]> git.proxmox.com Git - mirror_qemu.git/blobdiff - exec.c
virtio: do not take address of packed members
[mirror_qemu.git] / exec.c
diff --git a/exec.c b/exec.c
index c09bd93df31e9f43d790f8a2be995e9946eaaeb6..4f5df07b6a26a8aea869c6baac820a0097c79d51 100644 (file)
--- a/exec.c
+++ b/exec.c
@@ -104,6 +104,9 @@ static MemoryRegion io_mem_unassigned;
  * (Set during postcopy)
  */
 #define RAM_UF_ZEROPAGE (1 << 3)
+
+/* RAM can be migrated */
+#define RAM_MIGRATABLE (1 << 4)
 #endif
 
 #ifdef TARGET_PAGE_BITS_VARY
@@ -461,6 +464,79 @@ address_space_translate_internal(AddressSpaceDispatch *d, hwaddr addr, hwaddr *x
     return section;
 }
 
+/**
+ * address_space_translate_iommu - translate an address through an IOMMU
+ * memory region and then through the target address space.
+ *
+ * @iommu_mr: the IOMMU memory region that we start the translation from
+ * @addr: the address to be translated through the MMU
+ * @xlat: the translated address offset within the destination memory region.
+ *        It cannot be %NULL.
+ * @plen_out: valid read/write length of the translated address. It
+ *            cannot be %NULL.
+ * @page_mask_out: page mask for the translated address. This
+ *            should only be meaningful for IOMMU translated
+ *            addresses, since there may be huge pages that this bit
+ *            would tell. It can be %NULL if we don't care about it.
+ * @is_write: whether the translation operation is for write
+ * @is_mmio: whether this can be MMIO, set true if it can
+ * @target_as: the address space targeted by the IOMMU
+ * @attrs: transaction attributes
+ *
+ * This function is called from RCU critical section.  It is the common
+ * part of flatview_do_translate and address_space_translate_cached.
+ */
+static MemoryRegionSection address_space_translate_iommu(IOMMUMemoryRegion *iommu_mr,
+                                                         hwaddr *xlat,
+                                                         hwaddr *plen_out,
+                                                         hwaddr *page_mask_out,
+                                                         bool is_write,
+                                                         bool is_mmio,
+                                                         AddressSpace **target_as,
+                                                         MemTxAttrs attrs)
+{
+    MemoryRegionSection *section;
+    hwaddr page_mask = (hwaddr)-1;
+
+    do {
+        hwaddr addr = *xlat;
+        IOMMUMemoryRegionClass *imrc = memory_region_get_iommu_class_nocheck(iommu_mr);
+        int iommu_idx = 0;
+        IOMMUTLBEntry iotlb;
+
+        if (imrc->attrs_to_index) {
+            iommu_idx = imrc->attrs_to_index(iommu_mr, attrs);
+        }
+
+        iotlb = imrc->translate(iommu_mr, addr, is_write ?
+                                IOMMU_WO : IOMMU_RO, iommu_idx);
+
+        if (!(iotlb.perm & (1 << is_write))) {
+            goto unassigned;
+        }
+
+        addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
+                | (addr & iotlb.addr_mask));
+        page_mask &= iotlb.addr_mask;
+        *plen_out = MIN(*plen_out, (addr | iotlb.addr_mask) - addr + 1);
+        *target_as = iotlb.target_as;
+
+        section = address_space_translate_internal(
+                address_space_to_dispatch(iotlb.target_as), addr, xlat,
+                plen_out, is_mmio);
+
+        iommu_mr = memory_region_get_iommu(section->mr);
+    } while (unlikely(iommu_mr));
+
+    if (page_mask_out) {
+        *page_mask_out = page_mask;
+    }
+    return *section;
+
+unassigned:
+    return (MemoryRegionSection) { .mr = &io_mem_unassigned };
+}
+
 /**
  * flatview_do_translate - translate an address in FlatView
  *
@@ -476,6 +552,8 @@ address_space_translate_internal(AddressSpaceDispatch *d, hwaddr addr, hwaddr *x
  *            would tell. It can be @NULL if we don't care about it.
  * @is_write: whether the translation operation is for write
  * @is_mmio: whether this can be MMIO, set true if it can
+ * @target_as: the address space targeted by the IOMMU
+ * @attrs: memory transaction attributes
  *
  * This function is called from RCU critical section
  */
@@ -486,68 +564,39 @@ static MemoryRegionSection flatview_do_translate(FlatView *fv,
                                                  hwaddr *page_mask_out,
                                                  bool is_write,
                                                  bool is_mmio,
-                                                 AddressSpace **target_as)
+                                                 AddressSpace **target_as,
+                                                 MemTxAttrs attrs)
 {
-    IOMMUTLBEntry iotlb;
     MemoryRegionSection *section;
     IOMMUMemoryRegion *iommu_mr;
-    IOMMUMemoryRegionClass *imrc;
-    hwaddr page_mask = (hwaddr)(-1);
     hwaddr plen = (hwaddr)(-1);
 
-    if (plen_out) {
-        plen = *plen_out;
-    }
-
-    for (;;) {
-        section = address_space_translate_internal(
-                flatview_to_dispatch(fv), addr, &addr,
-                &plen, is_mmio);
-
-        iommu_mr = memory_region_get_iommu(section->mr);
-        if (!iommu_mr) {
-            break;
-        }
-        imrc = memory_region_get_iommu_class_nocheck(iommu_mr);
-
-        iotlb = imrc->translate(iommu_mr, addr, is_write ?
-                                IOMMU_WO : IOMMU_RO);
-        addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
-                | (addr & iotlb.addr_mask));
-        page_mask &= iotlb.addr_mask;
-        plen = MIN(plen, (addr | iotlb.addr_mask) - addr + 1);
-        if (!(iotlb.perm & (1 << is_write))) {
-            goto translate_fail;
-        }
-
-        fv = address_space_to_flatview(iotlb.target_as);
-        *target_as = iotlb.target_as;
+    if (!plen_out) {
+        plen_out = &plen;
     }
 
