| 1 | /* |
| 2 | * QEMU Xen emulation: Grant table support |
| 3 | * |
| 4 | * Copyright © 2022 Amazon.com, Inc. or its affiliates. All Rights Reserved. |
| 5 | * |
| 6 | * Authors: David Woodhouse <dwmw2@infradead.org> |
| 7 | * |
| 8 | * This work is licensed under the terms of the GNU GPL, version 2 or later. |
| 9 | * See the COPYING file in the top-level directory. |
| 10 | */ |
| 11 | |
| 12 | #include "qemu/osdep.h" |
| 13 | #include "qemu/host-utils.h" |
| 14 | #include "qemu/module.h" |
| 15 | #include "qemu/lockable.h" |
| 16 | #include "qemu/main-loop.h" |
| 17 | #include "qapi/error.h" |
| 18 | #include "qom/object.h" |
| 19 | #include "exec/target_page.h" |
| 20 | #include "system/address-spaces.h" |
| 21 | #include "migration/vmstate.h" |
| 22 | |
| 23 | #include "hw/core/sysbus.h" |
| 24 | #include "hw/xen/xen.h" |
| 25 | #include "hw/xen/xen_backend_ops.h" |
| 26 | #include "xen_overlay.h" |
| 27 | #include "xen_gnttab.h" |
| 28 | #include "xen_primary_console.h" |
| 29 | |
| 30 | #include "system/kvm.h" |
| 31 | #include "system/kvm_xen.h" |
| 32 | |
| 33 | #include "hw/xen/interface/memory.h" |
| 34 | #include "hw/xen/interface/grant_table.h" |
| 35 | |
| 36 | #define TYPE_XEN_GNTTAB "xen-gnttab" |
| 37 | OBJECT_DECLARE_SIMPLE_TYPE(XenGnttabState, XEN_GNTTAB) |
| 38 | |
| 39 | #define ENTRIES_PER_FRAME_V1 (XEN_PAGE_SIZE / sizeof(grant_entry_v1_t)) |
| 40 | |
| 41 | static struct gnttab_backend_ops emu_gnttab_backend_ops; |
| 42 | |
| 43 | struct XenGnttabState { |
| 44 | /*< private >*/ |
| 45 | SysBusDevice busdev; |
| 46 | /*< public >*/ |
| 47 | |
| 48 | QemuMutex gnt_lock; |
| 49 | |
| 50 | uint32_t nr_frames; |
| 51 | uint32_t max_frames; |
| 52 | |
| 53 | union { |
| 54 | grant_entry_v1_t *v1; |
| 55 | /* Theoretically, v2 support could be added here. */ |
| 56 | } entries; |
| 57 | |
| 58 | MemoryRegion gnt_frames; |
| 59 | MemoryRegion *gnt_aliases; |
| 60 | uint64_t *gnt_frame_gpas; |
| 61 | |
| 62 | uint8_t *map_track; |
| 63 | }; |
| 64 | |
| 65 | struct XenGnttabState *xen_gnttab_singleton; |
| 66 | |
| 67 | static void xen_gnttab_realize(DeviceState *dev, Error **errp) |
| 68 | { |
| 69 | XenGnttabState *s = XEN_GNTTAB(dev); |
| 70 | int i; |
| 71 | |
| 72 | if (xen_mode != XEN_EMULATE) { |
| 73 | error_setg(errp, "Xen grant table support is for Xen emulation"); |
| 74 | return; |
| 75 | } |
| 76 | s->max_frames = kvm_xen_get_gnttab_max_frames(); |
| 77 | memory_region_init_ram(&s->gnt_frames, OBJECT(dev), "xen:grant_table", |
| 78 | XEN_PAGE_SIZE * s->max_frames, &error_abort); |
| 79 | memory_region_set_enabled(&s->gnt_frames, true); |
| 80 | s->entries.v1 = memory_region_get_ram_ptr(&s->gnt_frames); |
| 81 | |
| 82 | /* Create individual page-sizes aliases for overlays */ |
| 83 | s->gnt_aliases = (void *)g_new0(MemoryRegion, s->max_frames); |
| 84 | s->gnt_frame_gpas = (void *)g_new(uint64_t, s->max_frames); |
