Linux Audio

Check our new training course

Embedded Linux training

Mar 10-20, 2025, special US time zones
Register
Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  KVM guest address space mapping code
   4 *
   5 *    Copyright IBM Corp. 2007, 2020
   6 *    Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
   7 *		 David Hildenbrand <david@redhat.com>
   8 *		 Janosch Frank <frankja@linux.vnet.ibm.com>
   9 */
  10
  11#include <linux/kernel.h>
  12#include <linux/pagewalk.h>
  13#include <linux/swap.h>
  14#include <linux/smp.h>
  15#include <linux/spinlock.h>
  16#include <linux/slab.h>
  17#include <linux/swapops.h>
  18#include <linux/ksm.h>
  19#include <linux/mman.h>
  20#include <linux/pgtable.h>
  21
  22#include <asm/pgalloc.h>
  23#include <asm/gmap.h>
 
  24#include <asm/tlb.h>
  25
  26#define GMAP_SHADOW_FAKE_TABLE 1ULL
  27
 
 
 
 
 
 
 
 
 
 
 
  28/**
  29 * gmap_alloc - allocate and initialize a guest address space
  30 * @limit: maximum address of the gmap address space
  31 *
  32 * Returns a guest address space structure.
  33 */
  34static struct gmap *gmap_alloc(unsigned long limit)
  35{
  36	struct gmap *gmap;
  37	struct page *page;
  38	unsigned long *table;
  39	unsigned long etype, atype;
  40
  41	if (limit < _REGION3_SIZE) {
  42		limit = _REGION3_SIZE - 1;
  43		atype = _ASCE_TYPE_SEGMENT;
  44		etype = _SEGMENT_ENTRY_EMPTY;
  45	} else if (limit < _REGION2_SIZE) {
  46		limit = _REGION2_SIZE - 1;
  47		atype = _ASCE_TYPE_REGION3;
  48		etype = _REGION3_ENTRY_EMPTY;
  49	} else if (limit < _REGION1_SIZE) {
  50		limit = _REGION1_SIZE - 1;
  51		atype = _ASCE_TYPE_REGION2;
  52		etype = _REGION2_ENTRY_EMPTY;
  53	} else {
  54		limit = -1UL;
  55		atype = _ASCE_TYPE_REGION1;
  56		etype = _REGION1_ENTRY_EMPTY;
  57	}
  58	gmap = kzalloc(sizeof(struct gmap), GFP_KERNEL_ACCOUNT);
  59	if (!gmap)
  60		goto out;
  61	INIT_LIST_HEAD(&gmap->crst_list);
  62	INIT_LIST_HEAD(&gmap->children);
  63	INIT_LIST_HEAD(&gmap->pt_list);
  64	INIT_RADIX_TREE(&gmap->guest_to_host, GFP_KERNEL_ACCOUNT);
  65	INIT_RADIX_TREE(&gmap->host_to_guest, GFP_ATOMIC | __GFP_ACCOUNT);
  66	INIT_RADIX_TREE(&gmap->host_to_rmap, GFP_ATOMIC | __GFP_ACCOUNT);
  67	spin_lock_init(&gmap->guest_table_lock);
  68	spin_lock_init(&gmap->shadow_lock);
  69	refcount_set(&gmap->ref_count, 1);
  70	page = alloc_pages(GFP_KERNEL_ACCOUNT, CRST_ALLOC_ORDER);
  71	if (!page)
  72		goto out_free;
  73	page->index = 0;
  74	list_add(&page->lru, &gmap->crst_list);
  75	table = page_to_virt(page);
  76	crst_table_init(table, etype);
  77	gmap->table = table;
  78	gmap->asce = atype | _ASCE_TABLE_LENGTH |
  79		_ASCE_USER_BITS | __pa(table);
  80	gmap->asce_end = limit;
  81	return gmap;
  82
  83out_free:
  84	kfree(gmap);
  85out:
  86	return NULL;
  87}
  88
  89/**
  90 * gmap_create - create a guest address space
  91 * @mm: pointer to the parent mm_struct
  92 * @limit: maximum size of the gmap address space
  93 *
  94 * Returns a guest address space structure.
  95 */
  96struct gmap *gmap_create(struct mm_struct *mm, unsigned long limit)
  97{
  98	struct gmap *gmap;
  99	unsigned long gmap_asce;
 100
 101	gmap = gmap_alloc(limit);
 102	if (!gmap)
 103		return NULL;
 104	gmap->mm = mm;
 105	spin_lock(&mm->context.lock);
 106	list_add_rcu(&gmap->list, &mm->context.gmap_list);
 107	if (list_is_singular(&mm->context.gmap_list))
 108		gmap_asce = gmap->asce;
 109	else
 110		gmap_asce = -1UL;
 111	WRITE_ONCE(mm->context.gmap_asce, gmap_asce);
 112	spin_unlock(&mm->context.lock);
 113	return gmap;
 114}
 115EXPORT_SYMBOL_GPL(gmap_create);
 116
 117static void gmap_flush_tlb(struct gmap *gmap)
 118{
 119	if (MACHINE_HAS_IDTE)
 120		__tlb_flush_idte(gmap->asce);
 121	else
 122		__tlb_flush_global();
 123}
 124
 125static void gmap_radix_tree_free(struct radix_tree_root *root)
 126{
 127	struct radix_tree_iter iter;
 128	unsigned long indices[16];
 129	unsigned long index;
 130	void __rcu **slot;
 131	int i, nr;
 132
 133	/* A radix tree is freed by deleting all of its entries */
 134	index = 0;
 135	do {
 136		nr = 0;
 137		radix_tree_for_each_slot(slot, root, &iter, index) {
 138			indices[nr] = iter.index;
 139			if (++nr == 16)
 140				break;
 141		}
 142		for (i = 0; i < nr; i++) {
 143			index = indices[i];
 144			radix_tree_delete(root, index);
 145		}
 146	} while (nr > 0);
 147}
 148
 149static void gmap_rmap_radix_tree_free(struct radix_tree_root *root)
 150{
 151	struct gmap_rmap *rmap, *rnext, *head;
 152	struct radix_tree_iter iter;
 153	unsigned long indices[16];
 154	unsigned long index;
 155	void __rcu **slot;
 156	int i, nr;
 157
 158	/* A radix tree is freed by deleting all of its entries */
 159	index = 0;
 160	do {
 161		nr = 0;
 162		radix_tree_for_each_slot(slot, root, &iter, index) {
 163			indices[nr] = iter.index;
 164			if (++nr == 16)
 165				break;
 166		}
 167		for (i = 0; i < nr; i++) {
 168			index = indices[i];
 169			head = radix_tree_delete(root, index);
 170			gmap_for_each_rmap_safe(rmap, rnext, head)
 171				kfree(rmap);
 172		}
 173	} while (nr > 0);
 174}
 175
 176/**
 177 * gmap_free - free a guest address space
 178 * @gmap: pointer to the guest address space structure
 179 *
 180 * No locks required. There are no references to this gmap anymore.
 181 */
 182static void gmap_free(struct gmap *gmap)
 183{
 184	struct page *page, *next;
 185
 186	/* Flush tlb of all gmaps (if not already done for shadows) */
 187	if (!(gmap_is_shadow(gmap) && gmap->removed))
 188		gmap_flush_tlb(gmap);
 189	/* Free all segment & region tables. */
 190	list_for_each_entry_safe(page, next, &gmap->crst_list, lru)
 191		__free_pages(page, CRST_ALLOC_ORDER);
 192	gmap_radix_tree_free(&gmap->guest_to_host);
 193	gmap_radix_tree_free(&gmap->host_to_guest);
 194
 195	/* Free additional data for a shadow gmap */
 196	if (gmap_is_shadow(gmap)) {
 
 
 197		/* Free all page tables. */
 198		list_for_each_entry_safe(page, next, &gmap->pt_list, lru)
 199			page_table_free_pgste(page);
 200		gmap_rmap_radix_tree_free(&gmap->host_to_rmap);
 201		/* Release reference to the parent */
 202		gmap_put(gmap->parent);
 203	}
 204
 205	kfree(gmap);
 206}
 207
 208/**
 209 * gmap_get - increase reference counter for guest address space
 210 * @gmap: pointer to the guest address space structure
 211 *
 212 * Returns the gmap pointer
 213 */
 214struct gmap *gmap_get(struct gmap *gmap)
 215{
 216	refcount_inc(&gmap->ref_count);
 217	return gmap;
 218}
 219EXPORT_SYMBOL_GPL(gmap_get);
 220
 221/**
 222 * gmap_put - decrease reference counter for guest address space
 223 * @gmap: pointer to the guest address space structure
 224 *
 225 * If the reference counter reaches zero the guest address space is freed.
 226 */
 227void gmap_put(struct gmap *gmap)
 228{
 229	if (refcount_dec_and_test(&gmap->ref_count))
 230		gmap_free(gmap);
 231}
 232EXPORT_SYMBOL_GPL(gmap_put);
 233
 234/**
 235 * gmap_remove - remove a guest address space but do not free it yet
 236 * @gmap: pointer to the guest address space structure
 237 */
 238void gmap_remove(struct gmap *gmap)
 239{
 240	struct gmap *sg, *next;
 241	unsigned long gmap_asce;
 242
 243	/* Remove all shadow gmaps linked to this gmap */
 244	if (!list_empty(&gmap->children)) {
 245		spin_lock(&gmap->shadow_lock);
 246		list_for_each_entry_safe(sg, next, &gmap->children, list) {
 247			list_del(&sg->list);
 248			gmap_put(sg);
 249		}
 250		spin_unlock(&gmap->shadow_lock);
 251	}
 252	/* Remove gmap from the pre-mm list */
 253	spin_lock(&gmap->mm->context.lock);
 254	list_del_rcu(&gmap->list);
 255	if (list_empty(&gmap->mm->context.gmap_list))
 256		gmap_asce = 0;
 257	else if (list_is_singular(&gmap->mm->context.gmap_list))
 258		gmap_asce = list_first_entry(&gmap->mm->context.gmap_list,
 259					     struct gmap, list)->asce;
 260	else
 261		gmap_asce = -1UL;
 262	WRITE_ONCE(gmap->mm->context.gmap_asce, gmap_asce);
 263	spin_unlock(&gmap->mm->context.lock);
 264	synchronize_rcu();
 265	/* Put reference */
 266	gmap_put(gmap);
 267}
 268EXPORT_SYMBOL_GPL(gmap_remove);
 269
 270/**
 271 * gmap_enable - switch primary space to the guest address space
 272 * @gmap: pointer to the guest address space structure
 273 */
 274void gmap_enable(struct gmap *gmap)
 275{
 276	S390_lowcore.gmap = (unsigned long) gmap;
 277}
 278EXPORT_SYMBOL_GPL(gmap_enable);
 279
 280/**
 281 * gmap_disable - switch back to the standard primary address space
 282 * @gmap: pointer to the guest address space structure
 283 */
 284void gmap_disable(struct gmap *gmap)
 285{
 286	S390_lowcore.gmap = 0UL;
 287}
 288EXPORT_SYMBOL_GPL(gmap_disable);
 289
 290/**
 291 * gmap_get_enabled - get a pointer to the currently enabled gmap
 292 *
 293 * Returns a pointer to the currently enabled gmap. 0 if none is enabled.