-    *xlat = addr;
+    section = address_space_translate_internal(
+            flatview_to_dispatch(fv), addr, xlat,
+            plen_out, is_mmio);
 
-    if (page_mask == (hwaddr)(-1)) {
-        /* Not behind an IOMMU, use default page size. */
-        page_mask = ~TARGET_PAGE_MASK;
+    iommu_mr = memory_region_get_iommu(section->mr);
+    if (unlikely(iommu_mr)) {
+        return address_space_translate_iommu(iommu_mr, xlat,
+                                             plen_out, page_mask_out,
+                                             is_write, is_mmio,
+                                             target_as, attrs);
     }
-
     if (page_mask_out) {
-        *page_mask_out = page_mask;
-    }
-
-    if (plen_out) {
-        *plen_out = plen;
+        /* Not behind an IOMMU, use default page size. */
+        *page_mask_out = ~TARGET_PAGE_MASK;
     }
 
     return *section;
-
-translate_fail:
-    return (MemoryRegionSection) { .mr = &io_mem_unassigned };
 }
 
 /* Called from RCU critical section */
 IOMMUTLBEntry address_space_get_iotlb_entry(AddressSpace *as, hwaddr addr,
-                                            bool is_write)
+                                            bool is_write, MemTxAttrs attrs)
 {
     MemoryRegionSection section;
     hwaddr xlat, page_mask;
@@ -557,7 +606,8 @@ IOMMUTLBEntry address_space_get_iotlb_entry(AddressSpace *as, hwaddr addr,
      * but page mask.
      */
     section = flatview_do_translate(address_space_to_flatview(as), addr, &xlat,
-                                    NULL, &page_mask, is_write, false, &as);
+                                    NULL, &page_mask, is_write, false, &as,
+                                    attrs);
 
     /* Illegal translation */
     if (section.mr == &io_mem_unassigned) {
@@ -583,7 +633,8 @@ iotlb_fail:
 
 /* Called from RCU critical section */
 MemoryRegion *flatview_translate(FlatView *fv, hwaddr addr, hwaddr *xlat,
-                                 hwaddr *plen, bool is_write)
+                                 hwaddr *plen, bool is_write,
+                                 MemTxAttrs attrs)
 {
     MemoryRegion *mr;
     MemoryRegionSection section;
@@ -591,7 +642,7 @@ MemoryRegion *flatview_translate(FlatView *fv, hwaddr addr, hwaddr *xlat,
 
     /* This can be MMIO, so setup MMIO bit. */
     section = flatview_do_translate(fv, addr, xlat, plen, NULL,
-                                    is_write, true, &as);
+                                    is_write, true, &as, attrs);
     mr = section.mr;
 
     if (xen_enabled() && memory_access_is_direct(mr, is_write)) {
@@ -602,18 +653,144 @@ MemoryRegion *flatview_translate(FlatView *fv, hwaddr addr, hwaddr *xlat,
     return mr;
 }
 
+typedef struct TCGIOMMUNotifier {
+    IOMMUNotifier n;
+    MemoryRegion *mr;
+    CPUState *cpu;
+    int iommu_idx;
+    bool active;
+} TCGIOMMUNotifier;
+
+static void tcg_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
+{
+    TCGIOMMUNotifier *notifier = container_of(n, TCGIOMMUNotifier, n);
+
+    if (!notifier->active) {
+        return;
+    }
+    tlb_flush(notifier->cpu);
+    notifier->active = false;
+    /* We leave the notifier struct on the list to avoid reallocating it later.
+     * Generally the number of IOMMUs a CPU deals with will be small.
+     * In any case we can't unregister the iommu notifier from a notify
+     * callback.
+     */
+}
+
+static void tcg_register_iommu_notifier(CPUState *cpu,
+                                        IOMMUMemoryRegion *iommu_mr,
+                                        int iommu_idx)
+{
+    /* Make sure this CPU has an IOMMU notifier registered for this
+     * IOMMU/IOMMU index combination, so that we can flush its TLB
+     * when the IOMMU tells us the mappings we've cached have changed.
+     */
+    MemoryRegion *mr = MEMORY_REGION(iommu_mr);
+    TCGIOMMUNotifier *notifier;
+    int i;
+
+    for (i = 0; i < cpu->iommu_notifiers->len; i++) {
+        notifier = &g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier, i);
+        if (notifier->mr == mr && notifier->iommu_idx == iommu_idx) {
+            break;
+        }
+    }
+    if (i == cpu->iommu_notifiers->len) {
+        /* Not found, add a new entry at the end of the array */
+        cpu->iommu_notifiers = g_array_set_size(cpu->iommu_notifiers, i + 1);
+        notifier = &g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier, i);
+
+        notifier->mr = mr;
+        notifier->iommu_idx = iommu_idx;
+        notifier->cpu = cpu;
+        /* Rather than trying to register interest in the specific part
+         * of the iommu's address space that we've accessed and then
+         * expand it later as subsequent accesses touch more of it, we
+         * just register interest in the whole thing, on the assumption
+         * that iommu reconfiguration will be rare.
+         */
+        iommu_notifier_init(&notifier->n,
+                            tcg_iommu_unmap_notify,
+                            IOMMU_NOTIFIER_UNMAP,
+                            0,
+                            HWADDR_MAX,
+                            iommu_idx);
+        memory_region_register_iommu_notifier(notifier->mr, &notifier->n);
+    }
+
+    if (!notifier->active) {
+        notifier->active = true;
+    }
+}
+
+static void tcg_iommu_free_notifier_list(CPUState *cpu)
+{
+    /* Destroy the CPU's notifier list */
+    int i;
+    TCGIOMMUNotifier *notifier;
+
+    for (i = 0; i < cpu->iommu_notifiers->len; i++) {
+        notifier = &g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier, i);
+        memory_region_unregister_iommu_notifier(notifier->mr, &notifier->n);
+    }
+    g_array_free(cpu->iommu_notifiers, true);
+}
+
 /* Called from RCU critical section */
 MemoryRegionSection *
 address_space_translate_for_iotlb(CPUState *cpu, int asidx, hwaddr addr,
-                                  hwaddr *xlat, hwaddr *plen)
+                                  hwaddr *xlat, hwaddr *plen,
+                                  MemTxAttrs attrs, int *prot)
 {
     MemoryRegionSection *section;
+    IOMMUMemoryRegion *iommu_mr;
+    IOMMUMemoryRegionClass *imrc;
+    IOMMUTLBEntry iotlb;
+    int iommu_idx;
     AddressSpaceDispatch *d = atomic_rcu_read(&cpu->cpu_ases[asidx].memory_dispatch);
 