| 85 | for (i = 0; i < s->max_frames; i++) { |
| 86 | memory_region_init_alias(&s->gnt_aliases[i], OBJECT(dev), |
| 87 | NULL, &s->gnt_frames, |
| 88 | i * XEN_PAGE_SIZE, XEN_PAGE_SIZE); |
| 89 | s->gnt_frame_gpas[i] = INVALID_GPA; |
| 90 | } |
| 91 | |
| 92 | s->nr_frames = 0; |
| 93 | memset(s->entries.v1, 0, XEN_PAGE_SIZE * s->max_frames); |
| 94 | s->entries.v1[GNTTAB_RESERVED_XENSTORE].flags = GTF_permit_access; |
| 95 | s->entries.v1[GNTTAB_RESERVED_XENSTORE].frame = XEN_SPECIAL_PFN(XENSTORE); |
| 96 | |
| 97 | qemu_mutex_init(&s->gnt_lock); |
| 98 | |
| 99 | xen_gnttab_singleton = s; |
| 100 | |
| 101 | s->map_track = g_new0(uint8_t, s->max_frames * ENTRIES_PER_FRAME_V1); |
| 102 | |
| 103 | xen_gnttab_ops = &emu_gnttab_backend_ops; |
| 104 | } |
| 105 | |
| 106 | static int xen_gnttab_post_load(void *opaque, int version_id) |
| 107 | { |
| 108 | XenGnttabState *s = XEN_GNTTAB(opaque); |
| 109 | uint32_t i; |
| 110 | |
| 111 | for (i = 0; i < s->nr_frames; i++) { |
| 112 | if (s->gnt_frame_gpas[i] != INVALID_GPA) { |
| 113 | xen_overlay_do_map_page(&s->gnt_aliases[i], s->gnt_frame_gpas[i]); |
| 114 | } |
| 115 | } |
| 116 | return 0; |
| 117 | } |
| 118 | |
| 119 | static bool xen_gnttab_is_needed(void *opaque) |
| 120 | { |
| 121 | return xen_mode == XEN_EMULATE; |
| 122 | } |
| 123 | |
| 124 | static const VMStateDescription xen_gnttab_vmstate = { |
| 125 | .name = "xen_gnttab", |
| 126 | .version_id = 1, |
| 127 | .minimum_version_id = 1, |
| 128 | .needed = xen_gnttab_is_needed, |
| 129 | .post_load = xen_gnttab_post_load, |
| 130 | .fields = (const VMStateField[]) { |
| 131 | VMSTATE_UINT32(nr_frames, XenGnttabState), |
| 132 | VMSTATE_VARRAY_UINT32(gnt_frame_gpas, XenGnttabState, nr_frames, 0, |
| 133 | vmstate_info_uint64, uint64_t), |
| 134 | VMSTATE_END_OF_LIST() |
| 135 | } |
| 136 | }; |
| 137 | |
| 138 | static void xen_gnttab_class_init(ObjectClass *klass, const void *data) |
| 139 | { |
| 140 | DeviceClass *dc = DEVICE_CLASS(klass); |
| 141 | |
| 142 | dc->realize = xen_gnttab_realize; |
| 143 | dc->vmsd = &xen_gnttab_vmstate; |
| 144 | } |
| 145 | |
| 146 | static const TypeInfo xen_gnttab_info = { |
| 147 | .name = TYPE_XEN_GNTTAB, |
| 148 | .parent = TYPE_SYS_BUS_DEVICE, |
| 149 | .instance_size = sizeof(XenGnttabState), |
| 150 | .class_init = xen_gnttab_class_init, |
| 151 | }; |
| 152 | |
| 153 | void xen_gnttab_create(void) |
| 154 | { |
| 155 | xen_gnttab_singleton = XEN_GNTTAB(sysbus_create_simple(TYPE_XEN_GNTTAB, |
| 156 | -1, NULL)); |
| 157 | } |
| 158 | |
| 159 | static void xen_gnttab_register_types(void) |
| 160 | { |
| 161 | type_register_static(&xen_gnttab_info); |
| 162 | } |
| 163 | |
| 164 | type_init(xen_gnttab_register_types) |
| 165 | |
| 166 | int xen_gnttab_map_page(uint64_t idx, uint64_t gfn) |
| 167 | { |
| 168 | XenGnttabState *s = xen_gnttab_singleton; |
| 169 | uint64_t gpa = gfn << XEN_PAGE_SHIFT; |
| 170 | |