 294 */
 295struct gmap *gmap_get_enabled(void)
 296{
 297	return (struct gmap *) S390_lowcore.gmap;
 298}
 299EXPORT_SYMBOL_GPL(gmap_get_enabled);
 300
 301/*
 302 * gmap_alloc_table is assumed to be called with mmap_lock held
 303 */
 304static int gmap_alloc_table(struct gmap *gmap, unsigned long *table,
 305			    unsigned long init, unsigned long gaddr)
 306{
 307	struct page *page;
 308	unsigned long *new;
 309
 310	/* since we dont free the gmap table until gmap_free we can unlock */
 311	page = alloc_pages(GFP_KERNEL_ACCOUNT, CRST_ALLOC_ORDER);
 312	if (!page)
 313		return -ENOMEM;
 314	new = page_to_virt(page);
 315	crst_table_init(new, init);
 316	spin_lock(&gmap->guest_table_lock);
 317	if (*table & _REGION_ENTRY_INVALID) {
 318		list_add(&page->lru, &gmap->crst_list);
 319		*table = __pa(new) | _REGION_ENTRY_LENGTH |
 320			(*table & _REGION_ENTRY_TYPE_MASK);
 321		page->index = gaddr;
 322		page = NULL;
 323	}
 324	spin_unlock(&gmap->guest_table_lock);
 325	if (page)
 326		__free_pages(page, CRST_ALLOC_ORDER);
 327	return 0;
 328}
 329
 330/**
 331 * __gmap_segment_gaddr - find virtual address from segment pointer
 332 * @entry: pointer to a segment table entry in the guest address space
 333 *
 334 * Returns the virtual address in the guest address space for the segment
 335 */
 336static unsigned long __gmap_segment_gaddr(unsigned long *entry)
 337{
 338	struct page *page;
 339	unsigned long offset;
 340
 341	offset = (unsigned long) entry / sizeof(unsigned long);
 342	offset = (offset & (PTRS_PER_PMD - 1)) * PMD_SIZE;
 343	page = pmd_pgtable_page((pmd_t *) entry);
 344	return page->index + offset;
 345}
 346
 347/**
 348 * __gmap_unlink_by_vmaddr - unlink a single segment via a host address
 349 * @gmap: pointer to the guest address space structure
 350 * @vmaddr: address in the host process address space
 351 *
 352 * Returns 1 if a TLB flush is required
 353 */
 354static int __gmap_unlink_by_vmaddr(struct gmap *gmap, unsigned long vmaddr)
 355{
 356	unsigned long *entry;
 357	int flush = 0;
 358
 359	BUG_ON(gmap_is_shadow(gmap));
 360	spin_lock(&gmap->guest_table_lock);
 361	entry = radix_tree_delete(&gmap->host_to_guest, vmaddr >> PMD_SHIFT);
 362	if (entry) {
 363		flush = (*entry != _SEGMENT_ENTRY_EMPTY);
 364		*entry = _SEGMENT_ENTRY_EMPTY;
 365	}
 366	spin_unlock(&gmap->guest_table_lock);
 367	return flush;
 368}
 369
 370/**
 371 * __gmap_unmap_by_gaddr - unmap a single segment via a guest address
 372 * @gmap: pointer to the guest address space structure
 373 * @gaddr: address in the guest address space
 374 *
 375 * Returns 1 if a TLB flush is required
 376 */
 377static int __gmap_unmap_by_gaddr(struct gmap *gmap, unsigned long gaddr)
 378{
 379	unsigned long vmaddr;
 380
 381	vmaddr = (unsigned long) radix_tree_delete(&gmap->guest_to_host,
 382						   gaddr >> PMD_SHIFT);
 383	return vmaddr ? __gmap_unlink_by_vmaddr(gmap, vmaddr) : 0;
 384}
 385
 386/**
 387 * gmap_unmap_segment - unmap segment from the guest address space
 388 * @gmap: pointer to the guest address space structure
 389 * @to: address in the guest address space
 390 * @len: length of the memory area to unmap
 391 *
 392 * Returns 0 if the unmap succeeded, -EINVAL if not.
 393 */
 394int gmap_unmap_segment(struct gmap *gmap, unsigned long to, unsigned long len)
 395{
 396	unsigned long off;
 397	int flush;
 398
 399	BUG_ON(gmap_is_shadow(gmap));
 400	if ((to | len) & (PMD_SIZE - 1))
 401		return -EINVAL;
 402	if (len == 0 || to + len < to)
 403		return -EINVAL;
 404
 405	flush = 0;
 406	mmap_write_lock(gmap->mm);
 407	for (off = 0; off < len; off += PMD_SIZE)
 408		flush |= __gmap_unmap_by_gaddr(gmap, to + off);
 409	mmap_write_unlock(gmap->mm);
 410	if (flush)
 411		gmap_flush_tlb(gmap);
 412	return 0;
 413}
 414EXPORT_SYMBOL_GPL(gmap_unmap_segment);
 415
 416/**
 417 * gmap_map_segment - map a segment to the guest address space
 418 * @gmap: pointer to the guest address space structure
 419 * @from: source address in the parent address space
 420 * @to: target address in the guest address space
 421 * @len: length of the memory area to map
 422 *
 423 * Returns 0 if the mmap succeeded, -EINVAL or -ENOMEM if not.
 424 */
 425int gmap_map_segment(struct gmap *gmap, unsigned long from,
 426		     unsigned long to, unsigned long len)
 427{
 428	unsigned long off;
 429	int flush;
 430
 431	BUG_ON(gmap_is_shadow(gmap));
 432	if ((from | to | len) & (PMD_SIZE - 1))
 433		return -EINVAL;
 434	if (len == 0 || from + len < from || to + len < to ||
 435	    from + len - 1 > TASK_SIZE_MAX || to + len - 1 > gmap->asce_end)
 436		return -EINVAL;
 437
 438	flush = 0;
 439	mmap_write_lock(gmap->mm);
 440	for (off = 0; off < len; off += PMD_SIZE) {
 441		/* Remove old translation */
 442		flush |= __gmap_unmap_by_gaddr(gmap, to + off);
 443		/* Store new translation */
 444		if (radix_tree_insert(&gmap->guest_to_host,
 445				      (to + off) >> PMD_SHIFT,
 446				      (void *) from + off))
 447			break;
 448	}
 449	mmap_write_unlock(gmap->mm);
 450	if (flush)
 451		gmap_flush_tlb(gmap);
 452	if (off >= len)
 453		return 0;
 454	gmap_unmap_segment(gmap, to, len);
 455	return -ENOMEM;
 456}
 457EXPORT_SYMBOL_GPL(gmap_map_segment);
 458
 459/**
 460 * __gmap_translate - translate a guest address to a user space address
 461 * @gmap: pointer to guest mapping meta data structure
 462 * @gaddr: guest address
 463 *
 464 * Returns user space address which corresponds to the guest address or
 465 * -EFAULT if no such mapping exists.
 466 * This function does not establish potentially missing page table entries.
 467 * The mmap_lock of the mm that belongs to the address space must be held
 468 * when this function gets called.
 469 *
 470 * Note: Can also be called for shadow gmaps.
 471 */
 472unsigned long __gmap_translate(struct gmap *gmap, unsigned long gaddr)
 473{
 474	unsigned long vmaddr;
 475
 476	vmaddr = (unsigned long)
 477		radix_tree_lookup(&gmap->guest_to_host, gaddr >> PMD_SHIFT);
 478	/* Note: guest_to_host is empty for a shadow gmap */
 479	return vmaddr ? (vmaddr | (gaddr & ~PMD_MASK)) : -EFAULT;
 480}
 481EXPORT_SYMBOL_GPL(__gmap_translate);
 482
 483/**
 484 * gmap_translate - translate a guest address to a user space address
 485 * @gmap: pointer to guest mapping meta data structure
 486 * @gaddr: guest address
 487 *
 488 * Returns user space address which corresponds to the guest address or
 489 * -EFAULT if no such mapping exists.
 490 * This function does not establish potentially missing page table entries.
 491 */
 492unsigned long gmap_translate(struct gmap *gmap, unsigned long gaddr)
 493{
 494	unsigned long rc;
 495
 496	mmap_read_lock(gmap->mm);
 497	rc = __gmap_translate(gmap, gaddr);
 498	mmap_read_unlock(gmap->mm);
 499	return rc;
 500}
 501EXPORT_SYMBOL_GPL(gmap_translate);
 502
 503/**
 504 * gmap_unlink - disconnect a page table from the gmap shadow tables
 505 * @mm: pointer to the parent mm_struct
 506 * @table: pointer to the host page table
 507 * @vmaddr: vm address associated with the host page table
 508 */
 509void gmap_unlink(struct mm_struct *mm, unsigned long *table,
 510		 unsigned long vmaddr)
 511{
 512	struct gmap *gmap;
 513	int flush;
 514
 515	rcu_read_lock();
 516	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
 517		flush = __gmap_unlink_by_vmaddr(gmap, vmaddr);
 518		if (flush)
 519			gmap_flush_tlb(gmap);
 520	}
 521	rcu_read_unlock();
 522}
 523
 524static void gmap_pmdp_xchg(struct gmap *gmap, pmd_t *old, pmd_t new,
 525			   unsigned long gaddr);
 526
 527/**
 528 * __gmap_link - set up shadow page tables to connect a host to a guest address
 529 * @gmap: pointer to guest mapping meta data structure
 530 * @gaddr: guest address
 531 * @vmaddr: vm address
 532 *
 533 * Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
 534 * if the vm address is already mapped to a different guest segment.
 535 * The mmap_lock of the mm that belongs to the address space must be held
 536 * when this function gets called.