-    section = address_space_translate_internal(d, addr, xlat, plen, false);
+    for (;;) {
+        section = address_space_translate_internal(d, addr, &addr, plen, false);
+
+        iommu_mr = memory_region_get_iommu(section->mr);
+        if (!iommu_mr) {
+            break;
+        }
+
+        imrc = memory_region_get_iommu_class_nocheck(iommu_mr);
+
+        iommu_idx = imrc->attrs_to_index(iommu_mr, attrs);
+        tcg_register_iommu_notifier(cpu, iommu_mr, iommu_idx);
+        /* We need all the permissions, so pass IOMMU_NONE so the IOMMU
+         * doesn't short-cut its translation table walk.
+         */
+        iotlb = imrc->translate(iommu_mr, addr, IOMMU_NONE, iommu_idx);
+        addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
+                | (addr & iotlb.addr_mask));
+        /* Update the caller's prot bits to remove permissions the IOMMU
+         * is giving us a failure response for. If we get down to no
+         * permissions left at all we can give up now.
+         */
+        if (!(iotlb.perm & IOMMU_RO)) {
+            *prot &= ~(PAGE_READ | PAGE_EXEC);
+        }
+        if (!(iotlb.perm & IOMMU_WO)) {
+            *prot &= ~PAGE_WRITE;
+        }
+
+        if (!*prot) {
+            goto translate_fail;
+        }
+
+        d = flatview_to_dispatch(address_space_to_flatview(iotlb.target_as));
+    }
 
     assert(!memory_region_is_iommu(section->mr));
+    *xlat = addr;
     return section;
+
+translate_fail:
+    return &d->map.sections[PHYS_SECTION_UNASSIGNED];
 }
 #endif
 
@@ -772,6 +949,9 @@ void cpu_exec_unrealizefn(CPUState *cpu)
     if (qdev_get_vmsd(DEVICE(cpu)) == NULL) {
         vmstate_unregister(NULL, &vmstate_cpu_common, cpu);
     }
+#ifndef CONFIG_USER_ONLY
+    tcg_iommu_free_notifier_list(cpu);
+#endif
 }
 
 Property cpu_common_props[] = {
@@ -819,6 +999,8 @@ void cpu_exec_realizefn(CPUState *cpu, Error **errp)
     if (cc->vmsd != NULL) {
         vmstate_register(NULL, cpu->cpu_index, cc->vmsd, cpu);
     }
+
+    cpu->iommu_notifiers = g_array_new(false, true, sizeof(TCGIOMMUNotifier));
 #endif
 }
 
@@ -846,15 +1028,40 @@ const char *parse_cpu_model(const char *cpu_model)
 }
 
 #if defined(CONFIG_USER_ONLY)
-static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
+void tb_invalidate_phys_addr(target_ulong addr)
 {
     mmap_lock();
-    tb_lock();
-    tb_invalidate_phys_page_range(pc, pc + 1, 0);
-    tb_unlock();
+    tb_invalidate_phys_page_range(addr, addr + 1, 0);
     mmap_unlock();
 }
+
+static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
+{
+    tb_invalidate_phys_addr(pc);
+}
 #else
+void tb_invalidate_phys_addr(AddressSpace *as, hwaddr addr, MemTxAttrs attrs)
+{
+    ram_addr_t ram_addr;
+    MemoryRegion *mr;
+    hwaddr l = 1;
+
+    if (!tcg_enabled()) {
+        return;
+    }
+
+    rcu_read_lock();
+    mr = address_space_translate(as, addr, &addr, &l, false, attrs);
+    if (!(memory_region_is_ram(mr)
+          || memory_region_is_romd(mr))) {
+        rcu_read_unlock();
+        return;
+    }
+    ram_addr = memory_region_get_ram_addr(mr) + addr;
+    tb_invalidate_phys_page_range(ram_addr, ram_addr + 1, 0);
+    rcu_read_unlock();
+}
+
 static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
 {
     MemTxAttrs attrs;
@@ -863,7 +1070,7 @@ static void breakpoint_invalidate(CPUState *cpu, target_ulong pc)
     if (phys != -1) {
         /* Locks grabbed by tb_invalidate_phys_addr */
         tb_invalidate_phys_addr(cpu->cpu_ases[asidx].as,
-                                phys | (pc & ~TARGET_PAGE_MASK));
+                                phys | (pc & ~TARGET_PAGE_MASK), attrs);
     }
 }
 #endif
@@ -1089,6 +1296,7 @@ void cpu_abort(CPUState *cpu, const char *fmt, ...)
         struct sigaction act;
         sigfillset(&act.sa_mask);
         act.sa_handler = SIG_DFL;
+        act.sa_flags = 0;
         sigaction(SIGABRT, &act, NULL);
     }
 #endif
@@ -1142,6 +1350,7 @@ static void tlb_reset_dirty_range_all(ram_addr_t start, ram_addr_t length)
     RAMBlock *block;
     ram_addr_t end;
 