| 171 | if (!s) { |
| 172 | return -ENOTSUP; |
| 173 | } |
| 174 | |
| 175 | if (idx >= s->max_frames) { |
| 176 | return -EINVAL; |
| 177 | } |
| 178 | |
| 179 | BQL_LOCK_GUARD(); |
| 180 | QEMU_LOCK_GUARD(&s->gnt_lock); |
| 181 | |
| 182 | xen_overlay_do_map_page(&s->gnt_aliases[idx], gpa); |
| 183 | |
| 184 | s->gnt_frame_gpas[idx] = gpa; |
| 185 | |
| 186 | if (s->nr_frames <= idx) { |
| 187 | s->nr_frames = idx + 1; |
| 188 | } |
| 189 | |
| 190 | return 0; |
| 191 | } |
| 192 | |
| 193 | int xen_gnttab_set_version_op(struct gnttab_set_version *set) |
| 194 | { |
| 195 | int ret; |
| 196 | |
| 197 | switch (set->version) { |
| 198 | case 1: |
| 199 | ret = 0; |
| 200 | break; |
| 201 | |
| 202 | case 2: |
| 203 | /* Behave as before set_version was introduced. */ |
| 204 | ret = -ENOSYS; |
| 205 | break; |
| 206 | |
| 207 | default: |
| 208 | ret = -EINVAL; |
| 209 | } |
| 210 | |
| 211 | set->version = 1; |
| 212 | return ret; |
| 213 | } |
| 214 | |
| 215 | int xen_gnttab_get_version_op(struct gnttab_get_version *get) |
| 216 | { |
| 217 | if (get->dom != DOMID_SELF && get->dom != xen_domid) { |
| 218 | return -ESRCH; |
| 219 | } |
| 220 | |
| 221 | get->version = 1; |
| 222 | return 0; |
| 223 | } |
| 224 | |
| 225 | int xen_gnttab_query_size_op(struct gnttab_query_size *size) |
| 226 | { |
| 227 | XenGnttabState *s = xen_gnttab_singleton; |
| 228 | |
| 229 | if (!s) { |
| 230 | return -ENOTSUP; |
| 231 | } |
| 232 | |
| 233 | if (size->dom != DOMID_SELF && size->dom != xen_domid) { |
| 234 | size->status = GNTST_bad_domain; |
| 235 | return 0; |
| 236 | } |
| 237 | |
| 238 | size->status = GNTST_okay; |
| 239 | size->nr_frames = s->nr_frames; |
| 240 | size->max_nr_frames = s->max_frames; |
| 241 | return 0; |
| 242 | } |
| 243 | |
| 244 | /* Track per-open refs, to allow close() to clean up. */ |
| 245 | struct active_ref { |
| 246 | MemoryRegionSection mrs; |
| 247 | void *virtaddr; |
| 248 | uint32_t refcnt; |
| 249 | int prot; |
| 250 | }; |
| 251 | |
| 252 | static void gnt_unref(XenGnttabState *s, grant_ref_t ref, |
| 253 | MemoryRegionSection *mrs, int prot) |
| 254 | { |
| 255 | if (mrs && mrs->mr) { |
| 256 | if (prot & PROT_WRITE) { |
| 257 | memory_region_set_dirty(mrs->mr, mrs->offset_within_region, |
| 258 | XEN_PAGE_SIZE); |
| 259 | } |
| 260 | memory_region_unref(mrs->mr); |
| 261 | mrs->mr = NULL; |
| 262 | } |
| 263 | assert(s->map_track[ref] != 0); |
| 264 | |
| 265 | if (--s->map_track[ref] == 0) { |
| 266 | grant_entry_v1_t *gnt_p = &s->entries.v1[ref]; |
| 267 | qatomic_and(&gnt_p->flags, (uint16_t)~(GTF_reading | GTF_writing)); |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | static uint64_t gnt_ref(XenGnttabState *s, grant_ref_t ref, int prot) |
| 272 | { |
| 273 | uint16_t mask = GTF_type_mask | GTF_sub_page; |
| 274 | grant_entry_v1_t gnt, *gnt_p; |
| 275 | int retries = 0; |
| 276 | |