 537 */
 538int __gmap_link(struct gmap *gmap, unsigned long gaddr, unsigned long vmaddr)
 539{
 540	struct mm_struct *mm;
 541	unsigned long *table;
 542	spinlock_t *ptl;
 543	pgd_t *pgd;
 544	p4d_t *p4d;
 545	pud_t *pud;
 546	pmd_t *pmd;
 547	u64 unprot;
 548	int rc;
 549
 550	BUG_ON(gmap_is_shadow(gmap));
 551	/* Create higher level tables in the gmap page table */
 552	table = gmap->table;
 553	if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION1) {
 554		table += (gaddr & _REGION1_INDEX) >> _REGION1_SHIFT;
 555		if ((*table & _REGION_ENTRY_INVALID) &&
 556		    gmap_alloc_table(gmap, table, _REGION2_ENTRY_EMPTY,
 557				     gaddr & _REGION1_MASK))
 558			return -ENOMEM;
 559		table = __va(*table & _REGION_ENTRY_ORIGIN);
 560	}
 561	if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION2) {
 562		table += (gaddr & _REGION2_INDEX) >> _REGION2_SHIFT;
 563		if ((*table & _REGION_ENTRY_INVALID) &&
 564		    gmap_alloc_table(gmap, table, _REGION3_ENTRY_EMPTY,
 565				     gaddr & _REGION2_MASK))
 566			return -ENOMEM;
 567		table = __va(*table & _REGION_ENTRY_ORIGIN);
 568	}
 569	if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION3) {
 570		table += (gaddr & _REGION3_INDEX) >> _REGION3_SHIFT;
 571		if ((*table & _REGION_ENTRY_INVALID) &&
 572		    gmap_alloc_table(gmap, table, _SEGMENT_ENTRY_EMPTY,
 573				     gaddr & _REGION3_MASK))
 574			return -ENOMEM;
 575		table = __va(*table & _REGION_ENTRY_ORIGIN);
 576	}
 577	table += (gaddr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
 578	/* Walk the parent mm page table */
 579	mm = gmap->mm;
 580	pgd = pgd_offset(mm, vmaddr);
 581	VM_BUG_ON(pgd_none(*pgd));
 582	p4d = p4d_offset(pgd, vmaddr);
 583	VM_BUG_ON(p4d_none(*p4d));
 584	pud = pud_offset(p4d, vmaddr);
 585	VM_BUG_ON(pud_none(*pud));
 586	/* large puds cannot yet be handled */
 587	if (pud_large(*pud))
 588		return -EFAULT;
 589	pmd = pmd_offset(pud, vmaddr);
 590	VM_BUG_ON(pmd_none(*pmd));
 591	/* Are we allowed to use huge pages? */
 592	if (pmd_large(*pmd) && !gmap->mm->context.allow_gmap_hpage_1m)
 593		return -EFAULT;
 594	/* Link gmap segment table entry location to page table. */
 595	rc = radix_tree_preload(GFP_KERNEL_ACCOUNT);
 596	if (rc)
 597		return rc;
 598	ptl = pmd_lock(mm, pmd);
 599	spin_lock(&gmap->guest_table_lock);
 600	if (*table == _SEGMENT_ENTRY_EMPTY) {
 601		rc = radix_tree_insert(&gmap->host_to_guest,
 602				       vmaddr >> PMD_SHIFT, table);
 603		if (!rc) {
 604			if (pmd_large(*pmd)) {
 605				*table = (pmd_val(*pmd) &
 606					  _SEGMENT_ENTRY_HARDWARE_BITS_LARGE)
 607					| _SEGMENT_ENTRY_GMAP_UC;
 608			} else
 609				*table = pmd_val(*pmd) &
 610					_SEGMENT_ENTRY_HARDWARE_BITS;
 611		}
 612	} else if (*table & _SEGMENT_ENTRY_PROTECT &&
 613		   !(pmd_val(*pmd) & _SEGMENT_ENTRY_PROTECT)) {
 614		unprot = (u64)*table;
 615		unprot &= ~_SEGMENT_ENTRY_PROTECT;
 616		unprot |= _SEGMENT_ENTRY_GMAP_UC;
 617		gmap_pmdp_xchg(gmap, (pmd_t *)table, __pmd(unprot), gaddr);
 618	}
 619	spin_unlock(&gmap->guest_table_lock);
 620	spin_unlock(ptl);
 621	radix_tree_preload_end();
 622	return rc;
 623}
 624
 625/**
 626 * gmap_fault - resolve a fault on a guest address
 627 * @gmap: pointer to guest mapping meta data structure
 628 * @gaddr: guest address
 629 * @fault_flags: flags to pass down to handle_mm_fault()
 630 *
 631 * Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
 632 * if the vm address is already mapped to a different guest segment.
 633 */
 634int gmap_fault(struct gmap *gmap, unsigned long gaddr,
 635	       unsigned int fault_flags)
 636{
 637	unsigned long vmaddr;
 638	int rc;
 639	bool unlocked;
 640
 641	mmap_read_lock(gmap->mm);
 642
 643retry:
 644	unlocked = false;
 645	vmaddr = __gmap_translate(gmap, gaddr);
 646	if (IS_ERR_VALUE(vmaddr)) {
 647		rc = vmaddr;
 648		goto out_up;
 649	}
 650	if (fixup_user_fault(gmap->mm, vmaddr, fault_flags,
 651			     &unlocked)) {
 652		rc = -EFAULT;
 653		goto out_up;
 654	}
 655	/*
 656	 * In the case that fixup_user_fault unlocked the mmap_lock during
 657	 * faultin redo __gmap_translate to not race with a map/unmap_segment.
 658	 */
 659	if (unlocked)
 660		goto retry;
 661
 662	rc = __gmap_link(gmap, gaddr, vmaddr);
 663out_up:
 664	mmap_read_unlock(gmap->mm);
 665	return rc;
 666}
 667EXPORT_SYMBOL_GPL(gmap_fault);
 668
 669/*
 670 * this function is assumed to be called with mmap_lock held
 671 */
 672void __gmap_zap(struct gmap *gmap, unsigned long gaddr)
 673{
 674	struct vm_area_struct *vma;
 675	unsigned long vmaddr;
 676	spinlock_t *ptl;
 677	pte_t *ptep;
 678
 679	/* Find the vm address for the guest address */
 680	vmaddr = (unsigned long) radix_tree_lookup(&gmap->guest_to_host,
 681						   gaddr >> PMD_SHIFT);
 682	if (vmaddr) {
 683		vmaddr |= gaddr & ~PMD_MASK;
 684
 685		vma = vma_lookup(gmap->mm, vmaddr);
 686		if (!vma || is_vm_hugetlb_page(vma))
 687			return;
 688
 689		/* Get pointer to the page table entry */
 690		ptep = get_locked_pte(gmap->mm, vmaddr, &ptl);
 691		if (likely(ptep)) {
 692			ptep_zap_unused(gmap->mm, vmaddr, ptep, 0);
 693			pte_unmap_unlock(ptep, ptl);
 694		}
 695	}
 696}
 697EXPORT_SYMBOL_GPL(__gmap_zap);
 698
 699void gmap_discard(struct gmap *gmap, unsigned long from, unsigned long to)
 700{
 701	unsigned long gaddr, vmaddr, size;
 702	struct vm_area_struct *vma;
 703
 704	mmap_read_lock(gmap->mm);
 705	for (gaddr = from; gaddr < to;
 706	     gaddr = (gaddr + PMD_SIZE) & PMD_MASK) {
 707		/* Find the vm address for the guest address */
 708		vmaddr = (unsigned long)
 709			radix_tree_lookup(&gmap->guest_to_host,
 710					  gaddr >> PMD_SHIFT);
 711		if (!vmaddr)
 712			continue;
 713		vmaddr |= gaddr & ~PMD_MASK;
 714		/* Find vma in the parent mm */
 715		vma = find_vma(gmap->mm, vmaddr);
 716		if (!vma)
 717			continue;
 718		/*
 719		 * We do not discard pages that are backed by
 720		 * hugetlbfs, so we don't have to refault them.
 721		 */
 722		if (is_vm_hugetlb_page(vma))
 723			continue;
 724		size = min(to - gaddr, PMD_SIZE - (gaddr & ~PMD_MASK));
 725		zap_page_range(vma, vmaddr, size);
 726	}
 727	mmap_read_unlock(gmap->mm);
 728}
 729EXPORT_SYMBOL_GPL(gmap_discard);
 730
 731static LIST_HEAD(gmap_notifier_list);
 732static DEFINE_SPINLOCK(gmap_notifier_lock);
 733
 734/**
 735 * gmap_register_pte_notifier - register a pte invalidation callback
 736 * @nb: pointer to the gmap notifier block
 737 */
 738void gmap_register_pte_notifier(struct gmap_notifier *nb)
 739{
 740	spin_lock(&gmap_notifier_lock);
 741	list_add_rcu(&nb->list, &gmap_notifier_list);
 742	spin_unlock(&gmap_notifier_lock);
 743}
 744EXPORT_SYMBOL_GPL(gmap_register_pte_notifier);
 745
 746/**
 747 * gmap_unregister_pte_notifier - remove a pte invalidation callback
 748 * @nb: pointer to the gmap notifier block
 749 */
 750void gmap_unregister_pte_notifier(struct gmap_notifier *nb)
 751{
 752	spin_lock(&gmap_notifier_lock);
 753	list_del_rcu(&nb->list);
 754	spin_unlock(&gmap_notifier_lock);
 755	synchronize_rcu();
 756}
 757EXPORT_SYMBOL_GPL(gmap_unregister_pte_notifier);
 758
 759/**
 760 * gmap_call_notifier - call all registered invalidation callbacks
 761 * @gmap: pointer to guest mapping meta data structure
 762 * @start: start virtual address in the guest address space
 763 * @end: end virtual address in the guest address space
 764 */
 765static void gmap_call_notifier(struct gmap *gmap, unsigned long start,
 766			       unsigned long end)
 767{
 768	struct gmap_notifier *nb;
 769
 770	list_for_each_entry(nb, &gmap_notifier_list, list)
 771		nb->notifier_call(gmap, start, end);
 772}
 773
 774/**
 775 * gmap_table_walk - walk the gmap page tables
 776 * @gmap: pointer to guest mapping meta data structure
 777 * @gaddr: virtual address in the guest address space
 778 * @level: page table level to stop at
 779 *
 780 * Returns a table entry pointer for the given guest address and @level
 781 * @level=0 : returns a pointer to a page table table entry (or NULL)
 782 * @level=1 : returns a pointer to a segment table entry (or NULL)
 783 * @level=2 : returns a pointer to a region-3 table entry (or NULL)
 784 * @level=3 : returns a pointer to a region-2 table entry (or NULL)
 785 * @level=4 : returns a pointer to a region-1 table entry (or NULL)
 786 *
 787 * Returns NULL if the gmap page tables could not be walked to the
 788 * requested level.
 789 *
 790 * Note: Can also be called for shadow gmaps.
