target/i386/hax-mem.c

   1 /*
   2  * HAX memory mapping operations
   3  *
   4  * Copyright (c) 2015-16 Intel Corporation
   5  * Copyright 2016 Google, Inc.
   6  *
   7  * This work is licensed under the terms of the GNU GPL, version 2.  See
   8  * the COPYING file in the top-level directory.
   9  */
  10
  11 #include "qemu/osdep.h"
  12 #include "cpu.h"
  13 #include "exec/address-spaces.h"
  14 #include "exec/exec-all.h"
  15 #include "qemu/error-report.h"
  16
  17 #include "target/i386/hax-i386.h"
  18 #include "qemu/queue.h"
  19
  20 #define DEBUG_HAX_MEM 0
  21
  22 #define DPRINTF(fmt, ...) \
  23     do { \
  24         if (DEBUG_HAX_MEM) { \
  25             fprintf(stdout, fmt, ## __VA_ARGS__); \
  26         } \
  27     } while (0)
  28
  29 /**
  30  * HAXMapping: describes a pending guest physical memory mapping
  31  *
  32  * @start_pa: a guest physical address marking the start of the region; must be
  33  *            page-aligned
  34  * @size: a guest physical address marking the end of the region; must be
  35  *          page-aligned
  36  * @host_va: the host virtual address of the start of the mapping
  37  * @flags: mapping parameters e.g. HAX_RAM_INFO_ROM or HAX_RAM_INFO_INVALID
  38  * @entry: additional fields for linking #HAXMapping instances together
  39  */
  40 typedef struct HAXMapping {
  41     uint64_t start_pa;
  42     uint32_t size;
  43     uint64_t host_va;
  44     int flags;
  45     QTAILQ_ENTRY(HAXMapping) entry;
  46 } HAXMapping;
  47
  48 /*
  49  * A doubly-linked list (actually a tail queue) of the pending page mappings
  50  * for the ongoing memory transaction.
  51  *
  52  * It is used to optimize the number of page mapping updates done through the
  53  * kernel module. For example, it's effective when a driver is digging an MMIO
  54  * hole inside an existing memory mapping. It will get a deletion of the whole
  55  * region, then the addition of the 2 remaining RAM areas around the hole and
  56  * finally the memory transaction commit. During the commit, it will effectively
  57  * send to the kernel only the removal of the pages from the MMIO hole after
  58  * having computed locally the result of the deletion and additions.
  59  */
  60 static QTAILQ_HEAD(HAXMappingListHead, HAXMapping) mappings =
  61     QTAILQ_HEAD_INITIALIZER(mappings);
  62
  63 /**
  64  * hax_mapping_dump_list: dumps @mappings to stdout (for debugging)
  65  */
  66 static void hax_mapping_dump_list(void)
  67 {
  68     HAXMapping *entry;
  69
  70     DPRINTF("%s updates:\n", __func__);
  71     QTAILQ_FOREACH(entry, &mappings, entry) {
  72         DPRINTF("\t%c 0x%016" PRIx64 "->0x%016" PRIx64 " VA 0x%016" PRIx64
  73                 "%s\n", entry->flags & HAX_RAM_INFO_INVALID ? '-' : '+',
  74                 entry->start_pa, entry->start_pa + entry->size, entry->host_va,
  75                 entry->flags & HAX_RAM_INFO_ROM ? " ROM" : "");
  76     }
  77 }
  78
  79 static void hax_insert_mapping_before(HAXMapping *next, uint64_t start_pa,
  80                                       uint32_t size, uint64_t host_va,
  81                                       uint8_t flags)
  82 {
  83     HAXMapping *entry;
  84
  85     entry = g_malloc0(sizeof(*entry));
  86     entry->start_pa = start_pa;
  87     entry->size = size;
  88     entry->host_va = host_va;
  89     entry->flags = flags;
  90     if (!next) {
  91         QTAILQ_INSERT_TAIL(&mappings, entry, entry);
  92     } else {
  93         QTAILQ_INSERT_BEFORE(next, entry, entry);
  94     }
  95 }
  96
  97 static bool hax_mapping_is_opposite(HAXMapping *entry, uint64_t host_va,
  98                                     uint8_t flags)
  99 {
 100     /* removed then added without change for the read-only flag */