| 277 | if (ref >= s->max_frames * ENTRIES_PER_FRAME_V1 || |
| 278 | s->map_track[ref] == UINT8_MAX) { |
| 279 | return INVALID_GPA; |
| 280 | } |
| 281 | |
| 282 | if (prot & PROT_WRITE) { |
| 283 | mask |= GTF_readonly; |
| 284 | } |
| 285 | |
| 286 | gnt_p = &s->entries.v1[ref]; |
| 287 | |
| 288 | /* |
| 289 | * The guest can legitimately be changing the GTF_readonly flag. Allow |
| 290 | * that, but don't let a malicious guest cause a livelock. |
| 291 | */ |
| 292 | for (retries = 0; retries < 5; retries++) { |
| 293 | uint16_t new_flags; |
| 294 | |
| 295 | /* Read the entry before an atomic operation on its flags */ |
| 296 | gnt = *(volatile grant_entry_v1_t *)gnt_p; |
| 297 | |
| 298 | if ((gnt.flags & mask) != GTF_permit_access || |
| 299 | gnt.domid != DOMID_QEMU) { |
| 300 | return INVALID_GPA; |
| 301 | } |
| 302 | |
| 303 | new_flags = gnt.flags | GTF_reading; |
| 304 | if (prot & PROT_WRITE) { |
| 305 | new_flags |= GTF_writing; |
| 306 | } |
| 307 | |
| 308 | if (qatomic_cmpxchg(&gnt_p->flags, gnt.flags, new_flags) == gnt.flags) { |
| 309 | return (uint64_t)gnt.frame << XEN_PAGE_SHIFT; |
| 310 | } |
| 311 | } |
| 312 | |
| 313 | return INVALID_GPA; |
| 314 | } |
| 315 | |
| 316 | struct xengntdev_handle { |
| 317 | GHashTable *active_maps; |
| 318 | }; |
| 319 | |
| 320 | static int xen_be_gnttab_set_max_grants(struct xengntdev_handle *xgt, |
| 321 | uint32_t nr_grants) |
| 322 | { |
| 323 | return 0; |
| 324 | } |
| 325 | |
| 326 | static void *xen_be_gnttab_map_refs(struct xengntdev_handle *xgt, |
| 327 | uint32_t count, uint32_t domid, |
| 328 | uint32_t *refs, int prot) |
| 329 | { |
| 330 | XenGnttabState *s = xen_gnttab_singleton; |
| 331 | struct active_ref *act; |
| 332 | |
| 333 | if (!s) { |
| 334 | errno = ENOTSUP; |
| 335 | return NULL; |
| 336 | } |
| 337 | |
| 338 | if (domid != xen_domid) { |
| 339 | errno = EINVAL; |
| 340 | return NULL; |
| 341 | } |
| 342 | |
| 343 | if (!count || count > 4096) { |
| 344 | errno = EINVAL; |
| 345 | return NULL; |
| 346 | } |
| 347 | |
| 348 | /* |
| 349 | * Making a contiguous mapping from potentially discontiguous grant |
| 350 | * references would be... distinctly non-trivial. We don't support it. |
| 351 | * Even changing the API to return an array of pointers, one per page, |
| 352 | * wouldn't be simple to use in PV backends because some structures |
| 353 | * actually cross page boundaries (e.g. 32-bit blkif_response ring |
| 354 | * entries are 12 bytes). |
| 355 | */ |
| 356 | if (count != 1) { |
| 357 | errno = EINVAL; |
| 358 | return NULL; |
| 359 | } |
| 360 | |
| 361 | QEMU_LOCK_GUARD(&s->gnt_lock); |
| 362 | |
| 363 | act = g_hash_table_lookup(xgt->active_maps, GINT_TO_POINTER(refs[0])); |
| 364 | if (act) { |
| 365 | if ((prot & PROT_WRITE) && !(act->prot & PROT_WRITE)) { |
| 366 | if (gnt_ref(s, refs[0], prot) == INVALID_GPA) { |
| 367 | return NULL; |
| 368 | } |
| 369 | act->prot |= PROT_WRITE; |