 791 */
 792static inline unsigned long *gmap_table_walk(struct gmap *gmap,
 793					     unsigned long gaddr, int level)
 794{
 795	const int asce_type = gmap->asce & _ASCE_TYPE_MASK;
 796	unsigned long *table = gmap->table;
 797
 798	if (gmap_is_shadow(gmap) && gmap->removed)
 799		return NULL;
 800
 801	if (WARN_ON_ONCE(level > (asce_type >> 2) + 1))
 802		return NULL;
 803
 804	if (asce_type != _ASCE_TYPE_REGION1 &&
 805	    gaddr & (-1UL << (31 + (asce_type >> 2) * 11)))
 806		return NULL;
 807
 808	switch (asce_type) {
 809	case _ASCE_TYPE_REGION1:
 810		table += (gaddr & _REGION1_INDEX) >> _REGION1_SHIFT;
 811		if (level == 4)
 812			break;
 813		if (*table & _REGION_ENTRY_INVALID)
 814			return NULL;
 815		table = __va(*table & _REGION_ENTRY_ORIGIN);
 816		fallthrough;
 817	case _ASCE_TYPE_REGION2:
 818		table += (gaddr & _REGION2_INDEX) >> _REGION2_SHIFT;
 819		if (level == 3)
 820			break;
 821		if (*table & _REGION_ENTRY_INVALID)
 822			return NULL;
 823		table = __va(*table & _REGION_ENTRY_ORIGIN);
 824		fallthrough;
 825	case _ASCE_TYPE_REGION3:
 826		table += (gaddr & _REGION3_INDEX) >> _REGION3_SHIFT;
 827		if (level == 2)
 828			break;
 829		if (*table & _REGION_ENTRY_INVALID)
 830			return NULL;
 831		table = __va(*table & _REGION_ENTRY_ORIGIN);
 832		fallthrough;
 833	case _ASCE_TYPE_SEGMENT:
 834		table += (gaddr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
 835		if (level == 1)
 836			break;
 837		if (*table & _REGION_ENTRY_INVALID)
 838			return NULL;
 839		table = __va(*table & _SEGMENT_ENTRY_ORIGIN);
 840		table += (gaddr & _PAGE_INDEX) >> _PAGE_SHIFT;
 841	}
 842	return table;
 843}
 844
 845/**
 846 * gmap_pte_op_walk - walk the gmap page table, get the page table lock
 847 *		      and return the pte pointer
 848 * @gmap: pointer to guest mapping meta data structure
 849 * @gaddr: virtual address in the guest address space
 850 * @ptl: pointer to the spinlock pointer
 851 *
 852 * Returns a pointer to the locked pte for a guest address, or NULL
 853 */
 854static pte_t *gmap_pte_op_walk(struct gmap *gmap, unsigned long gaddr,
 855			       spinlock_t **ptl)
 856{
 857	unsigned long *table;
 858
 859	BUG_ON(gmap_is_shadow(gmap));
 860	/* Walk the gmap page table, lock and get pte pointer */
 861	table = gmap_table_walk(gmap, gaddr, 1); /* get segment pointer */
 862	if (!table || *table & _SEGMENT_ENTRY_INVALID)
 863		return NULL;
 864	return pte_alloc_map_lock(gmap->mm, (pmd_t *) table, gaddr, ptl);
 865}
 866
 867/**
 868 * gmap_pte_op_fixup - force a page in and connect the gmap page table
 869 * @gmap: pointer to guest mapping meta data structure
 870 * @gaddr: virtual address in the guest address space
 871 * @vmaddr: address in the host process address space
 872 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
 873 *
 874 * Returns 0 if the caller can retry __gmap_translate (might fail again),
 875 * -ENOMEM if out of memory and -EFAULT if anything goes wrong while fixing
 876 * up or connecting the gmap page table.
 877 */
 878static int gmap_pte_op_fixup(struct gmap *gmap, unsigned long gaddr,
 879			     unsigned long vmaddr, int prot)
 880{
 881	struct mm_struct *mm = gmap->mm;
 882	unsigned int fault_flags;
 883	bool unlocked = false;
 884
 885	BUG_ON(gmap_is_shadow(gmap));
 886	fault_flags = (prot == PROT_WRITE) ? FAULT_FLAG_WRITE : 0;
 887	if (fixup_user_fault(mm, vmaddr, fault_flags, &unlocked))
 888		return -EFAULT;
 889	if (unlocked)
 890		/* lost mmap_lock, caller has to retry __gmap_translate */
 891		return 0;
 892	/* Connect the page tables */
 893	return __gmap_link(gmap, gaddr, vmaddr);
 894}
 895
 896/**
 897 * gmap_pte_op_end - release the page table lock
 898 * @ptl: pointer to the spinlock pointer
 
 899 */
 900static void gmap_pte_op_end(spinlock_t *ptl)
 901{
 902	if (ptl)
 903		spin_unlock(ptl);
 904}
 905
 906/**
 907 * gmap_pmd_op_walk - walk the gmap tables, get the guest table lock
 908 *		      and return the pmd pointer
 909 * @gmap: pointer to guest mapping meta data structure
 910 * @gaddr: virtual address in the guest address space
 911 *
 912 * Returns a pointer to the pmd for a guest address, or NULL
 913 */
 914static inline pmd_t *gmap_pmd_op_walk(struct gmap *gmap, unsigned long gaddr)
 915{
 916	pmd_t *pmdp;
 917
 918	BUG_ON(gmap_is_shadow(gmap));
 919	pmdp = (pmd_t *) gmap_table_walk(gmap, gaddr, 1);
 920	if (!pmdp)
 921		return NULL;
 922
 923	/* without huge pages, there is no need to take the table lock */
 924	if (!gmap->mm->context.allow_gmap_hpage_1m)
 925		return pmd_none(*pmdp) ? NULL : pmdp;
 926
 927	spin_lock(&gmap->guest_table_lock);
 928	if (pmd_none(*pmdp)) {
 929		spin_unlock(&gmap->guest_table_lock);
 930		return NULL;
 931	}
 932
 933	/* 4k page table entries are locked via the pte (pte_alloc_map_lock). */
 934	if (!pmd_large(*pmdp))
 935		spin_unlock(&gmap->guest_table_lock);
 936	return pmdp;
 937}
 938
 939/**
 940 * gmap_pmd_op_end - release the guest_table_lock if needed
 941 * @gmap: pointer to the guest mapping meta data structure
 942 * @pmdp: pointer to the pmd
 943 */
 944static inline void gmap_pmd_op_end(struct gmap *gmap, pmd_t *pmdp)
 945{
 946	if (pmd_large(*pmdp))
 947		spin_unlock(&gmap->guest_table_lock);
 948}
 949
 950/*
 951 * gmap_protect_pmd - remove access rights to memory and set pmd notification bits
 952 * @pmdp: pointer to the pmd to be protected
 953 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
 954 * @bits: notification bits to set
 955 *
 956 * Returns:
 957 * 0 if successfully protected
 958 * -EAGAIN if a fixup is needed
 959 * -EINVAL if unsupported notifier bits have been specified
 960 *
 961 * Expected to be called with sg->mm->mmap_lock in read and
 962 * guest_table_lock held.
 963 */
 964static int gmap_protect_pmd(struct gmap *gmap, unsigned long gaddr,
 965			    pmd_t *pmdp, int prot, unsigned long bits)
 966{
 967	int pmd_i = pmd_val(*pmdp) & _SEGMENT_ENTRY_INVALID;
 968	int pmd_p = pmd_val(*pmdp) & _SEGMENT_ENTRY_PROTECT;
 969	pmd_t new = *pmdp;
 970
 971	/* Fixup needed */
 972	if ((pmd_i && (prot != PROT_NONE)) || (pmd_p && (prot == PROT_WRITE)))
 973		return -EAGAIN;
 974
 975	if (prot == PROT_NONE && !pmd_i) {
 976		new = set_pmd_bit(new, __pgprot(_SEGMENT_ENTRY_INVALID));
 977		gmap_pmdp_xchg(gmap, pmdp, new, gaddr);
 978	}
 979
 980	if (prot == PROT_READ && !pmd_p) {
 981		new = clear_pmd_bit(new, __pgprot(_SEGMENT_ENTRY_INVALID));
 982		new = set_pmd_bit(new, __pgprot(_SEGMENT_ENTRY_PROTECT));
 983		gmap_pmdp_xchg(gmap, pmdp, new, gaddr);
 984	}
 985
 986	if (bits & GMAP_NOTIFY_MPROT)
 987		set_pmd(pmdp, set_pmd_bit(*pmdp, __pgprot(_SEGMENT_ENTRY_GMAP_IN)));
 988
 989	/* Shadow GMAP protection needs split PMDs */
 990	if (bits & GMAP_NOTIFY_SHADOW)
 991		return -EINVAL;
 992
 993	return 0;
 994}
 995
 996/*
 997 * gmap_protect_pte - remove access rights to memory and set pgste bits
 998 * @gmap: pointer to guest mapping meta data structure
 999 * @gaddr: virtual address in the guest address space
1000 * @pmdp: pointer to the pmd associated with the pte
1001 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
1002 * @bits: notification bits to set
1003 *
1004 * Returns 0 if successfully protected, -ENOMEM if out of memory and
1005 * -EAGAIN if a fixup is needed.
1006 *
1007 * Expected to be called with sg->mm->mmap_lock in read
1008 */
1009static int gmap_protect_pte(struct gmap *gmap, unsigned long gaddr,
1010			    pmd_t *pmdp, int prot, unsigned long bits)
1011{
1012	int rc;
1013	pte_t *ptep;
1014	spinlock_t *ptl = NULL;
1015	unsigned long pbits = 0;
1016
1017	if (pmd_val(*pmdp) & _SEGMENT_ENTRY_INVALID)
1018		return -EAGAIN;
1019
1020	ptep = pte_alloc_map_lock(gmap->mm, pmdp, gaddr, &ptl);
1021	if (!ptep)
1022		return -ENOMEM;
1023
1024	pbits |= (bits & GMAP_NOTIFY_MPROT) ? PGSTE_IN_BIT : 0;
1025	pbits |= (bits & GMAP_NOTIFY_SHADOW) ? PGSTE_VSIE_BIT : 0;
1026	/* Protect and unlock. */
1027	rc = ptep_force_prot(gmap->mm, gaddr, ptep, prot, pbits);
1028	gmap_pte_op_end(ptl);
1029	return rc;
1030}
1031
1032/*
1033 * gmap_protect_range - remove access rights to memory and set pgste bits
1034 * @gmap: pointer to guest mapping meta data structure
1035 * @gaddr: virtual address in the guest address space
1036 * @len: size of area
1037 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
1038 * @bits: pgste notification bits to set
1039 *
1040 * Returns 0 if successfully protected, -ENOMEM if out of memory and
1041 * -EFAULT if gaddr is invalid (or mapping for shadows is missing).
1042 *
1043 * Called with sg->mm->mmap_lock in read.
