
2463 n = dummy_section(&d->map, fv, &io_mem_unassigned);
2464 assert(n == PHYS_SECTION_UNASSIGNED);
2466 d->phys_map = (PhysPageEntry) { .ptr = PHYS_MAP_NODE_NIL, .skip = 1 };
2471 void address_space_dispatch_free(AddressSpaceDispatch *d)
2473 phys_sections_free(&d->map);
2477 static void do_nothing(CPUState *cpu, run_on_cpu_data d)
2481 static void tcg_log_global_after_sync(MemoryListener *listener)
2483 CPUAddressSpace *cpuas;
2485 /* Wait for the CPU to end the current TB. This avoids the following
2489 * ---------------------- -------------------------
2490 * TLB check -> slow path
2491 * notdirty_mem_write
2495 * TLB check -> fast path
2499 * by pushing the migration thread's memory read after the vCPU thread has
2500 * written the memory.
2502 if (replay_mode == REPLAY_MODE_NONE) {
2504 * VGA can make calls to this function while updating the screen.
2505 * In record/replay mode this causes a deadlock, because
2506 * run_on_cpu waits for rr mutex. Therefore no races are possible
2507 * in this case and no need for making run_on_cpu when
2508 * record/replay is enabled.
2510 cpuas = container_of(listener, CPUAddressSpace, tcg_as_listener);
2511 run_on_cpu(cpuas->cpu, do_nothing, RUN_ON_CPU_NULL);
2515 static void tcg_commit_cpu(CPUState *cpu, run_on_cpu_data data)
2517 CPUAddressSpace *cpuas = data.host_ptr;
2519 cpuas->memory_dispatch = address_space_to_dispatch(cpuas->as);
2523 static void tcg_commit(MemoryListener *listener)
2525 CPUAddressSpace *cpuas;
2528 assert(tcg_enabled());
2529 /* since each CPU stores ram addresses in its TLB cache, we must
2530 reset the modified entries */
2531 cpuas = container_of(listener, CPUAddressSpace, tcg_as_listener);
2535 * Defer changes to as->memory_dispatch until the cpu is quiescent.
2536 * Otherwise we race between (1) other cpu threads and (2) ongoing
2537 * i/o for the current cpu thread, with data cached by mmu_lookup().
2539 * In addition, queueing the work function will kick the cpu back to
2540 * the main loop, which will end the RCU critical section and reclaim
2541 * the memory data structures.
2543 * That said, the listener is also called during realize, before
2544 * all of the tcg machinery for run-on is initialized: thus halt_cond.
2546 if (cpu->halt_cond) {
2547 async_run_on_cpu(cpu, tcg_commit_cpu, RUN_ON_CPU_HOST_PTR(cpuas));
2549 tcg_commit_cpu(cpu, RUN_ON_CPU_HOST_PTR(cpuas));
2553 static void memory_map_init(void)
2555 system_memory = g_malloc(sizeof(*system_memory));
2557 memory_region_init(system_memory, NULL, "system", UINT64_MAX);
2558 address_space_init(&address_space_memory, system_memory, "memory");
2560 system_io = g_malloc(sizeof(*system_io));
2561 memory_region_init_io(system_io, NULL, &unassigned_io_ops, NULL, "io",
2563 address_space_init(&address_space_io, system_io, "I/O");
2566 MemoryRegion *get_system_memory(void)
2568 return system_memory;
2571 MemoryRegion *get_system_io(void)
2576 static void invalidate_and_set_dirty(MemoryRegion *mr, hwaddr addr,
2579 uint8_t dirty_log_mask = memory_region_get_dirty_log_mask(mr);
2580 addr += memory_region_get_ram_addr(mr);
2582 /* No early return if dirty_log_mask is or becomes 0, because
2583 * cpu_physical_memory_set_dirty_range will still call
2584 * xen_modified_memory.
2586 if (dirty_log_mask) {
2588 cpu_physical_memory_range_includes_clean(addr, length, dirty_log_mask);
2590 if (dirty_log_mask & (1 << DIRTY_MEMORY_CODE)) {
2591 assert(tcg_enabled());
2592 tb_invalidate_phys_range(addr, addr + length - 1);
2593 dirty_log_mask &= ~(1 << DIRTY_MEMORY_CODE);
2595 cpu_physical_memory_set_dirty_range(addr, length, dirty_log_mask);
2598 void memory_region_flush_rom_device(MemoryRegion *mr, hwaddr addr, hwaddr size)
2601 * In principle this function would work on other memory region types too,
2602 * but the ROM device use case is the only one where this operation is
2603 * necessary. Other memory regions should use the
2604 * address_space_read/write() APIs.