 101     bool nop_flags = (entry->flags ^ flags) == HAX_RAM_INFO_INVALID;
 102
 103     return (entry->host_va == host_va) && nop_flags;
 104 }
 105
 106 static void hax_update_mapping(uint64_t start_pa, uint32_t size,
 107                                uint64_t host_va, uint8_t flags)
 108 {
 109     uint64_t end_pa = start_pa + size;
 110     HAXMapping *entry, *next;
 111
 112     QTAILQ_FOREACH_SAFE(entry, &mappings, entry, next) {
 113         uint32_t chunk_sz;
 114         if (start_pa >= entry->start_pa + entry->size) {
 115             continue;
 116         }
 117         if (start_pa < entry->start_pa) {
 118             chunk_sz = end_pa <= entry->start_pa ? size
 119                                                  : entry->start_pa - start_pa;
 120             hax_insert_mapping_before(entry, start_pa, chunk_sz,
 121                                       host_va, flags);
 122             start_pa += chunk_sz;
 123             host_va += chunk_sz;
 124             size -= chunk_sz;
 125         } else if (start_pa > entry->start_pa) {
 126             /* split the existing chunk at start_pa */
 127             chunk_sz = start_pa - entry->start_pa;
 128             hax_insert_mapping_before(entry, entry->start_pa, chunk_sz,
 129                                       entry->host_va, entry->flags);
 130             entry->start_pa += chunk_sz;
 131             entry->host_va += chunk_sz;
 132             entry->size -= chunk_sz;
 133         }
 134         /* now start_pa == entry->start_pa */
 135         chunk_sz = MIN(size, entry->size);
 136         if (chunk_sz) {
 137             bool nop = hax_mapping_is_opposite(entry, host_va, flags);
 138             bool partial = chunk_sz < entry->size;
 139             if (partial) {
 140                 /* remove the beginning of the existing chunk */
 141                 entry->start_pa += chunk_sz;
 142                 entry->host_va += chunk_sz;
 143                 entry->size -= chunk_sz;
 144                 if (!nop) {
 145                     hax_insert_mapping_before(entry, start_pa, chunk_sz,
 146                                               host_va, flags);
 147                 }
 148             } else { /* affects the full mapping entry */
 149                 if (nop) { /* no change to this mapping, remove it */
 150                     QTAILQ_REMOVE(&mappings, entry, entry);
 151                     g_free(entry);
 152                 } else { /* update mapping properties */
 153                     entry->host_va = host_va;
 154                     entry->flags = flags;
 155                 }
 156             }
 157             start_pa += chunk_sz;
 158             host_va += chunk_sz;
 159             size -= chunk_sz;
 160         }
 161         if (!size) { /* we are done */
 162             break;
 163         }
 164     }
 165     if (size) { /* add the leftover */
 166         hax_insert_mapping_before(NULL, start_pa, size, host_va, flags);
 167     }
 168 }
 169
 170 static void hax_process_section(MemoryRegionSection *section, uint8_t flags)
 171 {
 172     MemoryRegion *mr = section->mr;
 173     hwaddr start_pa = section->offset_within_address_space;
 174     ram_addr_t size = int128_get64(section->size);
 175     unsigned int delta;
 176     uint64_t host_va;
 177
 178     /* We only care about RAM and ROM regions */
 179     if (!memory_region_is_ram(mr)) {
 180         if (memory_region_is_romd(mr)) {
 181             /* HAXM kernel module does not support ROMD yet  */
 182             warn_report("Ignoring ROMD region 0x%016" PRIx64 "->0x%016" PRIx64,
 183                         start_pa, start_pa + size);
 184         }
 185         return;
 186     }
 187
 188     /* Adjust start_pa and size so that they are page-aligned. (Cf
 189      * kvm_set_phys_mem() in kvm-all.c).