1044 */
1045static int gmap_protect_range(struct gmap *gmap, unsigned long gaddr,
1046			      unsigned long len, int prot, unsigned long bits)
1047{
1048	unsigned long vmaddr, dist;
1049	pmd_t *pmdp;
1050	int rc;
1051
1052	BUG_ON(gmap_is_shadow(gmap));
1053	while (len) {
1054		rc = -EAGAIN;
1055		pmdp = gmap_pmd_op_walk(gmap, gaddr);
1056		if (pmdp) {
1057			if (!pmd_large(*pmdp)) {
1058				rc = gmap_protect_pte(gmap, gaddr, pmdp, prot,
1059						      bits);
1060				if (!rc) {
1061					len -= PAGE_SIZE;
1062					gaddr += PAGE_SIZE;
1063				}
1064			} else {
1065				rc = gmap_protect_pmd(gmap, gaddr, pmdp, prot,
1066						      bits);
1067				if (!rc) {
1068					dist = HPAGE_SIZE - (gaddr & ~HPAGE_MASK);
1069					len = len < dist ? 0 : len - dist;
1070					gaddr = (gaddr & HPAGE_MASK) + HPAGE_SIZE;
1071				}
1072			}
1073			gmap_pmd_op_end(gmap, pmdp);
1074		}
1075		if (rc) {
1076			if (rc == -EINVAL)
1077				return rc;
1078
1079			/* -EAGAIN, fixup of userspace mm and gmap */
1080			vmaddr = __gmap_translate(gmap, gaddr);
1081			if (IS_ERR_VALUE(vmaddr))
1082				return vmaddr;
1083			rc = gmap_pte_op_fixup(gmap, gaddr, vmaddr, prot);
1084			if (rc)
1085				return rc;
1086		}
1087	}
1088	return 0;
1089}
1090
1091/**
1092 * gmap_mprotect_notify - change access rights for a range of ptes and
1093 *                        call the notifier if any pte changes again
1094 * @gmap: pointer to guest mapping meta data structure
1095 * @gaddr: virtual address in the guest address space
1096 * @len: size of area
1097 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
1098 *
1099 * Returns 0 if for each page in the given range a gmap mapping exists,
1100 * the new access rights could be set and the notifier could be armed.
1101 * If the gmap mapping is missing for one or more pages -EFAULT is
1102 * returned. If no memory could be allocated -ENOMEM is returned.
1103 * This function establishes missing page table entries.
1104 */
1105int gmap_mprotect_notify(struct gmap *gmap, unsigned long gaddr,
1106			 unsigned long len, int prot)
1107{
1108	int rc;
1109
1110	if ((gaddr & ~PAGE_MASK) || (len & ~PAGE_MASK) || gmap_is_shadow(gmap))
1111		return -EINVAL;
1112	if (!MACHINE_HAS_ESOP && prot == PROT_READ)
1113		return -EINVAL;
1114	mmap_read_lock(gmap->mm);
1115	rc = gmap_protect_range(gmap, gaddr, len, prot, GMAP_NOTIFY_MPROT);
1116	mmap_read_unlock(gmap->mm);
1117	return rc;
1118}
1119EXPORT_SYMBOL_GPL(gmap_mprotect_notify);
1120
1121/**
1122 * gmap_read_table - get an unsigned long value from a guest page table using
1123 *                   absolute addressing, without marking the page referenced.
1124 * @gmap: pointer to guest mapping meta data structure
1125 * @gaddr: virtual address in the guest address space
1126 * @val: pointer to the unsigned long value to return
1127 *
1128 * Returns 0 if the value was read, -ENOMEM if out of memory and -EFAULT
1129 * if reading using the virtual address failed. -EINVAL if called on a gmap
1130 * shadow.
1131 *
1132 * Called with gmap->mm->mmap_lock in read.
1133 */
1134int gmap_read_table(struct gmap *gmap, unsigned long gaddr, unsigned long *val)
1135{
1136	unsigned long address, vmaddr;
1137	spinlock_t *ptl;
1138	pte_t *ptep, pte;
1139	int rc;
1140
1141	if (gmap_is_shadow(gmap))
1142		return -EINVAL;
1143
1144	while (1) {
1145		rc = -EAGAIN;
1146		ptep = gmap_pte_op_walk(gmap, gaddr, &ptl);
1147		if (ptep) {
1148			pte = *ptep;
1149			if (pte_present(pte) && (pte_val(pte) & _PAGE_READ)) {
1150				address = pte_val(pte) & PAGE_MASK;
1151				address += gaddr & ~PAGE_MASK;
1152				*val = *(unsigned long *)__va(address);
1153				set_pte(ptep, set_pte_bit(*ptep, __pgprot(_PAGE_YOUNG)));
1154				/* Do *NOT* clear the _PAGE_INVALID bit! */
1155				rc = 0;
1156			}
1157			gmap_pte_op_end(ptl);
1158		}
1159		if (!rc)
1160			break;
1161		vmaddr = __gmap_translate(gmap, gaddr);
1162		if (IS_ERR_VALUE(vmaddr)) {
1163			rc = vmaddr;
1164			break;
1165		}
1166		rc = gmap_pte_op_fixup(gmap, gaddr, vmaddr, PROT_READ);
1167		if (rc)
1168			break;
1169	}
1170	return rc;
1171}
1172EXPORT_SYMBOL_GPL(gmap_read_table);
1173
1174/**
1175 * gmap_insert_rmap - add a rmap to the host_to_rmap radix tree
1176 * @sg: pointer to the shadow guest address space structure
1177 * @vmaddr: vm address associated with the rmap
1178 * @rmap: pointer to the rmap structure
1179 *
1180 * Called with the sg->guest_table_lock
1181 */
1182static inline void gmap_insert_rmap(struct gmap *sg, unsigned long vmaddr,
1183				    struct gmap_rmap *rmap)
1184{
1185	struct gmap_rmap *temp;
1186	void __rcu **slot;
1187
1188	BUG_ON(!gmap_is_shadow(sg));
1189	slot = radix_tree_lookup_slot(&sg->host_to_rmap, vmaddr >> PAGE_SHIFT);
1190	if (slot) {
1191		rmap->next = radix_tree_deref_slot_protected(slot,
1192							&sg->guest_table_lock);
1193		for (temp = rmap->next; temp; temp = temp->next) {
1194			if (temp->raddr == rmap->raddr) {
1195				kfree(rmap);
1196				return;
1197			}
1198		}
1199		radix_tree_replace_slot(&sg->host_to_rmap, slot, rmap);
1200	} else {
1201		rmap->next = NULL;
1202		radix_tree_insert(&sg->host_to_rmap, vmaddr >> PAGE_SHIFT,
1203				  rmap);
1204	}
1205}
1206
1207/**
1208 * gmap_protect_rmap - restrict access rights to memory (RO) and create an rmap
1209 * @sg: pointer to the shadow guest address space structure
1210 * @raddr: rmap address in the shadow gmap
1211 * @paddr: address in the parent guest address space
1212 * @len: length of the memory area to protect
1213 *
1214 * Returns 0 if successfully protected and the rmap was created, -ENOMEM
1215 * if out of memory and -EFAULT if paddr is invalid.
1216 */
1217static int gmap_protect_rmap(struct gmap *sg, unsigned long raddr,
1218			     unsigned long paddr, unsigned long len)
1219{
1220	struct gmap *parent;
1221	struct gmap_rmap *rmap;
1222	unsigned long vmaddr;
1223	spinlock_t *ptl;
1224	pte_t *ptep;
1225	int rc;
1226
1227	BUG_ON(!gmap_is_shadow(sg));
1228	parent = sg->parent;
1229	while (len) {
1230		vmaddr = __gmap_translate(parent, paddr);
1231		if (IS_ERR_VALUE(vmaddr))
1232			return vmaddr;
1233		rmap = kzalloc(sizeof(*rmap), GFP_KERNEL_ACCOUNT);
1234		if (!rmap)
1235			return -ENOMEM;
1236		rmap->raddr = raddr;
1237		rc = radix_tree_preload(GFP_KERNEL_ACCOUNT);
1238		if (rc) {
1239			kfree(rmap);
1240			return rc;
1241		}
1242		rc = -EAGAIN;
1243		ptep = gmap_pte_op_walk(parent, paddr, &ptl);
1244		if (ptep) {
1245			spin_lock(&sg->guest_table_lock);
1246			rc = ptep_force_prot(parent->mm, paddr, ptep, PROT_READ,
1247					     PGSTE_VSIE_BIT);
1248			if (!rc)
1249				gmap_insert_rmap(sg, vmaddr, rmap);
1250			spin_unlock(&sg->guest_table_lock);
1251			gmap_pte_op_end(ptl);
1252		}
1253		radix_tree_preload_end();
1254		if (rc) {
1255			kfree(rmap);
1256			rc = gmap_pte_op_fixup(parent, paddr, vmaddr, PROT_READ);
1257			if (rc)
1258				return rc;
1259			continue;
1260		}
1261		paddr += PAGE_SIZE;
1262		len -= PAGE_SIZE;
1263	}
1264	return 0;
1265}
1266
1267#define _SHADOW_RMAP_MASK	0x7
1268#define _SHADOW_RMAP_REGION1	0x5
1269#define _SHADOW_RMAP_REGION2	0x4
1270#define _SHADOW_RMAP_REGION3	0x3
1271#define _SHADOW_RMAP_SEGMENT	0x2
1272#define _SHADOW_RMAP_PGTABLE	0x1
1273
1274/**
1275 * gmap_idte_one - invalidate a single region or segment table entry
1276 * @asce: region or segment table *origin* + table-type bits
1277 * @vaddr: virtual address to identify the table entry to flush
1278 *
1279 * The invalid bit of a single region or segment table entry is set
1280 * and the associated TLB entries depending on the entry are flushed.
1281 * The table-type of the @asce identifies the portion of the @vaddr
1282 * that is used as the invalidation index.