2606 assert(memory_region_is_romd(mr));
2608 invalidate_and_set_dirty(mr, addr, size);
2611 int memory_access_size(MemoryRegion *mr, unsigned l, hwaddr addr)
2613 unsigned access_size_max = mr->ops->valid.max_access_size;
2615 /* Regions are assumed to support 1-4 byte accesses unless
2616 otherwise specified. */
2617 if (access_size_max == 0) {
2618 access_size_max = 4;
2621 /* Bound the maximum access by the alignment of the address. */
2622 if (!mr->ops->impl.unaligned) {
2623 unsigned align_size_max = addr & -addr;
2624 if (align_size_max != 0 && align_size_max < access_size_max) {
2625 access_size_max = align_size_max;
2629 /* Don't attempt accesses larger than the maximum. */
2630 if (l > access_size_max) {
2631 l = access_size_max;
2638 bool prepare_mmio_access(MemoryRegion *mr)
2640 bool release_lock = false;
2642 if (!qemu_mutex_iothread_locked()) {
2643 qemu_mutex_lock_iothread();
2644 release_lock = true;
2646 if (mr->flush_coalesced_mmio) {
2647 qemu_flush_coalesced_mmio_buffer();
2650 return release_lock;
2654 * flatview_access_allowed
2655 * @mr: #MemoryRegion to be accessed
2656 * @attrs: memory transaction attributes
2657 * @addr: address within that memory region
2658 * @len: the number of bytes to access
2660 * Check if a memory transaction is allowed.
2662 * Returns: true if transaction is allowed, false if denied.
2664 static bool flatview_access_allowed(MemoryRegion *mr, MemTxAttrs attrs,
2665 hwaddr addr, hwaddr len)
2667 if (likely(!attrs.memory)) {
2670 if (memory_region_is_ram(mr)) {
2673 qemu_log_mask(LOG_GUEST_ERROR,
2674 "Invalid access to non-RAM device at "
2675 "addr 0x%" HWADDR_PRIX ", size %" HWADDR_PRIu ", "
2676 "region '%s'\n", addr, len, memory_region_name(mr));
2680 /* Called within RCU critical section. */
2681 static MemTxResult flatview_write_continue(FlatView *fv, hwaddr addr,
2684 hwaddr len, hwaddr addr1,
2685 hwaddr l, MemoryRegion *mr)
2689 MemTxResult result = MEMTX_OK;
2690 bool release_lock = false;
2691 const uint8_t *buf = ptr;
2694 if (!flatview_access_allowed(mr, attrs, addr1, l)) {
2695 result |= MEMTX_ACCESS_ERROR;
2697 } else if (!memory_access_is_direct(mr, true)) {
2698 release_lock |= prepare_mmio_access(mr);
2699 l = memory_access_size(mr, l, addr1);
2700 /* XXX: could force current_cpu to NULL to avoid
2702 val = ldn_he_p(buf, l);
2703 result |= memory_region_dispatch_write(mr, addr1, val,
2704 size_memop(l), attrs);
2707 ram_ptr = qemu_ram_ptr_length(mr->ram_block, addr1, &l, false);
2708 memmove(ram_ptr, buf, l);
2709 invalidate_and_set_dirty(mr, addr1, l);
2713 qemu_mutex_unlock_iothread();
2714 release_lock = false;
2726 mr = flatview_translate(fv, addr, &addr1, &l, true, attrs);
2732 /* Called from RCU critical section. */
2733 static MemTxResult flatview_write(FlatView *fv, hwaddr addr, MemTxAttrs attrs,
2734 const void *buf, hwaddr len)
2741 mr = flatview_translate(fv, addr, &addr1, &l, true, attrs);
2742 if (!flatview_access_allowed(mr, attrs, addr, len)) {
2743 return MEMTX_ACCESS_ERROR;
2745 return flatview_write_continue(fv, addr, attrs, buf, len,
2749 /* Called within RCU critical section. */
2750 MemTxResult flatview_read_continue(FlatView *fv, hwaddr addr,
2751 MemTxAttrs attrs, void *ptr,
2752 hwaddr len, hwaddr addr1, hwaddr l,
2757 MemTxResult result = MEMTX_OK;
2758 bool release_lock = false;
2761 fuzz_dma_read_cb(addr, len, mr);
2763 if (!flatview_access_allowed(mr, attrs, addr1, l)) {