+    assert(tcg_enabled());
     end = TARGET_PAGE_ALIGN(start + length);
     start &= TARGET_PAGE_MASK;
 
@@ -1488,18 +1697,13 @@ void ram_block_dump(Monitor *mon)
  */
 static int find_max_supported_pagesize(Object *obj, void *opaque)
 {
-    char *mem_path;
     long *hpsize_min = opaque;
 
     if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
-        mem_path = object_property_get_str(obj, "mem-path", NULL);
-        if (mem_path) {
-            long hpsize = qemu_mempath_getpagesize(mem_path);
-            if (hpsize < *hpsize_min) {
-                *hpsize_min = hpsize;
-            }
-        } else {
-            *hpsize_min = getpagesize();
+        long hpsize = host_memory_backend_pagesize(MEMORY_BACKEND(obj));
+
+        if (hpsize < *hpsize_min) {
+            *hpsize_min = hpsize;
         }
     }
 
@@ -1512,11 +1716,7 @@ long qemu_getrampagesize(void)
     long mainrampagesize;
     Object *memdev_root;
 
-    if (mem_path) {
-        mainrampagesize = qemu_mempath_getpagesize(mem_path);
-    } else {
-        mainrampagesize = getpagesize();
-    }
+    mainrampagesize = qemu_mempath_getpagesize(mem_path);
 
     /* it's possible we have memory-backend objects with
      * hugepage-backed RAM. these may get mapped into system
@@ -1645,6 +1845,10 @@ static void *file_ram_alloc(RAMBlock *block,
                    " must be multiples of page size 0x%zx",
                    block->mr->align, block->page_size);
         return NULL;
+    } else if (block->mr->align && !is_power_of_2(block->mr->align)) {
+        error_setg(errp, "alignment 0x%" PRIx64
+                   " must be a power of two", block->mr->align);
+        return NULL;
     }
     block->mr->align = MAX(block->page_size, block->mr->align);
 #if defined(__s390x__)
@@ -1757,7 +1961,7 @@ static ram_addr_t find_ram_offset(ram_addr_t size)
     return offset;
 }
 
-unsigned long last_ram_page(void)
+static unsigned long last_ram_page(void)
 {
     RAMBlock *block;
     ram_addr_t last = 0;
@@ -1806,6 +2010,21 @@ void qemu_ram_set_uf_zeroable(RAMBlock *rb)
     rb->flags |= RAM_UF_ZEROPAGE;
 }
 
+bool qemu_ram_is_migratable(RAMBlock *rb)
+{
+    return rb->flags & RAM_MIGRATABLE;
+}
+
+void qemu_ram_set_migratable(RAMBlock *rb)
+{
+    rb->flags |= RAM_MIGRATABLE;
+}
+
+void qemu_ram_unset_migratable(RAMBlock *rb)
+{
+    rb->flags &= ~RAM_MIGRATABLE;
+}
+
 /* Called with iothread lock held.  */
 void qemu_ram_set_idstr(RAMBlock *new_block, const char *name, DeviceState *dev)
 {
@@ -2455,21 +2674,22 @@ void memory_notdirty_write_prepare(NotDirtyInfo *ndi,
     ndi->ram_addr = ram_addr;
     ndi->mem_vaddr = mem_vaddr;
     ndi->size = size;
-    ndi->locked = false;
+    ndi->pages = NULL;
 
     assert(tcg_enabled());
     if (!cpu_physical_memory_get_dirty_flag(ram_addr, DIRTY_MEMORY_CODE)) {
-        ndi->locked = true;
-        tb_lock();
-        tb_invalidate_phys_page_fast(ram_addr, size);
+        ndi->pages = page_collection_lock(ram_addr, ram_addr + size);
+        tb_invalidate_phys_page_fast(ndi->pages, ram_addr, size);
     }
 }
 
 /* Called within RCU critical section. */
 void memory_notdirty_write_complete(NotDirtyInfo *ndi)
 {
-    if (ndi->locked) {
-        tb_unlock();
+    if (ndi->pages) {
+        assert(tcg_enabled());
+        page_collection_unlock(ndi->pages);
+        ndi->pages = NULL;
     }
 
     /* Set both VGA and migration bits for simplicity and to remove
@@ -2493,27 +2713,13 @@ static void notdirty_mem_write(void *opaque, hwaddr ram_addr,
     memory_notdirty_write_prepare(&ndi, current_cpu, current_cpu->mem_io_vaddr,
                          ram_addr, size);
 