1283 */
1284static inline void gmap_idte_one(unsigned long asce, unsigned long vaddr)
1285{
1286	asm volatile(
1287		"	idte	%0,0,%1"
1288		: : "a" (asce), "a" (vaddr) : "cc", "memory");
1289}
1290
1291/**
1292 * gmap_unshadow_page - remove a page from a shadow page table
1293 * @sg: pointer to the shadow guest address space structure
1294 * @raddr: rmap address in the shadow guest address space
1295 *
1296 * Called with the sg->guest_table_lock
1297 */
1298static void gmap_unshadow_page(struct gmap *sg, unsigned long raddr)
1299{
1300	unsigned long *table;
1301
1302	BUG_ON(!gmap_is_shadow(sg));
1303	table = gmap_table_walk(sg, raddr, 0); /* get page table pointer */
1304	if (!table || *table & _PAGE_INVALID)
1305		return;
1306	gmap_call_notifier(sg, raddr, raddr + _PAGE_SIZE - 1);
1307	ptep_unshadow_pte(sg->mm, raddr, (pte_t *) table);
1308}
1309
1310/**
1311 * __gmap_unshadow_pgt - remove all entries from a shadow page table
1312 * @sg: pointer to the shadow guest address space structure
1313 * @raddr: rmap address in the shadow guest address space
1314 * @pgt: pointer to the start of a shadow page table
1315 *
1316 * Called with the sg->guest_table_lock
1317 */
1318static void __gmap_unshadow_pgt(struct gmap *sg, unsigned long raddr,
1319				unsigned long *pgt)
1320{
1321	int i;
1322
1323	BUG_ON(!gmap_is_shadow(sg));
1324	for (i = 0; i < _PAGE_ENTRIES; i++, raddr += _PAGE_SIZE)
1325		pgt[i] = _PAGE_INVALID;
1326}
1327
1328/**
1329 * gmap_unshadow_pgt - remove a shadow page table from a segment entry
1330 * @sg: pointer to the shadow guest address space structure
1331 * @raddr: address in the shadow guest address space
1332 *
1333 * Called with the sg->guest_table_lock
1334 */
1335static void gmap_unshadow_pgt(struct gmap *sg, unsigned long raddr)
1336{
1337	unsigned long *ste;
1338	phys_addr_t sto, pgt;
1339	struct page *page;
1340
1341	BUG_ON(!gmap_is_shadow(sg));
1342	ste = gmap_table_walk(sg, raddr, 1); /* get segment pointer */
1343	if (!ste || !(*ste & _SEGMENT_ENTRY_ORIGIN))
1344		return;
1345	gmap_call_notifier(sg, raddr, raddr + _SEGMENT_SIZE - 1);
1346	sto = __pa(ste - ((raddr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT));
1347	gmap_idte_one(sto | _ASCE_TYPE_SEGMENT, raddr);
1348	pgt = *ste & _SEGMENT_ENTRY_ORIGIN;
1349	*ste = _SEGMENT_ENTRY_EMPTY;
1350	__gmap_unshadow_pgt(sg, raddr, __va(pgt));
1351	/* Free page table */
1352	page = phys_to_page(pgt);
1353	list_del(&page->lru);
1354	page_table_free_pgste(page);
1355}
1356
1357/**
1358 * __gmap_unshadow_sgt - remove all entries from a shadow segment table
1359 * @sg: pointer to the shadow guest address space structure
1360 * @raddr: rmap address in the shadow guest address space
1361 * @sgt: pointer to the start of a shadow segment table
1362 *
1363 * Called with the sg->guest_table_lock
1364 */
1365static void __gmap_unshadow_sgt(struct gmap *sg, unsigned long raddr,
1366				unsigned long *sgt)
1367{
1368	struct page *page;
1369	phys_addr_t pgt;
1370	int i;
1371
1372	BUG_ON(!gmap_is_shadow(sg));
1373	for (i = 0; i < _CRST_ENTRIES; i++, raddr += _SEGMENT_SIZE) {
1374		if (!(sgt[i] & _SEGMENT_ENTRY_ORIGIN))
1375			continue;
1376		pgt = sgt[i] & _REGION_ENTRY_ORIGIN;
1377		sgt[i] = _SEGMENT_ENTRY_EMPTY;
1378		__gmap_unshadow_pgt(sg, raddr, __va(pgt));
1379		/* Free page table */
1380		page = phys_to_page(pgt);
1381		list_del(&page->lru);
1382		page_table_free_pgste(page);
1383	}
1384}
1385
1386/**
1387 * gmap_unshadow_sgt - remove a shadow segment table from a region-3 entry
1388 * @sg: pointer to the shadow guest address space structure
1389 * @raddr: rmap address in the shadow guest address space
1390 *
1391 * Called with the shadow->guest_table_lock
1392 */
1393static void gmap_unshadow_sgt(struct gmap *sg, unsigned long raddr)
1394{
1395	unsigned long r3o, *r3e;
1396	phys_addr_t sgt;
1397	struct page *page;
1398
1399	BUG_ON(!gmap_is_shadow(sg));
1400	r3e = gmap_table_walk(sg, raddr, 2); /* get region-3 pointer */
1401	if (!r3e || !(*r3e & _REGION_ENTRY_ORIGIN))
1402		return;
1403	gmap_call_notifier(sg, raddr, raddr + _REGION3_SIZE - 1);
1404	r3o = (unsigned long) (r3e - ((raddr & _REGION3_INDEX) >> _REGION3_SHIFT));
1405	gmap_idte_one(__pa(r3o) | _ASCE_TYPE_REGION3, raddr);
1406	sgt = *r3e & _REGION_ENTRY_ORIGIN;
1407	*r3e = _REGION3_ENTRY_EMPTY;
1408	__gmap_unshadow_sgt(sg, raddr, __va(sgt));
1409	/* Free segment table */
1410	page = phys_to_page(sgt);
1411	list_del(&page->lru);
1412	__free_pages(page, CRST_ALLOC_ORDER);
1413}
1414
1415/**
1416 * __gmap_unshadow_r3t - remove all entries from a shadow region-3 table
1417 * @sg: pointer to the shadow guest address space structure
1418 * @raddr: address in the shadow guest address space
1419 * @r3t: pointer to the start of a shadow region-3 table
1420 *
1421 * Called with the sg->guest_table_lock
1422 */
1423static void __gmap_unshadow_r3t(struct gmap *sg, unsigned long raddr,
1424				unsigned long *r3t)
1425{
1426	struct page *page;
1427	phys_addr_t sgt;
1428	int i;
1429
1430	BUG_ON(!gmap_is_shadow(sg));
1431	for (i = 0; i < _CRST_ENTRIES; i++, raddr += _REGION3_SIZE) {
1432		if (!(r3t[i] & _REGION_ENTRY_ORIGIN))
1433			continue;
1434		sgt = r3t[i] & _REGION_ENTRY_ORIGIN;
1435		r3t[i] = _REGION3_ENTRY_EMPTY;
1436		__gmap_unshadow_sgt(sg, raddr, __va(sgt));
1437		/* Free segment table */
1438		page = phys_to_page(sgt);
1439		list_del(&page->lru);
1440		__free_pages(page, CRST_ALLOC_ORDER);
1441	}
1442}
1443
1444/**
1445 * gmap_unshadow_r3t - remove a shadow region-3 table from a region-2 entry
1446 * @sg: pointer to the shadow guest address space structure
1447 * @raddr: rmap address in the shadow guest address space
1448 *
1449 * Called with the sg->guest_table_lock
1450 */
1451static void gmap_unshadow_r3t(struct gmap *sg, unsigned long raddr)
1452{
1453	unsigned long r2o, *r2e;
1454	phys_addr_t r3t;
1455	struct page *page;
1456
1457	BUG_ON(!gmap_is_shadow(sg));
1458	r2e = gmap_table_walk(sg, raddr, 3); /* get region-2 pointer */
1459	if (!r2e || !(*r2e & _REGION_ENTRY_ORIGIN))
1460		return;
1461	gmap_call_notifier(sg, raddr, raddr + _REGION2_SIZE - 1);
1462	r2o = (unsigned long) (r2e - ((raddr & _REGION2_INDEX) >> _REGION2_SHIFT));
1463	gmap_idte_one(__pa(r2o) | _ASCE_TYPE_REGION2, raddr);
1464	r3t = *r2e & _REGION_ENTRY_ORIGIN;
1465	*r2e = _REGION2_ENTRY_EMPTY;
1466	__gmap_unshadow_r3t(sg, raddr, __va(r3t));
1467	/* Free region 3 table */
1468	page = phys_to_page(r3t);
1469	list_del(&page->lru);
1470	__free_pages(page, CRST_ALLOC_ORDER);
1471}
1472
1473/**
1474 * __gmap_unshadow_r2t - remove all entries from a shadow region-2 table
1475 * @sg: pointer to the shadow guest address space structure
1476 * @raddr: rmap address in the shadow guest address space
1477 * @r2t: pointer to the start of a shadow region-2 table
1478 *
1479 * Called with the sg->guest_table_lock
1480 */
1481static void __gmap_unshadow_r2t(struct gmap *sg, unsigned long raddr,
1482				unsigned long *r2t)
1483{
1484	phys_addr_t r3t;
1485	struct page *page;
1486	int i;
1487
1488	BUG_ON(!gmap_is_shadow(sg));
1489	for (i = 0; i < _CRST_ENTRIES; i++, raddr += _REGION2_SIZE) {
1490		if (!(r2t[i] & _REGION_ENTRY_ORIGIN))
1491			continue;
1492		r3t = r2t[i] & _REGION_ENTRY_ORIGIN;
1493		r2t[i] = _REGION2_ENTRY_EMPTY;
1494		__gmap_unshadow_r3t(sg, raddr, __va(r3t));
1495		/* Free region 3 table */
1496		page = phys_to_page(r3t);
1497		list_del(&page->lru);
1498		__free_pages(page, CRST_ALLOC_ORDER);
1499	}
1500}
1501
1502/**
1503 * gmap_unshadow_r2t - remove a shadow region-2 table from a region-1 entry
1504 * @sg: pointer to the shadow guest address space structure
1505 * @raddr: rmap address in the shadow guest address space
1506 *
1507 * Called with the sg->guest_table_lock
1508 */
1509static void gmap_unshadow_r2t(struct gmap *sg, unsigned long raddr)
1510{
1511	unsigned long r1o, *r1e;
1512	struct page *page;
1513	phys_addr_t r2t;
1514
1515	BUG_ON(!gmap_is_shadow(sg));
1516	r1e = gmap_table_walk(sg, raddr, 4); /* get region-1 pointer */
1517	if (!r1e || !(*r1e & _REGION_ENTRY_ORIGIN))
1518		return;
1519	gmap_call_notifier(sg, raddr, raddr + _REGION1_SIZE - 1);
1520	r1o = (unsigned long) (r1e - ((raddr & _REGION1_INDEX) >> _REGION1_SHIFT));
1521	gmap_idte_one(__pa(r1o) | _ASCE_TYPE_REGION1, raddr);
1522	r2t = *r1e & _REGION_ENTRY_ORIGIN;
1523	*r1e = _REGION1_ENTRY_EMPTY;
1524	__gmap_unshadow_r2t(sg, raddr, __va(r2t));
1525	/* Free region 2 table */
1526	page = phys_to_page(r2t);
1527	list_del(&page->lru);
1528	__free_pages(page, CRST_ALLOC_ORDER);
1529}
1530
1531/**
1532 * __gmap_unshadow_r1t - remove all entries from a shadow region-1 table
1533 * @sg: pointer to the shadow guest address space structure
1534 * @raddr: rmap address in the shadow guest address space
1535 * @r1t: pointer to the start of a shadow region-1 table
1536 *
1537 * Called with the shadow->guest_table_lock
1538 */
1539static void __gmap_unshadow_r1t(struct gmap *sg, unsigned long raddr,
1540				unsigned long *r1t)
1541{
1542	unsigned long asce;
1543	struct page *page;
1544	phys_addr_t r2t;
1545	int i;
1546
1547	BUG_ON(!gmap_is_shadow(sg));
1548	asce = __pa(r1t) | _ASCE_TYPE_REGION1;
1549	for (i = 0; i < _CRST_ENTRIES; i++, raddr += _REGION1_SIZE) {
1550		if (!(r1t[i] & _REGION_ENTRY_ORIGIN))
1551			continue;
1552		r2t = r1t[i] & _REGION_ENTRY_ORIGIN;
1553		__gmap_unshadow_r2t(sg, raddr, __va(r2t));
1554		/* Clear entry and flush translation r1t -> r2t */
1555		gmap_idte_one(asce, raddr);
1556		r1t[i] = _REGION1_ENTRY_EMPTY;
1557		/* Free region 2 table */
1558		page = phys_to_page(r2t);
1559		list_del(&page->lru);
1560		__free_pages(page, CRST_ALLOC_ORDER);
1561	}
1562}
1563
1564/**
1565 * gmap_unshadow - remove a shadow page table completely
1566 * @sg: pointer to the shadow guest address space structure
1567 *
1568 * Called with sg->guest_table_lock
1569 */
1570static void gmap_unshadow(struct gmap *sg)
1571{
1572	unsigned long *table;
1573
1574	BUG_ON(!gmap_is_shadow(sg));
1575	if (sg->removed)
1576		return;
1577	sg->removed = 1;
1578	gmap_call_notifier(sg, 0, -1UL);
1579	gmap_flush_tlb(sg);
1580	table = __va(sg->asce & _ASCE_ORIGIN);
1581	switch (sg->asce & _ASCE_TYPE_MASK) {
1582	case _ASCE_TYPE_REGION1:
1583		__gmap_unshadow_r1t(sg, 0, table);
1584		break;
1585	case _ASCE_TYPE_REGION2:
1586		__gmap_unshadow_r2t(sg, 0, table);
1587		break;
1588	case _ASCE_TYPE_REGION3:
1589		__gmap_unshadow_r3t(sg, 0, table);
1590		break;
1591	case _ASCE_TYPE_SEGMENT:
1592		__gmap_unshadow_sgt(sg, 0, table);
1593		break;
1594	}
1595}
1596
1597/**
1598 * gmap_find_shadow - find a specific asce in the list of shadow tables
1599 * @parent: pointer to the parent gmap
1600 * @asce: ASCE for which the shadow table is created
1601 * @edat_level: edat level to be used for the shadow translation
1602 *
1603 * Returns the pointer to a gmap if a shadow table with the given asce is
1604 * already available, ERR_PTR(-EAGAIN) if another one is just being created,
1605 * otherwise NULL
1606 */
1607static struct gmap *gmap_find_shadow(struct gmap *parent, unsigned long asce,
1608				     int edat_level)
1609{
1610	struct gmap *sg;
1611
1612	list_for_each_entry(sg, &parent->children, list) {
1613		if (sg->orig_asce != asce || sg->edat_level != edat_level ||
1614		    sg->removed)
1615			continue;
1616		if (!sg->initialized)
1617			return ERR_PTR(-EAGAIN);
1618		refcount_inc(&sg->ref_count);
1619		return sg;
1620	}
1621	return NULL;
1622}
1623
1624/**
1625 * gmap_shadow_valid - check if a shadow guest address space matches the
1626 *                     given properties and is still valid
1627 * @sg: pointer to the shadow guest address space structure
1628 * @asce: ASCE for which the shadow table is requested
1629 * @edat_level: edat level to be used for the shadow translation
1630 *
1631 * Returns 1 if the gmap shadow is still valid and matches the given
1632 * properties, the caller can continue using it. Returns 0 otherwise, the
1633 * caller has to request a new shadow gmap in this case.