2764 result |= MEMTX_ACCESS_ERROR;
2766 } else if (!memory_access_is_direct(mr, false)) {
2768 release_lock |= prepare_mmio_access(mr);
2769 l = memory_access_size(mr, l, addr1);
2770 result |= memory_region_dispatch_read(mr, addr1, &val,
2771 size_memop(l), attrs);
2772 stn_he_p(buf, l, val);
2775 ram_ptr = qemu_ram_ptr_length(mr->ram_block, addr1, &l, false);
2776 memcpy(buf, ram_ptr, l);
2780 qemu_mutex_unlock_iothread();
2781 release_lock = false;
2793 mr = flatview_translate(fv, addr, &addr1, &l, false, attrs);
2799 /* Called from RCU critical section. */
2800 static MemTxResult flatview_read(FlatView *fv, hwaddr addr,
2801 MemTxAttrs attrs, void *buf, hwaddr len)
2808 mr = flatview_translate(fv, addr, &addr1, &l, false, attrs);
2809 if (!flatview_access_allowed(mr, attrs, addr, len)) {
2810 return MEMTX_ACCESS_ERROR;
2812 return flatview_read_continue(fv, addr, attrs, buf, len,
2816 MemTxResult address_space_read_full(AddressSpace *as, hwaddr addr,
2817 MemTxAttrs attrs, void *buf, hwaddr len)
2819 MemTxResult result = MEMTX_OK;
2823 RCU_READ_LOCK_GUARD();
2824 fv = address_space_to_flatview(as);
2825 result = flatview_read(fv, addr, attrs, buf, len);
2831 MemTxResult address_space_write(AddressSpace *as, hwaddr addr,
2833 const void *buf, hwaddr len)
2835 MemTxResult result = MEMTX_OK;
2839 RCU_READ_LOCK_GUARD();
2840 fv = address_space_to_flatview(as);
2841 result = flatview_write(fv, addr, attrs, buf, len);
2847 MemTxResult address_space_rw(AddressSpace *as, hwaddr addr, MemTxAttrs attrs,
2848 void *buf, hwaddr len, bool is_write)
2851 return address_space_write(as, addr, attrs, buf, len);
2853 return address_space_read_full(as, addr, attrs, buf, len);
2857 MemTxResult address_space_set(AddressSpace *as, hwaddr addr,
2858 uint8_t c, hwaddr len, MemTxAttrs attrs)
2860 #define FILLBUF_SIZE 512
2861 uint8_t fillbuf[FILLBUF_SIZE];
2863 MemTxResult error = MEMTX_OK;
2865 memset(fillbuf, c, FILLBUF_SIZE);
2867 l = len < FILLBUF_SIZE ? len : FILLBUF_SIZE;
2868 error |= address_space_write(as, addr, attrs, fillbuf, l);
2876 void cpu_physical_memory_rw(hwaddr addr, void *buf,
2877 hwaddr len, bool is_write)
2879 address_space_rw(&address_space_memory, addr, MEMTXATTRS_UNSPECIFIED,
2880 buf, len, is_write);
2883 enum write_rom_type {
2888 static inline MemTxResult address_space_write_rom_internal(AddressSpace *as,
2893 enum write_rom_type type)
2899 const uint8_t *buf = ptr;
2901 RCU_READ_LOCK_GUARD();
2904 mr = address_space_translate(as, addr, &addr1, &l, true, attrs);
2906 if (!(memory_region_is_ram(mr) ||
2907 memory_region_is_romd(mr))) {
2908 l = memory_access_size(mr, l, addr1);
2911 ram_ptr = qemu_map_ram_ptr(mr->ram_block, addr1);
2914 memcpy(ram_ptr, buf, l);
2915 invalidate_and_set_dirty(mr, addr1, l);
2918 flush_idcache_range((uintptr_t)ram_ptr, (uintptr_t)ram_ptr, l);
2929 /* used for ROM loading : can write in RAM and ROM */
2930 MemTxResult address_space_write_rom(AddressSpace *as, hwaddr addr,
2932 const void *buf, hwaddr len)
2934 return address_space_write_rom_internal(as, addr, attrs,
2935 buf, len, WRITE_DATA);
2938 void cpu_flush_icache_range(hwaddr start, hwaddr len)
2941 * This function should do the same thing as an icache flush that was
2942 * triggered from within the guest. For TCG we are always cache coherent,
2943 * so there is no need to flush anything. For KVM / Xen we need to flush
2944 * the host's instruction cache at least.