-    switch (size) {
-    case 1:
-        stb_p(qemu_map_ram_ptr(NULL, ram_addr), val);
-        break;
-    case 2:
-        stw_p(qemu_map_ram_ptr(NULL, ram_addr), val);
-        break;
-    case 4:
-        stl_p(qemu_map_ram_ptr(NULL, ram_addr), val);
-        break;
-    case 8:
-        stq_p(qemu_map_ram_ptr(NULL, ram_addr), val);
-        break;
-    default:
-        abort();
-    }
+    stn_p(qemu_map_ram_ptr(NULL, ram_addr), size, val);
     memory_notdirty_write_complete(&ndi);
 }
 
 static bool notdirty_mem_accepts(void *opaque, hwaddr addr,
-                                 unsigned size, bool is_write)
+                                 unsigned size, bool is_write,
+                                 MemTxAttrs attrs)
 {
     return is_write;
 }
@@ -2570,18 +2776,16 @@ static void check_watchpoint(int offset, int len, MemTxAttrs attrs, int flags)
                 }
                 cpu->watchpoint_hit = wp;
 
-                /* Both tb_lock and iothread_mutex will be reset when
-                 * cpu_loop_exit or cpu_loop_exit_noexc longjmp
-                 * back into the cpu_exec main loop.
-                 */
-                tb_lock();
+                mmap_lock();
                 tb_check_watchpoint(cpu);
                 if (wp->flags & BP_STOP_BEFORE_ACCESS) {
                     cpu->exception_index = EXCP_DEBUG;
+                    mmap_unlock();
                     cpu_loop_exit(cpu);
                 } else {
                     /* Force execution of one insn next time.  */
                     cpu->cflags_next_tb = 1 | curr_cflags();
+                    mmap_unlock();
                     cpu_loop_exit_noexc(cpu);
                 }
             }
@@ -2670,7 +2874,7 @@ static MemTxResult flatview_read(FlatView *fv, hwaddr addr,
 static MemTxResult flatview_write(FlatView *fv, hwaddr addr, MemTxAttrs attrs,
                                   const uint8_t *buf, int len);
 static bool flatview_access_valid(FlatView *fv, hwaddr addr, int len,
-                                  bool is_write);
+                                  bool is_write, MemTxAttrs attrs);
 
 static MemTxResult subpage_read(void *opaque, hwaddr addr, uint64_t *data,
                                 unsigned len, MemTxAttrs attrs)
@@ -2687,22 +2891,8 @@ static MemTxResult subpage_read(void *opaque, hwaddr addr, uint64_t *data,
     if (res) {
         return res;
     }
-    switch (len) {
-    case 1:
-        *data = ldub_p(buf);
-        return MEMTX_OK;
-    case 2:
-        *data = lduw_p(buf);
-        return MEMTX_OK;
-    case 4:
-        *data = ldl_p(buf);
-        return MEMTX_OK;
-    case 8:
-        *data = ldq_p(buf);
-        return MEMTX_OK;
-    default:
-        abort();
-    }
+    *data = ldn_p(buf, len);
+    return MEMTX_OK;
 }
 
 static MemTxResult subpage_write(void *opaque, hwaddr addr,
@@ -2716,27 +2906,13 @@ static MemTxResult subpage_write(void *opaque, hwaddr addr,
            " value %"PRIx64"\n",
            __func__, subpage, len, addr, value);
 #endif
-    switch (len) {
-    case 1:
-        stb_p(buf, value);
-        break;
-    case 2:
-        stw_p(buf, value);
-        break;
-    case 4:
-        stl_p(buf, value);
-        break;
-    case 8:
-        stq_p(buf, value);
-        break;
-    default:
-        abort();
-    }
+    stn_p(buf, len, value);
     return flatview_write(subpage->fv, addr + subpage->base, attrs, buf, len);
 }
 
 static bool subpage_accepts(void *opaque, hwaddr addr,
-                            unsigned len, bool is_write)
+                            unsigned len, bool is_write,
+                            MemTxAttrs attrs)
 {
     subpage_t *subpage = opaque;
 #if defined(DEBUG_SUBPAGE)
@@ -2745,7 +2921,7 @@ static bool subpage_accepts(void *opaque, hwaddr addr,
 #endif
 
     return flatview_access_valid(subpage->fv, addr + subpage->base,
-                                 len, is_write);
+                                 len, is_write, attrs);
 }
 
 static const MemoryRegionOps subpage_ops = {
@@ -2819,7 +2995,8 @@ static void readonly_mem_write(void *opaque, hwaddr addr,
 }
 
 static bool readonly_mem_accepts(void *opaque, hwaddr addr,
-                                 unsigned size, bool is_write)
+                                 unsigned size, bool is_write,
+                                 MemTxAttrs attrs)
 {
     return is_write;
 }
@@ -2843,14 +3020,15 @@ static const MemoryRegionOps readonly_mem_ops = {
     },
 };
 
-MemoryRegion *iotlb_to_region(CPUState *cpu, hwaddr index, MemTxAttrs attrs)
+MemoryRegionSection *iotlb_to_section(CPUState *cpu,
+                                      hwaddr index, MemTxAttrs attrs)
 {
     int asidx = cpu_asidx_from_attrs(cpu, attrs);
     CPUAddressSpace *cpuas = &cpu->cpu_ases[asidx];
     AddressSpaceDispatch *d = atomic_rcu_read(&cpuas->memory_dispatch);
     MemoryRegionSection *sections = d->map.sections;
 
-    return sections[index & ~TARGET_PAGE_MASK].mr;
+    return &sections[index & ~TARGET_PAGE_MASK];
 }
 
 static void io_mem_init(void)
@@ -2901,6 +3079,7 @@ static void tcg_commit(MemoryListener *listener)
     CPUAddressSpace *cpuas;
     AddressSpaceDispatch *d;
 