1634 *
1635 */
1636int gmap_shadow_valid(struct gmap *sg, unsigned long asce, int edat_level)
1637{
1638	if (sg->removed)
1639		return 0;
1640	return sg->orig_asce == asce && sg->edat_level == edat_level;
1641}
1642EXPORT_SYMBOL_GPL(gmap_shadow_valid);
1643
1644/**
1645 * gmap_shadow - create/find a shadow guest address space
1646 * @parent: pointer to the parent gmap
1647 * @asce: ASCE for which the shadow table is created
1648 * @edat_level: edat level to be used for the shadow translation
1649 *
1650 * The pages of the top level page table referred by the asce parameter
1651 * will be set to read-only and marked in the PGSTEs of the kvm process.
1652 * The shadow table will be removed automatically on any change to the
1653 * PTE mapping for the source table.
1654 *
1655 * Returns a guest address space structure, ERR_PTR(-ENOMEM) if out of memory,
1656 * ERR_PTR(-EAGAIN) if the caller has to retry and ERR_PTR(-EFAULT) if the
1657 * parent gmap table could not be protected.
1658 */
1659struct gmap *gmap_shadow(struct gmap *parent, unsigned long asce,
1660			 int edat_level)
1661{
1662	struct gmap *sg, *new;
1663	unsigned long limit;
1664	int rc;
1665
1666	BUG_ON(parent->mm->context.allow_gmap_hpage_1m);
1667	BUG_ON(gmap_is_shadow(parent));
1668	spin_lock(&parent->shadow_lock);
1669	sg = gmap_find_shadow(parent, asce, edat_level);
1670	spin_unlock(&parent->shadow_lock);
1671	if (sg)
1672		return sg;
1673	/* Create a new shadow gmap */
1674	limit = -1UL >> (33 - (((asce & _ASCE_TYPE_MASK) >> 2) * 11));
1675	if (asce & _ASCE_REAL_SPACE)
1676		limit = -1UL;
1677	new = gmap_alloc(limit);
1678	if (!new)
1679		return ERR_PTR(-ENOMEM);
1680	new->mm = parent->mm;
1681	new->parent = gmap_get(parent);
 
1682	new->orig_asce = asce;
1683	new->edat_level = edat_level;
1684	new->initialized = false;
1685	spin_lock(&parent->shadow_lock);
1686	/* Recheck if another CPU created the same shadow */
1687	sg = gmap_find_shadow(parent, asce, edat_level);
1688	if (sg) {
1689		spin_unlock(&parent->shadow_lock);
1690		gmap_free(new);
1691		return sg;
1692	}
1693	if (asce & _ASCE_REAL_SPACE) {
1694		/* only allow one real-space gmap shadow */
1695		list_for_each_entry(sg, &parent->children, list) {
1696			if (sg->orig_asce & _ASCE_REAL_SPACE) {
1697				spin_lock(&sg->guest_table_lock);
1698				gmap_unshadow(sg);
1699				spin_unlock(&sg->guest_table_lock);
1700				list_del(&sg->list);
1701				gmap_put(sg);
1702				break;
1703			}
1704		}
1705	}
1706	refcount_set(&new->ref_count, 2);
1707	list_add(&new->list, &parent->children);
1708	if (asce & _ASCE_REAL_SPACE) {
1709		/* nothing to protect, return right away */
1710		new->initialized = true;
1711		spin_unlock(&parent->shadow_lock);
1712		return new;
1713	}
1714	spin_unlock(&parent->shadow_lock);
1715	/* protect after insertion, so it will get properly invalidated */
1716	mmap_read_lock(parent->mm);
1717	rc = gmap_protect_range(parent, asce & _ASCE_ORIGIN,
1718				((asce & _ASCE_TABLE_LENGTH) + 1) * PAGE_SIZE,
1719				PROT_READ, GMAP_NOTIFY_SHADOW);
1720	mmap_read_unlock(parent->mm);
1721	spin_lock(&parent->shadow_lock);
1722	new->initialized = true;
1723	if (rc) {
1724		list_del(&new->list);
1725		gmap_free(new);
1726		new = ERR_PTR(rc);
1727	}
1728	spin_unlock(&parent->shadow_lock);
1729	return new;
1730}
1731EXPORT_SYMBOL_GPL(gmap_shadow);
1732
1733/**
1734 * gmap_shadow_r2t - create an empty shadow region 2 table
1735 * @sg: pointer to the shadow guest address space structure
1736 * @saddr: faulting address in the shadow gmap
1737 * @r2t: parent gmap address of the region 2 table to get shadowed
1738 * @fake: r2t references contiguous guest memory block, not a r2t
1739 *
1740 * The r2t parameter specifies the address of the source table. The
1741 * four pages of the source table are made read-only in the parent gmap
1742 * address space. A write to the source table area @r2t will automatically
1743 * remove the shadow r2 table and all of its decendents.
1744 *
1745 * Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
1746 * shadow table structure is incomplete, -ENOMEM if out of memory and
1747 * -EFAULT if an address in the parent gmap could not be resolved.
1748 *
1749 * Called with sg->mm->mmap_lock in read.
1750 */
1751int gmap_shadow_r2t(struct gmap *sg, unsigned long saddr, unsigned long r2t,
1752		    int fake)
1753{
1754	unsigned long raddr, origin, offset, len;
1755	unsigned long *table;
1756	phys_addr_t s_r2t;
1757	struct page *page;
1758	int rc;
1759
1760	BUG_ON(!gmap_is_shadow(sg));
1761	/* Allocate a shadow region second table */
1762	page = alloc_pages(GFP_KERNEL_ACCOUNT, CRST_ALLOC_ORDER);
1763	if (!page)
1764		return -ENOMEM;
1765	page->index = r2t & _REGION_ENTRY_ORIGIN;
1766	if (fake)
1767		page->index |= GMAP_SHADOW_FAKE_TABLE;
1768	s_r2t = page_to_phys(page);
1769	/* Install shadow region second table */
1770	spin_lock(&sg->guest_table_lock);
1771	table = gmap_table_walk(sg, saddr, 4); /* get region-1 pointer */
1772	if (!table) {
1773		rc = -EAGAIN;		/* Race with unshadow */
1774		goto out_free;
1775	}
1776	if (!(*table & _REGION_ENTRY_INVALID)) {
1777		rc = 0;			/* Already established */
1778		goto out_free;
1779	} else if (*table & _REGION_ENTRY_ORIGIN) {
1780		rc = -EAGAIN;		/* Race with shadow */
1781		goto out_free;
1782	}
1783	crst_table_init(__va(s_r2t), _REGION2_ENTRY_EMPTY);
1784	/* mark as invalid as long as the parent table is not protected */
1785	*table = s_r2t | _REGION_ENTRY_LENGTH |
1786		 _REGION_ENTRY_TYPE_R1 | _REGION_ENTRY_INVALID;
1787	if (sg->edat_level >= 1)
1788		*table |= (r2t & _REGION_ENTRY_PROTECT);
1789	list_add(&page->lru, &sg->crst_list);
1790	if (fake) {
1791		/* nothing to protect for fake tables */
1792		*table &= ~_REGION_ENTRY_INVALID;
1793		spin_unlock(&sg->guest_table_lock);
1794		return 0;
1795	}
1796	spin_unlock(&sg->guest_table_lock);
1797	/* Make r2t read-only in parent gmap page table */
1798	raddr = (saddr & _REGION1_MASK) | _SHADOW_RMAP_REGION1;
1799	origin = r2t & _REGION_ENTRY_ORIGIN;
1800	offset = ((r2t & _REGION_ENTRY_OFFSET) >> 6) * PAGE_SIZE;
1801	len = ((r2t & _REGION_ENTRY_LENGTH) + 1) * PAGE_SIZE - offset;
1802	rc = gmap_protect_rmap(sg, raddr, origin + offset, len);
1803	spin_lock(&sg->guest_table_lock);
1804	if (!rc) {
1805		table = gmap_table_walk(sg, saddr, 4);
1806		if (!table || (*table & _REGION_ENTRY_ORIGIN) != s_r2t)
1807			rc = -EAGAIN;		/* Race with unshadow */
1808		else
1809			*table &= ~_REGION_ENTRY_INVALID;
1810	} else {
1811		gmap_unshadow_r2t(sg, raddr);
1812	}
1813	spin_unlock(&sg->guest_table_lock);
1814	return rc;
1815out_free:
1816	spin_unlock(&sg->guest_table_lock);
1817	__free_pages(page, CRST_ALLOC_ORDER);
1818	return rc;
1819}
1820EXPORT_SYMBOL_GPL(gmap_shadow_r2t);
1821
1822/**
1823 * gmap_shadow_r3t - create a shadow region 3 table
1824 * @sg: pointer to the shadow guest address space structure
1825 * @saddr: faulting address in the shadow gmap
1826 * @r3t: parent gmap address of the region 3 table to get shadowed
1827 * @fake: r3t references contiguous guest memory block, not a r3t
1828 *
1829 * Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
1830 * shadow table structure is incomplete, -ENOMEM if out of memory and
1831 * -EFAULT if an address in the parent gmap could not be resolved.