| 370 | } |
| 371 | act->refcnt++; |
| 372 | } else { |
| 373 | uint64_t gpa = gnt_ref(s, refs[0], prot); |
| 374 | if (gpa == INVALID_GPA) { |
| 375 | errno = EINVAL; |
| 376 | return NULL; |
| 377 | } |
| 378 | |
| 379 | act = g_new0(struct active_ref, 1); |
| 380 | act->prot = prot; |
| 381 | act->refcnt = 1; |
| 382 | act->mrs = memory_region_find(get_system_memory(), gpa, XEN_PAGE_SIZE); |
| 383 | |
| 384 | if (act->mrs.mr && |
| 385 | !int128_lt(act->mrs.size, int128_make64(XEN_PAGE_SIZE)) && |
| 386 | memory_region_get_ram_addr(act->mrs.mr) != RAM_ADDR_INVALID) { |
| 387 | act->virtaddr = qemu_map_ram_ptr(act->mrs.mr->ram_block, |
| 388 | act->mrs.offset_within_region); |
| 389 | } |
| 390 | if (!act->virtaddr) { |
| 391 | gnt_unref(s, refs[0], &act->mrs, 0); |
| 392 | g_free(act); |
| 393 | errno = EINVAL; |
| 394 | return NULL; |
| 395 | } |
| 396 | |
| 397 | s->map_track[refs[0]]++; |
| 398 | g_hash_table_insert(xgt->active_maps, GINT_TO_POINTER(refs[0]), act); |
| 399 | } |
| 400 | |
| 401 | return act->virtaddr; |
| 402 | } |
| 403 | |
| 404 | static gboolean do_unmap(gpointer key, gpointer value, gpointer user_data) |
| 405 | { |
| 406 | XenGnttabState *s = user_data; |
| 407 | grant_ref_t gref = GPOINTER_TO_INT(key); |
| 408 | struct active_ref *act = value; |
| 409 | |
| 410 | gnt_unref(s, gref, &act->mrs, act->prot); |
| 411 | g_free(act); |
| 412 | return true; |
| 413 | } |
| 414 | |
| 415 | static int xen_be_gnttab_unmap(struct xengntdev_handle *xgt, |
| 416 | void *start_address, uint32_t *refs, |
| 417 | uint32_t count) |
| 418 | { |
| 419 | XenGnttabState *s = xen_gnttab_singleton; |
| 420 | struct active_ref *act; |
| 421 | |
| 422 | if (!s) { |
| 423 | return -ENOTSUP; |
| 424 | } |
| 425 | |
| 426 | if (count != 1) { |
| 427 | return -EINVAL; |
| 428 | } |
| 429 | |
| 430 | QEMU_LOCK_GUARD(&s->gnt_lock); |
| 431 | |
| 432 | act = g_hash_table_lookup(xgt->active_maps, GINT_TO_POINTER(refs[0])); |
| 433 | if (!act) { |
| 434 | return -ENOENT; |
| 435 | } |
| 436 | |
| 437 | if (act->virtaddr != start_address) { |
| 438 | return -EINVAL; |
| 439 | } |
| 440 | |
| 441 | if (!--act->refcnt) { |
| 442 | do_unmap(GINT_TO_POINTER(refs[0]), act, s); |
| 443 | g_hash_table_remove(xgt->active_maps, GINT_TO_POINTER(refs[0])); |
| 444 | } |
| 445 | |
| 446 | return 0; |
| 447 | } |
| 448 | |
| 449 | /* |
| 450 | * This looks a bit like the one for true Xen in xen-operations.c but |
| 451 | * in emulation we don't support multi-page mappings. And under Xen we |
| 452 | * *want* the multi-page mappings so we have fewer bounces through the |
| 453 | * kernel and the hypervisor. So the code paths end up being similar, |
| 454 | * but different. |
| 455 | */ |
| 456 | static int xen_be_gnttab_copy(struct xengntdev_handle *xgt, bool to_domain, |
| 457 | uint32_t domid, XenGrantCopySegment *segs, |
| 458 | uint32_t nr_segs, Error **errp) |