+    assert(tcg_enabled());
     /* since each CPU stores ram addresses in its TLB cache, we must
        reset the modified entries */
     cpuas = container_of(listener, CPUAddressSpace, tcg_as_listener);
@@ -2998,9 +3177,7 @@ static void invalidate_and_set_dirty(MemoryRegion *mr, hwaddr addr,
     }
     if (dirty_log_mask & (1 << DIRTY_MEMORY_CODE)) {
         assert(tcg_enabled());
-        tb_lock();
         tb_invalidate_phys_range(addr, addr + length);
-        tb_unlock();
         dirty_log_mask &= ~(1 << DIRTY_MEMORY_CODE);
     }
     cpu_physical_memory_set_dirty_range(addr, length, dirty_log_mask);
@@ -3074,34 +3251,8 @@ static MemTxResult flatview_write_continue(FlatView *fv, hwaddr addr,
             l = memory_access_size(mr, l, addr1);
             /* XXX: could force current_cpu to NULL to avoid
                potential bugs */
-            switch (l) {
-            case 8:
-                /* 64 bit write access */
-                val = ldq_p(buf);
-                result |= memory_region_dispatch_write(mr, addr1, val, 8,
-                                                       attrs);
-                break;
-            case 4:
-                /* 32 bit write access */
-                val = (uint32_t)ldl_p(buf);
-                result |= memory_region_dispatch_write(mr, addr1, val, 4,
-                                                       attrs);
-                break;
-            case 2:
-                /* 16 bit write access */
-                val = lduw_p(buf);
-                result |= memory_region_dispatch_write(mr, addr1, val, 2,
-                                                       attrs);
-                break;
-            case 1:
-                /* 8 bit write access */
-                val = ldub_p(buf);
-                result |= memory_region_dispatch_write(mr, addr1, val, 1,
-                                                       attrs);
-                break;
-            default:
-                abort();
-            }
+            val = ldn_p(buf, l);
+            result |= memory_region_dispatch_write(mr, addr1, val, l, attrs);
         } else {
             /* RAM case */
             ptr = qemu_ram_ptr_length(mr->ram_block, addr1, &l, false);
@@ -3123,7 +3274,7 @@ static MemTxResult flatview_write_continue(FlatView *fv, hwaddr addr,
         }
 
         l = len;
-        mr = flatview_translate(fv, addr, &addr1, &l, true);
+        mr = flatview_translate(fv, addr, &addr1, &l, true, attrs);
     }
 
     return result;
@@ -3139,7 +3290,7 @@ static MemTxResult flatview_write(FlatView *fv, hwaddr addr, MemTxAttrs attrs,
     MemTxResult result = MEMTX_OK;
 
     l = len;
-    mr = flatview_translate(fv, addr, &addr1, &l, true);
+    mr = flatview_translate(fv, addr, &addr1, &l, true, attrs);
     result = flatview_write_continue(fv, addr, attrs, buf, len,
                                      addr1, l, mr);
 
@@ -3162,34 +3313,8 @@ MemTxResult flatview_read_continue(FlatView *fv, hwaddr addr,
             /* I/O case */
             release_lock |= prepare_mmio_access(mr);
             l = memory_access_size(mr, l, addr1);
-            switch (l) {
-            case 8:
-                /* 64 bit read access */
-                result |= memory_region_dispatch_read(mr, addr1, &val, 8,
-                                                      attrs);
-                stq_p(buf, val);
-                break;
-            case 4:
-                /* 32 bit read access */
-                result |= memory_region_dispatch_read(mr, addr1, &val, 4,
-                                                      attrs);
-                stl_p(buf, val);
-                break;
-            case 2:
-                /* 16 bit read access */
-                result |= memory_region_dispatch_read(mr, addr1, &val, 2,
-                                                      attrs);
-                stw_p(buf, val);
-                break;
-            case 1:
-                /* 8 bit read access */
-                result |= memory_region_dispatch_read(mr, addr1, &val, 1,
-                                                      attrs);
-                stb_p(buf, val);
-                break;
-            default:
-                abort();
-            }
+            result |= memory_region_dispatch_read(mr, addr1, &val, l, attrs);
+            stn_p(buf, l, val);
         } else {
             /* RAM case */
             ptr = qemu_ram_ptr_length(mr->ram_block, addr1, &l, false);
@@ -3210,7 +3335,7 @@ MemTxResult flatview_read_continue(FlatView *fv, hwaddr addr,
         }
 
         l = len;
-        mr = flatview_translate(fv, addr, &addr1, &l, false);
+        mr = flatview_translate(fv, addr, &addr1, &l, false, attrs);
     }
 
     return result;
@@ -3225,7 +3350,7 @@ static MemTxResult flatview_read(FlatView *fv, hwaddr addr,
     MemoryRegion *mr;
 
     l = len;
-    mr = flatview_translate(fv, addr, &addr1, &l, false);
+    mr = flatview_translate(fv, addr, &addr1, &l, false, attrs);
     return flatview_read_continue(fv, addr, attrs, buf, len,
                                   addr1, l, mr);
 }
@@ -3296,7 +3421,8 @@ static inline void cpu_physical_memory_write_rom_internal(AddressSpace *as,
     rcu_read_lock();
     while (len > 0) {
         l = len;
-        mr = address_space_translate(as, addr, &addr1, &l, true);
+        mr = address_space_translate(as, addr, &addr1, &l, true,
+                                     MEMTXATTRS_UNSPECIFIED);
 
         if (!(memory_region_is_ram(mr) ||
               memory_region_is_romd(mr))) {
@@ -3431,17 +3557,17 @@ static void cpu_notify_map_clients(void)
 }
 
 static bool flatview_access_valid(FlatView *fv, hwaddr addr, int len,
-                                  bool is_write)
+                                  bool is_write, MemTxAttrs attrs)
 {
     MemoryRegion *mr;
     hwaddr l, xlat;
 