 190      */
 191     delta = qemu_real_host_page_size - (start_pa & ~qemu_real_host_page_mask);
 192     delta &= ~qemu_real_host_page_mask;
 193     if (delta > size) {
 194         return;
 195     }
 196     start_pa += delta;
 197     size -= delta;
 198     size &= qemu_real_host_page_mask;
 199     if (!size || (start_pa & ~qemu_real_host_page_mask)) {
 200         return;
 201     }
 202
 203     host_va = (uintptr_t)memory_region_get_ram_ptr(mr)
 204             + section->offset_within_region + delta;
 205     if (memory_region_is_rom(section->mr)) {
 206         flags |= HAX_RAM_INFO_ROM;
 207     }
 208
 209     /* the kernel module interface uses 32-bit sizes (but we could split...) */
 210     g_assert(size <= UINT32_MAX);
 211
 212     hax_update_mapping(start_pa, size, host_va, flags);
 213 }
 214
 215 static void hax_region_add(MemoryListener *listener,
 216                            MemoryRegionSection *section)
 217 {
 218     memory_region_ref(section->mr);
 219     hax_process_section(section, 0);
 220 }
 221
 222 static void hax_region_del(MemoryListener *listener,
 223                            MemoryRegionSection *section)
 224 {
 225     hax_process_section(section, HAX_RAM_INFO_INVALID);
 226     memory_region_unref(section->mr);
 227 }
 228
 229 static void hax_transaction_begin(MemoryListener *listener)
 230 {
 231     g_assert(QTAILQ_EMPTY(&mappings));
 232 }
 233
 234 static void hax_transaction_commit(MemoryListener *listener)
 235 {
 236     if (!QTAILQ_EMPTY(&mappings)) {
 237         HAXMapping *entry, *next;
 238
 239         if (DEBUG_HAX_MEM) {
 240             hax_mapping_dump_list();
 241         }
 242         QTAILQ_FOREACH_SAFE(entry, &mappings, entry, next) {
 243             if (entry->flags & HAX_RAM_INFO_INVALID) {
 244                 /* for unmapping, put the values expected by the kernel */
 245                 entry->flags = HAX_RAM_INFO_INVALID;
 246                 entry->host_va = 0;
 247             }
 248             if (hax_set_ram(entry->start_pa, entry->size,
 249                             entry->host_va, entry->flags)) {
 250                 fprintf(stderr, "%s: Failed mapping @0x%016" PRIx64 "+0x%"
 251                         PRIx32 " flags %02x\n", __func__, entry->start_pa,
 252                         entry->size, entry->flags);
 253             }
 254             QTAILQ_REMOVE(&mappings, entry, entry);
 255             g_free(entry);
 256         }
 257     }
 258 }
 259
 260 /* currently we fake the dirty bitmap sync, always dirty */
 261 static void hax_log_sync(MemoryListener *listener,
 262                          MemoryRegionSection *section)
 263 {
 264     MemoryRegion *mr = section->mr;
 265
 266     if (!memory_region_is_ram(mr)) {
 267         /* Skip MMIO regions */
 268         return;
 269     }
 270
 271     memory_region_set_dirty(mr, 0, int128_get64(section->size));
 272 }
 273
 274 static MemoryListener hax_memory_listener = {
 275     .begin = hax_transaction_begin,
 276     .commit = hax_transaction_commit,
 277     .region_add = hax_region_add,
 278     .region_del = hax_region_del,
 279     .log_sync = hax_log_sync,
 280     .priority = 10,
 281 };
 282
 283 static void hax_ram_block_added(RAMBlockNotifier *n, void *host, size_t size)
 284 {
 285     /*
 286      * In HAX, QEMU allocates the virtual address, and HAX kernel
 287      * populates the memory with physical memory. Currently we have no
 288      * paging, so user should make sure enough free memory in advance.
 289      */
 290     if (hax_populate_ram((uint64_t)(uintptr_t)host, size) < 0) {
 291         fprintf(stderr, "HAX failed to populate RAM");
 292         abort();
 293     }
 294 }
 295
 296 static struct RAMBlockNotifier hax_ram_notifier = {
 297     .ram_block_added = hax_ram_block_added,
 298 };
 299
 300 void hax_memory_init(void)
 301 {
 302     ram_block_notifier_add(&hax_ram_notifier);
 303     memory_listener_register(&hax_memory_listener, &address_space_memory);
 304 }