| 459 | { |
| 460 | int prot = to_domain ? PROT_WRITE : PROT_READ; |
| 461 | unsigned int i; |
| 462 | |
| 463 | for (i = 0; i < nr_segs; i++) { |
| 464 | XenGrantCopySegment *seg = &segs[i]; |
| 465 | void *page; |
| 466 | uint32_t ref = to_domain ? seg->dest.foreign.ref : |
| 467 | seg->source.foreign.ref; |
| 468 | |
| 469 | page = xen_be_gnttab_map_refs(xgt, 1, domid, &ref, prot); |
| 470 | if (!page) { |
| 471 | if (errp) { |
| 472 | error_setg_errno(errp, errno, |
| 473 | "xen_be_gnttab_map_refs failed"); |
| 474 | } |
| 475 | return -errno; |
| 476 | } |
| 477 | |
| 478 | if (to_domain) { |
| 479 | memcpy(page + seg->dest.foreign.offset, seg->source.virt, |
| 480 | seg->len); |
| 481 | } else { |
| 482 | memcpy(seg->dest.virt, page + seg->source.foreign.offset, |
| 483 | seg->len); |
| 484 | } |
| 485 | |
| 486 | if (xen_be_gnttab_unmap(xgt, page, &ref, 1)) { |
| 487 | if (errp) { |
| 488 | error_setg_errno(errp, errno, "xen_be_gnttab_unmap failed"); |
| 489 | } |
| 490 | return -errno; |
| 491 | } |
| 492 | } |
| 493 | |
| 494 | return 0; |
| 495 | } |
| 496 | |
| 497 | static struct xengntdev_handle *xen_be_gnttab_open(void) |
| 498 | { |
| 499 | struct xengntdev_handle *xgt = g_new0(struct xengntdev_handle, 1); |
| 500 | |
| 501 | xgt->active_maps = g_hash_table_new(g_direct_hash, g_direct_equal); |
| 502 | return xgt; |
| 503 | } |
| 504 | |
| 505 | static int xen_be_gnttab_close(struct xengntdev_handle *xgt) |
| 506 | { |
| 507 | XenGnttabState *s = xen_gnttab_singleton; |
| 508 | |
| 509 | if (!s) { |
| 510 | return -ENOTSUP; |
| 511 | } |
| 512 | |
| 513 | g_hash_table_foreach_remove(xgt->active_maps, do_unmap, s); |
| 514 | g_hash_table_destroy(xgt->active_maps); |
| 515 | g_free(xgt); |
| 516 | return 0; |
| 517 | } |
| 518 | |
| 519 | static struct gnttab_backend_ops emu_gnttab_backend_ops = { |
| 520 | .open = xen_be_gnttab_open, |
| 521 | .close = xen_be_gnttab_close, |
| 522 | .grant_copy = xen_be_gnttab_copy, |
| 523 | .set_max_grants = xen_be_gnttab_set_max_grants, |
| 524 | .map_refs = xen_be_gnttab_map_refs, |
| 525 | .unmap = xen_be_gnttab_unmap, |
| 526 | }; |
| 527 | |
| 528 | int xen_gnttab_reset(void) |
| 529 | { |
| 530 | XenGnttabState *s = xen_gnttab_singleton; |
| 531 | |
| 532 | if (!s) { |
| 533 | return -ENOTSUP; |
| 534 | } |
| 535 | |
| 536 | QEMU_LOCK_GUARD(&s->gnt_lock); |
| 537 | |
| 538 | s->nr_frames = 0; |
| 539 | |
| 540 | memset(s->entries.v1, 0, XEN_PAGE_SIZE * s->max_frames); |
| 541 | s->entries.v1[GNTTAB_RESERVED_XENSTORE].flags = GTF_permit_access; |
| 542 | s->entries.v1[GNTTAB_RESERVED_XENSTORE].frame = XEN_SPECIAL_PFN(XENSTORE); |
| 543 | |
| 544 | if (xen_primary_console_get_pfn()) { |
| 545 | s->entries.v1[GNTTAB_RESERVED_CONSOLE].flags = GTF_permit_access; |
| 546 | s->entries.v1[GNTTAB_RESERVED_CONSOLE].frame = XEN_SPECIAL_PFN(CONSOLE); |
| 547 | } |
| 548 | |
| 549 | return 0; |
| 550 | } |