2946 if (tcg_enabled()) {
2950 address_space_write_rom_internal(&address_space_memory,
2951 start, MEMTXATTRS_UNSPECIFIED,
2952 NULL, len, FLUSH_CACHE);
2963 static BounceBuffer bounce;
2965 typedef struct MapClient {
2967 QLIST_ENTRY(MapClient) link;
2970 QemuMutex map_client_list_lock;
2971 static QLIST_HEAD(, MapClient) map_client_list
2972 = QLIST_HEAD_INITIALIZER(map_client_list);
2974 static void cpu_unregister_map_client_do(MapClient *client)
2976 QLIST_REMOVE(client, link);
2980 static void cpu_notify_map_clients_locked(void)
2984 while (!QLIST_EMPTY(&map_client_list)) {
2985 client = QLIST_FIRST(&map_client_list);
2986 qemu_bh_schedule(client->bh);
2987 cpu_unregister_map_client_do(client);
2991 void cpu_register_map_client(QEMUBH *bh)
2993 MapClient *client = g_malloc(sizeof(*client));
2995 qemu_mutex_lock(&map_client_list_lock);
2997 QLIST_INSERT_HEAD(&map_client_list, client, link);
2998 /* Write map_client_list before reading in_use. */
3000 if (!qatomic_read(&bounce.in_use)) {
3001 cpu_notify_map_clients_locked();
3003 qemu_mutex_unlock(&map_client_list_lock);
3006 void cpu_exec_init_all(void)
3008 qemu_mutex_init(&ram_list.mutex);
3009 /* The data structures we set up here depend on knowing the page size,
3010 * so no more changes can be made after this point.
3011 * In an ideal world, nothing we did before we had finished the
3012 * machine setup would care about the target page size, and we could
3013 * do this much later, rather than requiring board models to state
3014 * up front what their requirements are.
3016 finalize_target_page_bits();
3019 qemu_mutex_init(&map_client_list_lock);
3022 void cpu_unregister_map_client(QEMUBH *bh)
3026 qemu_mutex_lock(&map_client_list_lock);
3027 QLIST_FOREACH(client, &map_client_list, link) {
3028 if (client->bh == bh) {
3029 cpu_unregister_map_client_do(client);
3033 qemu_mutex_unlock(&map_client_list_lock);
3036 static void cpu_notify_map_clients(void)
3038 qemu_mutex_lock(&map_client_list_lock);
3039 cpu_notify_map_clients_locked();
3040 qemu_mutex_unlock(&map_client_list_lock);
3043 static bool flatview_access_valid(FlatView *fv, hwaddr addr, hwaddr len,
3044 bool is_write, MemTxAttrs attrs)
3051 mr = flatview_translate(fv, addr, &xlat, &l, is_write, attrs);
3052 if (!memory_access_is_direct(mr, is_write)) {
3053 l = memory_access_size(mr, l, addr);
3054 if (!memory_region_access_valid(mr, xlat, l, is_write, attrs)) {
3065 bool address_space_access_valid(AddressSpace *as, hwaddr addr,
3066 hwaddr len, bool is_write,
3071 RCU_READ_LOCK_GUARD();
3072 fv = address_space_to_flatview(as);
3073 return flatview_access_valid(fv, addr, len, is_write, attrs);
3077 flatview_extend_translation(FlatView *fv, hwaddr addr,
3079 MemoryRegion *mr, hwaddr base, hwaddr len,
3080 bool is_write, MemTxAttrs attrs)
3084 MemoryRegion *this_mr;
3090 if (target_len == 0) {
3095 this_mr = flatview_translate(fv, addr, &xlat,
3096 &len, is_write, attrs);
3097 if (this_mr != mr || xlat != base + done) {
3103 /* Map a physical memory region into a host virtual address.
3104 * May map a subset of the requested range, given by and returned in *plen.
3105 * May return NULL if resources needed to perform the mapping are exhausted.
3106 * Use only for reads OR writes - not for read-modify-write operations.
3107 * Use cpu_register_map_client() to know when retrying the map operation is
3108 * likely to succeed.
3110 void *address_space_map(AddressSpace *as,
3126 RCU_READ_LOCK_GUARD();
3127 fv = address_space_to_flatview(as);
3128 mr = flatview_translate(fv, addr, &xlat, &l, is_write, attrs);
3130 if (!memory_access_is_direct(mr, is_write)) {
3131 if (qatomic_xchg(&bounce.in_use, true)) {
3135 /* Avoid unbounded allocations */
3136 l = MIN(l, TARGET_PAGE_SIZE);
3137 bounce.buffer = qemu_memalign(TARGET_PAGE_SIZE, l);
3141 memory_region_ref(mr);
3144 flatview_read(fv, addr, MEMTXATTRS_UNSPECIFIED,
3149 return bounce.buffer;
3153 memory_region_ref(mr);
3154 *plen = flatview_extend_translation(fv, addr, len, mr, xlat,
3155 l, is_write, attrs);
3156 fuzz_dma_read_cb(addr, *plen, mr);
3157 return qemu_ram_ptr_length(mr->ram_block, xlat, plen, true);
3160 /* Unmaps a memory region previously mapped by address_space_map().