     while (len > 0) {
         l = len;
-        mr = flatview_translate(fv, addr, &xlat, &l, is_write);
+        mr = flatview_translate(fv, addr, &xlat, &l, is_write, attrs);
         if (!memory_access_is_direct(mr, is_write)) {
             l = memory_access_size(mr, l, addr);
-            if (!memory_region_access_valid(mr, xlat, l, is_write)) {
+            if (!memory_region_access_valid(mr, xlat, l, is_write, attrs)) {
                 return false;
             }
         }
@@ -3453,23 +3579,24 @@ static bool flatview_access_valid(FlatView *fv, hwaddr addr, int len,
 }
 
 bool address_space_access_valid(AddressSpace *as, hwaddr addr,
-                                int len, bool is_write)
+                                int len, bool is_write,
+                                MemTxAttrs attrs)
 {
     FlatView *fv;
     bool result;
 
     rcu_read_lock();
     fv = address_space_to_flatview(as);
-    result = flatview_access_valid(fv, addr, len, is_write);
+    result = flatview_access_valid(fv, addr, len, is_write, attrs);
     rcu_read_unlock();
     return result;
 }
 
 static hwaddr
 flatview_extend_translation(FlatView *fv, hwaddr addr,
-                                 hwaddr target_len,
-                                 MemoryRegion *mr, hwaddr base, hwaddr len,
-                                 bool is_write)
+                            hwaddr target_len,
+                            MemoryRegion *mr, hwaddr base, hwaddr len,
+                            bool is_write, MemTxAttrs attrs)
 {
     hwaddr done = 0;
     hwaddr xlat;
@@ -3485,7 +3612,7 @@ flatview_extend_translation(FlatView *fv, hwaddr addr,
 
         len = target_len;
         this_mr = flatview_translate(fv, addr, &xlat,
-                                                   &len, is_write);
+                                     &len, is_write, attrs);
         if (this_mr != mr || xlat != base + done) {
             return done;
         }
@@ -3502,7 +3629,8 @@ flatview_extend_translation(FlatView *fv, hwaddr addr,
 void *address_space_map(AddressSpace *as,
                         hwaddr addr,
                         hwaddr *plen,
-                        bool is_write)
+                        bool is_write,
+                        MemTxAttrs attrs)
 {
     hwaddr len = *plen;
     hwaddr l, xlat;
@@ -3517,7 +3645,7 @@ void *address_space_map(AddressSpace *as,
     l = len;
     rcu_read_lock();
     fv = address_space_to_flatview(as);
-    mr = flatview_translate(fv, addr, &xlat, &l, is_write);
+    mr = flatview_translate(fv, addr, &xlat, &l, is_write, attrs);
 
     if (!memory_access_is_direct(mr, is_write)) {
         if (atomic_xchg(&bounce.in_use, true)) {
@@ -3545,7 +3673,7 @@ void *address_space_map(AddressSpace *as,
 
     memory_region_ref(mr);
     *plen = flatview_extend_translation(fv, addr, len, mr, xlat,
-                                             l, is_write);
+                                        l, is_write, attrs);
     ptr = qemu_ram_ptr_length(mr->ram_block, xlat, plen, true);
     rcu_read_unlock();
 
@@ -3589,7 +3717,8 @@ void *cpu_physical_memory_map(hwaddr addr,
                               hwaddr *plen,
                               int is_write)
 {
-    return address_space_map(&address_space_memory, addr, plen, is_write);
+    return address_space_map(&address_space_memory, addr, plen, is_write,
+                             MEMTXATTRS_UNSPECIFIED);
 }
 
 void cpu_physical_memory_unmap(void *buffer, hwaddr len,
@@ -3602,9 +3731,6 @@ void cpu_physical_memory_unmap(void *buffer, hwaddr len,
 #define ARG1                     as
 #define SUFFIX
 #define TRANSLATE(...)           address_space_translate(as, __VA_ARGS__)
-#define IS_DIRECT(mr, is_write)  memory_access_is_direct(mr, is_write)
-#define MAP_RAM(mr, ofs)         qemu_map_ram_ptr((mr)->ram_block, ofs)
-#define INVALIDATE(mr, ofs, len) invalidate_and_set_dirty(mr, ofs, len)
 #define RCU_READ_LOCK(...)       rcu_read_lock()
 #define RCU_READ_UNLOCK(...)     rcu_read_unlock()
 #include "memory_ldst.inc.c"
@@ -3615,33 +3741,134 @@ int64_t address_space_cache_init(MemoryRegionCache *cache,
                                  hwaddr len,
                                  bool is_write)
 {
-    cache->len = len;
-    cache->as = as;
-    cache->xlat = addr;
-    return len;
+    AddressSpaceDispatch *d;
+    hwaddr l;
+    MemoryRegion *mr;
+
+    assert(len > 0);
+
+    l = len;
+    cache->fv = address_space_get_flatview(as);
+    d = flatview_to_dispatch(cache->fv);
+    cache->mrs = *address_space_translate_internal(d, addr, &cache->xlat, &l, true);
+
+    mr = cache->mrs.mr;
+    memory_region_ref(mr);
+    if (memory_access_is_direct(mr, is_write)) {
+        /* We don't care about the memory attributes here as we're only
+         * doing this if we found actual RAM, which behaves the same
+         * regardless of attributes; so UNSPECIFIED is fine.
+         */
+        l = flatview_extend_translation(cache->fv, addr, len, mr,
+                                        cache->xlat, l, is_write,
+                                        MEMTXATTRS_UNSPECIFIED);
+        cache->ptr = qemu_ram_ptr_length(mr->ram_block, cache->xlat, &l, true);
+    } else {
+        cache->ptr = NULL;
+    }
+
+    cache->len = l;
+    cache->is_write = is_write;
+    return l;
 }
 