3161 * Will also mark the memory as dirty if is_write is true. access_len gives
3162 * the amount of memory that was actually read or written by the caller.
3164 void address_space_unmap(AddressSpace *as, void *buffer, hwaddr len,
3165 bool is_write, hwaddr access_len)
3167 if (buffer != bounce.buffer) {
3171 mr = memory_region_from_host(buffer, &addr1);
3174 invalidate_and_set_dirty(mr, addr1, access_len);
3176 if (xen_enabled()) {
3177 xen_invalidate_map_cache_entry(buffer);
3179 memory_region_unref(mr);
3183 address_space_write(as, bounce.addr, MEMTXATTRS_UNSPECIFIED,
3184 bounce.buffer, access_len);
3186 qemu_vfree(bounce.buffer);
3187 bounce.buffer = NULL;
3188 memory_region_unref(bounce.mr);
3189 /* Clear in_use before reading map_client_list. */
3190 qatomic_set_mb(&bounce.in_use, false);
3191 cpu_notify_map_clients();
3194 void *cpu_physical_memory_map(hwaddr addr,
3198 return address_space_map(&address_space_memory, addr, plen, is_write,
3199 MEMTXATTRS_UNSPECIFIED);
3202 void cpu_physical_memory_unmap(void *buffer, hwaddr len,
3203 bool is_write, hwaddr access_len)
3205 return address_space_unmap(&address_space_memory, buffer, len, is_write, access_len);
3208 #define ARG1_DECL AddressSpace *as
3211 #define TRANSLATE(...) address_space_translate(as, __VA_ARGS__)
3212 #define RCU_READ_LOCK(...) rcu_read_lock()
3213 #define RCU_READ_UNLOCK(...) rcu_read_unlock()
3214 #include "memory_ldst.c.inc"
3216 int64_t address_space_cache_init(MemoryRegionCache *cache,
3222 AddressSpaceDispatch *d;
3230 cache->fv = address_space_get_flatview(as);
3231 d = flatview_to_dispatch(cache->fv);
3232 cache->mrs = *address_space_translate_internal(d, addr, &cache->xlat, &l, true);
3235 * cache->xlat is now relative to cache->mrs.mr, not to the section itself.
3236 * Take that into account to compute how many bytes are there between
3237 * cache->xlat and the end of the section.
3239 diff = int128_sub(cache->mrs.size,
3240 int128_make64(cache->xlat - cache->mrs.offset_within_region));
3241 l = int128_get64(int128_min(diff, int128_make64(l)));
3244 memory_region_ref(mr);
3245 if (memory_access_is_direct(mr, is_write)) {
3246 /* We don't care about the memory attributes here as we're only
3247 * doing this if we found actual RAM, which behaves the same
3248 * regardless of attributes; so UNSPECIFIED is fine.
3250 l = flatview_extend_translation(cache->fv, addr, len, mr,
3251 cache->xlat, l, is_write,
3252 MEMTXATTRS_UNSPECIFIED);
3253 cache->ptr = qemu_ram_ptr_length(mr->ram_block, cache->xlat, &l, true);
3259 cache->is_write = is_write;
3263 void address_space_cache_invalidate(MemoryRegionCache *cache,
3267 assert(cache->is_write);
3268 if (likely(cache->ptr)) {
3269 invalidate_and_set_dirty(cache->mrs.mr, addr + cache->xlat, access_len);
3273 void address_space_cache_destroy(MemoryRegionCache *cache)
3275 if (!cache->mrs.mr) {
3279 if (xen_enabled()) {
3280 xen_invalidate_map_cache_entry(cache->ptr);
3282 memory_region_unref(cache->mrs.mr);
3283 flatview_unref(cache->fv);
3284 cache->mrs.mr = NULL;
3288 /* Called from RCU critical section. This function has the same
3289 * semantics as address_space_translate, but it only works on a
3290 * predefined range of a MemoryRegion that was mapped with
3291 * address_space_cache_init.
3293 static inline MemoryRegion *address_space_translate_cached(
3294 MemoryRegionCache *cache, hwaddr addr, hwaddr *xlat,
3295 hwaddr *plen, bool is_write, MemTxAttrs attrs)
3297 MemoryRegionSection section;
3299 IOMMUMemoryRegion *iommu_mr;
3300 AddressSpace *target_as;
3302 assert(!cache->ptr);
3303 *xlat = addr + cache->xlat;
3306 iommu_mr = memory_region_get_iommu(mr);
3312 section = address_space_translate_iommu(iommu_mr, xlat, plen,
3313 NULL, is_write, true,
3318 /* Called from RCU critical section. address_space_read_cached uses this
3319 * out of line function when the target is an MMIO or IOMMU region.
3322 address_space_read_cached_slow(MemoryRegionCache *cache, hwaddr addr,
3323 void *buf, hwaddr len)
3329 mr = address_space_translate_cached(cache, addr, &addr1, &l, false,
3330 MEMTXATTRS_UNSPECIFIED);
3331 return flatview_read_continue(cache->fv,
3332 addr, MEMTXATTRS_UNSPECIFIED, buf, len,
3336 /* Called from RCU critical section. address_space_write_cached uses this
3337 * out of line function when the target is an MMIO or IOMMU region.
3340 address_space_write_cached_slow(MemoryRegionCache *cache, hwaddr addr,
3341 const void *buf, hwaddr len)
3347 mr = address_space_translate_cached(cache, addr, &addr1, &l, true,
3348 MEMTXATTRS_UNSPECIFIED);
3349 return flatview_write_continue(cache->fv,
3350 addr, MEMTXATTRS_UNSPECIFIED, buf, len,
3354 #define ARG1_DECL MemoryRegionCache *cache
3356 #define SUFFIX _cached_slow
3357 #define TRANSLATE(...) address_space_translate_cached(cache, __VA_ARGS__)
3358 #define RCU_READ_LOCK() ((void)0)
3359 #define RCU_READ_UNLOCK() ((void)0)
3360 #include "memory_ldst.c.inc"
3362 /* virtual memory access for debug (includes writing to ROM) */
3363 int cpu_memory_rw_debug(CPUState *cpu, vaddr addr,
3364 void *ptr, size_t len, bool is_write)
3370 cpu_synchronize_state(cpu);
3376 page = addr & TARGET_PAGE_MASK;
3377 phys_addr = cpu_get_phys_page_attrs_debug(cpu, page, &attrs);
3378 asidx = cpu_asidx_from_attrs(cpu, attrs);
3379 /* if no physical page mapped, return an error */
3380 if (phys_addr == -1)
3382 l = (page + TARGET_PAGE_SIZE) - addr;
3385 phys_addr += (addr & ~TARGET_PAGE_MASK);
3387 res = address_space_write_rom(cpu->cpu_ases[asidx].as, phys_addr,
3390 res = address_space_read(cpu->cpu_ases[asidx].as, phys_addr,
3393 if (res != MEMTX_OK) {
3404 * Allows code that needs to deal with migration bitmaps etc to still be built
3405 * target independent.
3407 size_t qemu_target_page_size(void)
3409 return TARGET_PAGE_SIZE;
3412 int qemu_target_page_mask(void)
3414 return TARGET_PAGE_MASK;
3417 int qemu_target_page_bits(void)
3419 return TARGET_PAGE_BITS;
3422 int qemu_target_page_bits_min(void)
3424 return TARGET_PAGE_BITS_MIN;
3427 /* Convert target pages to MiB (2**20). */
3428 size_t qemu_target_pages_to_MiB(size_t pages)
3430 int page_bits = TARGET_PAGE_BITS;
3432 /* So far, the largest (non-huge) page size is 64k, i.e. 16 bits. */
3433 g_assert(page_bits < 20);
3435 return pages >> (20 - page_bits);
3438 bool cpu_physical_memory_is_io(hwaddr phys_addr)
3443 RCU_READ_LOCK_GUARD();
3444 mr = address_space_translate(&address_space_memory,
3445 phys_addr, &phys_addr, &l, false,
3446 MEMTXATTRS_UNSPECIFIED);
3448 return !(memory_region_is_ram(mr) || memory_region_is_romd(mr));
3451 int qemu_ram_foreach_block(RAMBlockIterFunc func, void *opaque)
3456 RCU_READ_LOCK_GUARD();
3457 RAMBLOCK_FOREACH(block) {
3458 ret = func(block, opaque);
3467 * Unmap pages of memory from start to start+length such that
3468 * they a) read as 0, b) Trigger whatever fault mechanism
3469 * the OS provides for postcopy.
3470 * The pages must be unmapped by the end of the function.