 void address_space_cache_invalidate(MemoryRegionCache *cache,
                                     hwaddr addr,
                                     hwaddr access_len)
 {
+    assert(cache->is_write);
+    if (likely(cache->ptr)) {
+        invalidate_and_set_dirty(cache->mrs.mr, addr + cache->xlat, access_len);
+    }
 }
 
 void address_space_cache_destroy(MemoryRegionCache *cache)
 {
-    cache->as = NULL;
+    if (!cache->mrs.mr) {
+        return;
+    }
+
+    if (xen_enabled()) {
+        xen_invalidate_map_cache_entry(cache->ptr);
+    }
+    memory_region_unref(cache->mrs.mr);
+    flatview_unref(cache->fv);
+    cache->mrs.mr = NULL;
+    cache->fv = NULL;
+}
+
+/* Called from RCU critical section.  This function has the same
+ * semantics as address_space_translate, but it only works on a
+ * predefined range of a MemoryRegion that was mapped with
+ * address_space_cache_init.
+ */
+static inline MemoryRegion *address_space_translate_cached(
+    MemoryRegionCache *cache, hwaddr addr, hwaddr *xlat,
+    hwaddr *plen, bool is_write, MemTxAttrs attrs)
+{
+    MemoryRegionSection section;
+    MemoryRegion *mr;
+    IOMMUMemoryRegion *iommu_mr;
+    AddressSpace *target_as;
+
+    assert(!cache->ptr);
+    *xlat = addr + cache->xlat;
+
+    mr = cache->mrs.mr;
+    iommu_mr = memory_region_get_iommu(mr);
+    if (!iommu_mr) {
+        /* MMIO region.  */
+        return mr;
+    }
+
+    section = address_space_translate_iommu(iommu_mr, xlat, plen,
+                                            NULL, is_write, true,
+                                            &target_as, attrs);
+    return section.mr;
+}
+
+/* Called from RCU critical section. address_space_read_cached uses this
+ * out of line function when the target is an MMIO or IOMMU region.
+ */
+void
+address_space_read_cached_slow(MemoryRegionCache *cache, hwaddr addr,
+                                   void *buf, int len)
+{
+    hwaddr addr1, l;
+    MemoryRegion *mr;
+
+    l = len;
+    mr = address_space_translate_cached(cache, addr, &addr1, &l, false,
+                                        MEMTXATTRS_UNSPECIFIED);
+    flatview_read_continue(cache->fv,
+                           addr, MEMTXATTRS_UNSPECIFIED, buf, len,
+                           addr1, l, mr);
+}
+
+/* Called from RCU critical section. address_space_write_cached uses this
+ * out of line function when the target is an MMIO or IOMMU region.
+ */
+void
+address_space_write_cached_slow(MemoryRegionCache *cache, hwaddr addr,
+                                    const void *buf, int len)
+{
+    hwaddr addr1, l;
+    MemoryRegion *mr;
+
+    l = len;
+    mr = address_space_translate_cached(cache, addr, &addr1, &l, true,
+                                        MEMTXATTRS_UNSPECIFIED);
+    flatview_write_continue(cache->fv,
+                            addr, MEMTXATTRS_UNSPECIFIED, buf, len,
+                            addr1, l, mr);
 }
 
 #define ARG1_DECL                MemoryRegionCache *cache
 #define ARG1                     cache
-#define SUFFIX                   _cached
-#define TRANSLATE(addr, ...)     \
-    address_space_translate(cache->as, cache->xlat + (addr), __VA_ARGS__)
-#define IS_DIRECT(mr, is_write)  true
-#define MAP_RAM(mr, ofs)         qemu_map_ram_ptr((mr)->ram_block, ofs)
-#define INVALIDATE(mr, ofs, len) invalidate_and_set_dirty(mr, ofs, len)
-#define RCU_READ_LOCK()          rcu_read_lock()
-#define RCU_READ_UNLOCK()        rcu_read_unlock()
+#define SUFFIX                   _cached_slow
+#define TRANSLATE(...)           address_space_translate_cached(cache, __VA_ARGS__)
+#define RCU_READ_LOCK()          ((void)0)
+#define RCU_READ_UNLOCK()        ((void)0)
 #include "memory_ldst.inc.c"
 
 /* virtual memory access for debug (includes writing to ROM) */
@@ -3725,7 +3952,8 @@ bool cpu_physical_memory_is_io(hwaddr phys_addr)
 
     rcu_read_lock();
     mr = address_space_translate(&address_space_memory,
-                                 phys_addr, &phys_addr, &l, false);
+                                 phys_addr, &phys_addr, &l, false,
+                                 MEMTXATTRS_UNSPECIFIED);
 
     res = !(memory_region_is_ram(mr) || memory_region_is_romd(mr));
     rcu_read_unlock();
@@ -3749,6 +3977,26 @@ int qemu_ram_foreach_block(RAMBlockIterFunc func, void *opaque)
     return ret;
 }
 
+int qemu_ram_foreach_migratable_block(RAMBlockIterFunc func, void *opaque)
+{
+    RAMBlock *block;
+    int ret = 0;
+
+    rcu_read_lock();
+    RAMBLOCK_FOREACH(block) {
+        if (!qemu_ram_is_migratable(block)) {
+            continue;
+        }
+        ret = func(block->idstr, block->host, block->offset,
+                   block->used_length, opaque);
+        if (ret) {
+            break;
+        }
+    }
+    rcu_read_unlock();
+    return ret;
+}
+
 /*
  * Unmap pages of memory from start to start+length such that
  * they a) read as 0, b) Trigger whatever fault mechanism