3471 * Returns: 0 on success, none-0 on failure
3474 int ram_block_discard_range(RAMBlock *rb, uint64_t start, size_t length)
3478 uint8_t *host_startaddr = rb->host + start;
3480 if (!QEMU_PTR_IS_ALIGNED(host_startaddr, rb->page_size)) {
3481 error_report("ram_block_discard_range: Unaligned start address: %p",
3486 if ((start + length) <= rb->max_length) {
3487 bool need_madvise, need_fallocate;
3488 if (!QEMU_IS_ALIGNED(length, rb->page_size)) {
3489 error_report("ram_block_discard_range: Unaligned length: %zx",
3494 errno = ENOTSUP; /* If we are missing MADVISE etc */
3496 /* The logic here is messy;
3497 * madvise DONTNEED fails for hugepages
3498 * fallocate works on hugepages and shmem
3499 * shared anonymous memory requires madvise REMOVE
3501 need_madvise = (rb->page_size == qemu_host_page_size);
3502 need_fallocate = rb->fd != -1;
3503 if (need_fallocate) {
3504 /* For a file, this causes the area of the file to be zero'd
3505 * if read, and for hugetlbfs also causes it to be unmapped
3506 * so a userfault will trigger.
3508 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
3510 * fallocate() will fail with readonly files. Let's print a
3511 * proper error message.
3513 if (rb->flags & RAM_READONLY_FD) {
3514 error_report("ram_block_discard_range: Discarding RAM"
3515 " with readonly files is not supported");
3520 * We'll discard data from the actual file, even though we only
3521 * have a MAP_PRIVATE mapping, possibly messing with other
3522 * MAP_PRIVATE/MAP_SHARED mappings. There is no easy way to
3523 * change that behavior whithout violating the promised
3524 * semantics of ram_block_discard_range().
3526 * Only warn, because it works as long as nobody else uses that
3529 if (!qemu_ram_is_shared(rb)) {
3530 warn_report_once("ram_block_discard_range: Discarding RAM"
3531 " in private file mappings is possibly"
3532 " dangerous, because it will modify the"
3533 " underlying file and will affect other"
3534 " users of the file");
3537 ret = fallocate(rb->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
3541 error_report("ram_block_discard_range: Failed to fallocate "
3542 "%s:%" PRIx64 " +%zx (%d)",
3543 rb->idstr, start, length, ret);
3548 error_report("ram_block_discard_range: fallocate not available/file"
3549 "%s:%" PRIx64 " +%zx (%d)",
3550 rb->idstr, start, length, ret);
3555 /* For normal RAM this causes it to be unmapped,
3556 * for shared memory it causes the local mapping to disappear
3557 * and to fall back on the file contents (which we just
3558 * fallocate'd away).
3560 #if defined(CONFIG_MADVISE)
3561 if (qemu_ram_is_shared(rb) && rb->fd < 0) {
3562 ret = madvise(host_startaddr, length, QEMU_MADV_REMOVE);
3564 ret = madvise(host_startaddr, length, QEMU_MADV_DONTNEED);
3568 error_report("ram_block_discard_range: Failed to discard range "
3569 "%s:%" PRIx64 " +%zx (%d)",
3570 rb->idstr, start, length, ret);
3575 error_report("ram_block_discard_range: MADVISE not available"
3576 "%s:%" PRIx64 " +%zx (%d)",
3577 rb->idstr, start, length, ret);
3581 trace_ram_block_discard_range(rb->idstr, host_startaddr, length,
3582 need_madvise, need_fallocate, ret);
3584 error_report("ram_block_discard_range: Overrun block '%s' (%" PRIu64
3585 "/%zx/" RAM_ADDR_FMT")",
3586 rb->idstr, start, length, rb->max_length);
3593 bool ramblock_is_pmem(RAMBlock *rb)
3595 return rb->flags & RAM_PMEM;
3598 static void mtree_print_phys_entries(int start, int end, int skip, int ptr)
3600 if (start == end - 1) {
3601 qemu_printf("\t%3d ", start);
3603 qemu_printf("\t%3d..%-3d ", start, end - 1);
3605 qemu_printf(" skip=%d ", skip);
3606 if (ptr == PHYS_MAP_NODE_NIL) {
3607 qemu_printf(" ptr=NIL");
3609 qemu_printf(" ptr=#%d", ptr);
3611 qemu_printf(" ptr=[%d]", ptr);
3616 #define MR_SIZE(size) (int128_nz(size) ? (hwaddr)int128_get64( \
3617 int128_sub((size), int128_one())) : 0)
3619 void mtree_print_dispatch(AddressSpaceDispatch *d, MemoryRegion *root)
3623 qemu_printf(" Dispatch\n");
3624 qemu_printf(" Physical sections\n");
3626 for (i = 0; i < d->map.sections_nb; ++i) {
3627 MemoryRegionSection *s = d->map.sections + i;
3628 const char *names[] = { " [unassigned]", " [not dirty]",
3629 " [ROM]", " [watch]" };
3631 qemu_printf(" #%d @" HWADDR_FMT_plx ".." HWADDR_FMT_plx
3634 s->offset_within_address_space,
3635 s->offset_within_address_space + MR_SIZE(s->size),
3636 s->mr->name ? s->mr->name : "(noname)",
3637 i < ARRAY_SIZE(names) ? names[i] : "",
3638 s->mr == root ? " [ROOT]" : "",
3639 s == d->mru_section ? " [MRU]" : "",
3640 s->mr->is_iommu ? " [iommu]" : "");
3643 qemu_printf(" alias=%s", s->mr->alias->name ?
3644 s->mr->alias->name : "noname");
3649 qemu_printf(" Nodes (%d bits per level, %d levels) ptr=[%d] skip=%d\n",
3650 P_L2_BITS, P_L2_LEVELS, d->phys_map.ptr, d->phys_map.skip);
3651 for (i = 0; i < d->map.nodes_nb; ++i) {
3654 Node *n = d->map.nodes + i;
3656 qemu_printf(" [%d]\n", i);
3658 for (j = 0, jprev = 0, prev = *n[0]; j < ARRAY_SIZE(*n); ++j) {
3659 PhysPageEntry *pe = *n + j;
3661 if (pe->ptr == prev.ptr && pe->skip == prev.skip) {
3665 mtree_print_phys_entries(jprev, j, prev.skip, prev.ptr);
3671 if (jprev != ARRAY_SIZE(*n)) {
3672 mtree_print_phys_entries(jprev, j, prev.skip, prev.ptr);
3677 /* Require any discards to work. */
3678 static unsigned int ram_block_discard_required_cnt;
3679 /* Require only coordinated discards to work. */
3680 static unsigned int ram_block_coordinated_discard_required_cnt;
3681 /* Disable any discards. */
3682 static unsigned int ram_block_discard_disabled_cnt;
3683 /* Disable only uncoordinated discards. */
3684 static unsigned int ram_block_uncoordinated_discard_disabled_cnt;
3685 static QemuMutex ram_block_discard_disable_mutex;
3687 static void ram_block_discard_disable_mutex_lock(void)
3689 static gsize initialized;
3691 if (g_once_init_enter(&initialized)) {
3692 qemu_mutex_init(&ram_block_discard_disable_mutex);
3693 g_once_init_leave(&initialized, 1);
3695 qemu_mutex_lock(&ram_block_discard_disable_mutex);
3698 static void ram_block_discard_disable_mutex_unlock(void)
3700 qemu_mutex_unlock(&ram_block_discard_disable_mutex);
3703 int ram_block_discard_disable(bool state)
3707 ram_block_discard_disable_mutex_lock();
3709 ram_block_discard_disabled_cnt--;
3710 } else if (ram_block_discard_required_cnt ||
3711 ram_block_coordinated_discard_required_cnt) {
3714 ram_block_discard_disabled_cnt++;
3716 ram_block_discard_disable_mutex_unlock();
3720 int ram_block_uncoordinated_discard_disable(bool state)
3724 ram_block_discard_disable_mutex_lock();
3726 ram_block_uncoordinated_discard_disabled_cnt--;
3727 } else if (ram_block_discard_required_cnt) {
3730 ram_block_uncoordinated_discard_disabled_cnt++;
3732 ram_block_discard_disable_mutex_unlock();
3736 int ram_block_discard_require(bool state)
3740 ram_block_discard_disable_mutex_lock();
3742 ram_block_discard_required_cnt--;
3743 } else if (ram_block_discard_disabled_cnt ||
3744 ram_block_uncoordinated_discard_disabled_cnt) {
3747 ram_block_discard_required_cnt++;
3749 ram_block_discard_disable_mutex_unlock();
3753 int ram_block_coordinated_discard_require(bool state)
3757 ram_block_discard_disable_mutex_lock();
3759 ram_block_coordinated_discard_required_cnt--;
3760 } else if (ram_block_discard_disabled_cnt) {
3763 ram_block_coordinated_discard_required_cnt++;
3765 ram_block_discard_disable_mutex_unlock();
3769 bool ram_block_discard_is_disabled(void)
3771 return qatomic_read(&ram_block_discard_disabled_cnt) ||
3772 qatomic_read(&ram_block_uncoordinated_discard_disabled_cnt);
3775 bool ram_block_discard_is_required(void)
3777 return qatomic_read(&ram_block_discard_required_cnt) ||
3778 qatomic_read(&ram_block_coordinated_discard_required_cnt);