1832 *
1833 * Called with sg->mm->mmap_lock in read.
1834 */
1835int gmap_shadow_r3t(struct gmap *sg, unsigned long saddr, unsigned long r3t,
1836		    int fake)
1837{
1838	unsigned long raddr, origin, offset, len;
1839	unsigned long *table;
1840	phys_addr_t s_r3t;
1841	struct page *page;
1842	int rc;
1843
1844	BUG_ON(!gmap_is_shadow(sg));
1845	/* Allocate a shadow region second table */
1846	page = alloc_pages(GFP_KERNEL_ACCOUNT, CRST_ALLOC_ORDER);
1847	if (!page)
1848		return -ENOMEM;
1849	page->index = r3t & _REGION_ENTRY_ORIGIN;
1850	if (fake)
1851		page->index |= GMAP_SHADOW_FAKE_TABLE;
1852	s_r3t = page_to_phys(page);
1853	/* Install shadow region second table */
1854	spin_lock(&sg->guest_table_lock);
1855	table = gmap_table_walk(sg, saddr, 3); /* get region-2 pointer */
1856	if (!table) {
1857		rc = -EAGAIN;		/* Race with unshadow */
1858		goto out_free;
1859	}
1860	if (!(*table & _REGION_ENTRY_INVALID)) {
1861		rc = 0;			/* Already established */
1862		goto out_free;
1863	} else if (*table & _REGION_ENTRY_ORIGIN) {
1864		rc = -EAGAIN;		/* Race with shadow */
1865		goto out_free;
1866	}
1867	crst_table_init(__va(s_r3t), _REGION3_ENTRY_EMPTY);
1868	/* mark as invalid as long as the parent table is not protected */
1869	*table = s_r3t | _REGION_ENTRY_LENGTH |
1870		 _REGION_ENTRY_TYPE_R2 | _REGION_ENTRY_INVALID;
1871	if (sg->edat_level >= 1)
1872		*table |= (r3t & _REGION_ENTRY_PROTECT);
1873	list_add(&page->lru, &sg->crst_list);
1874	if (fake) {
1875		/* nothing to protect for fake tables */
1876		*table &= ~_REGION_ENTRY_INVALID;
1877		spin_unlock(&sg->guest_table_lock);
1878		return 0;
1879	}
1880	spin_unlock(&sg->guest_table_lock);
1881	/* Make r3t read-only in parent gmap page table */
1882	raddr = (saddr & _REGION2_MASK) | _SHADOW_RMAP_REGION2;
1883	origin = r3t & _REGION_ENTRY_ORIGIN;
1884	offset = ((r3t & _REGION_ENTRY_OFFSET) >> 6) * PAGE_SIZE;
1885	len = ((r3t & _REGION_ENTRY_LENGTH) + 1) * PAGE_SIZE - offset;
1886	rc = gmap_protect_rmap(sg, raddr, origin + offset, len);
1887	spin_lock(&sg->guest_table_lock);
1888	if (!rc) {
1889		table = gmap_table_walk(sg, saddr, 3);
1890		if (!table || (*table & _REGION_ENTRY_ORIGIN) != s_r3t)
1891			rc = -EAGAIN;		/* Race with unshadow */
1892		else
1893			*table &= ~_REGION_ENTRY_INVALID;
1894	} else {
1895		gmap_unshadow_r3t(sg, raddr);
1896	}
1897	spin_unlock(&sg->guest_table_lock);
1898	return rc;
1899out_free:
1900	spin_unlock(&sg->guest_table_lock);
1901	__free_pages(page, CRST_ALLOC_ORDER);
1902	return rc;
1903}
1904EXPORT_SYMBOL_GPL(gmap_shadow_r3t);
1905
1906/**
1907 * gmap_shadow_sgt - create a shadow segment table
1908 * @sg: pointer to the shadow guest address space structure
1909 * @saddr: faulting address in the shadow gmap
1910 * @sgt: parent gmap address of the segment table to get shadowed
1911 * @fake: sgt references contiguous guest memory block, not a sgt
1912 *
1913 * Returns: 0 if successfully shadowed or already shadowed, -EAGAIN if the
1914 * shadow table structure is incomplete, -ENOMEM if out of memory and
1915 * -EFAULT if an address in the parent gmap could not be resolved.
1916 *
1917 * Called with sg->mm->mmap_lock in read.
1918 */
1919int gmap_shadow_sgt(struct gmap *sg, unsigned long saddr, unsigned long sgt,
1920		    int fake)
1921{
1922	unsigned long raddr, origin, offset, len;
1923	unsigned long *table;
1924	phys_addr_t s_sgt;
1925	struct page *page;
1926	int rc;
1927
1928	BUG_ON(!gmap_is_shadow(sg) || (sgt & _REGION3_ENTRY_LARGE));
1929	/* Allocate a shadow segment table */
1930	page = alloc_pages(GFP_KERNEL_ACCOUNT, CRST_ALLOC_ORDER);
1931	if (!page)
1932		return -ENOMEM;
1933	page->index = sgt & _REGION_ENTRY_ORIGIN;
1934	if (fake)
1935		page->index |= GMAP_SHADOW_FAKE_TABLE;
1936	s_sgt = page_to_phys(page);
1937	/* Install shadow region second table */
1938	spin_lock(&sg->guest_table_lock);
1939	table = gmap_table_walk(sg, saddr, 2); /* get region-3 pointer */
1940	if (!table) {
1941		rc = -EAGAIN;		/* Race with unshadow */
1942		goto out_free;
1943	}
1944	if (!(*table & _REGION_ENTRY_INVALID)) {
1945		rc = 0;			/* Already established */
1946		goto out_free;
1947	} else if (*table & _REGION_ENTRY_ORIGIN) {
1948		rc = -EAGAIN;		/* Race with shadow */
1949		goto out_free;
1950	}
1951	crst_table_init(__va(s_sgt), _SEGMENT_ENTRY_EMPTY);
1952	/* mark as invalid as long as the parent table is not protected */
1953	*table = s_sgt | _REGION_ENTRY_LENGTH |
1954		 _REGION_ENTRY_TYPE_R3 | _REGION_ENTRY_INVALID;
1955	if (sg->edat_level >= 1)
1956		*table |= sgt & _REGION_ENTRY_PROTECT;
1957	list_add(&page->lru, &sg->crst_list);
1958	if (fake) {
1959		/* nothing to protect for fake tables */
1960		*table &= ~_REGION_ENTRY_INVALID;
1961		spin_unlock(&sg->guest_table_lock);
1962		return 0;
1963	}
1964	spin_unlock(&sg->guest_table_lock);
1965	/* Make sgt read-only in parent gmap page table */
1966	raddr = (saddr & _REGION3_MASK) | _SHADOW_RMAP_REGION3;
1967	origin = sgt & _REGION_ENTRY_ORIGIN;
1968	offset = ((sgt & _REGION_ENTRY_OFFSET) >> 6) * PAGE_SIZE;
1969	len = ((sgt & _REGION_ENTRY_LENGTH) + 1) * PAGE_SIZE - offset;
1970	rc = gmap_protect_rmap(sg, raddr, origin + offset, len);
1971	spin_lock(&sg->guest_table_lock);
1972	if (!rc) {
1973		table = gmap_table_walk(sg, saddr, 2);
1974		if (!table || (*table & _REGION_ENTRY_ORIGIN) != s_sgt)
1975			rc = -EAGAIN;		/* Race with unshadow */
1976		else
1977			*table &= ~_REGION_ENTRY_INVALID;
1978	} else {
1979		gmap_unshadow_sgt(sg, raddr);
1980	}
1981	spin_unlock(&sg->guest_table_lock);
1982	return rc;
1983out_free:
1984	spin_unlock(&sg->guest_table_lock);
1985	__free_pages(page, CRST_ALLOC_ORDER);
1986	return rc;
1987}
1988EXPORT_SYMBOL_GPL(gmap_shadow_sgt);
1989
1990/**
1991 * gmap_shadow_pgt_lookup - find a shadow page table
1992 * @sg: pointer to the shadow guest address space structure
1993 * @saddr: the address in the shadow aguest address space
1994 * @pgt: parent gmap address of the page table to get shadowed
1995 * @dat_protection: if the pgtable is marked as protected by dat
1996 * @fake: pgt references contiguous guest memory block, not a pgtable
1997 *
1998 * Returns 0 if the shadow page table was found and -EAGAIN if the page
1999 * table was not found.
2000 *
2001 * Called with sg->mm->mmap_lock in read.
2002 */
2003int gmap_shadow_pgt_lookup(struct gmap *sg, unsigned long saddr,
2004			   unsigned long *pgt, int *dat_protection,
2005			   int *fake)
2006{
2007	unsigned long *table;
2008	struct page *page;
2009	int rc;
2010
2011	BUG_ON(!gmap_is_shadow(sg));
2012	spin_lock(&sg->guest_table_lock);
2013	table = gmap_table_walk(sg, saddr, 1); /* get segment pointer */
2014	if (table && !(*table & _SEGMENT_ENTRY_INVALID)) {
2015		/* Shadow page tables are full pages (pte+pgste) */
2016		page = pfn_to_page(*table >> PAGE_SHIFT);
2017		*pgt = page->index & ~GMAP_SHADOW_FAKE_TABLE;
2018		*dat_protection = !!(*table & _SEGMENT_ENTRY_PROTECT);
2019		*fake = !!(page->index & GMAP_SHADOW_FAKE_TABLE);
2020		rc = 0;
2021	} else  {
2022		rc = -EAGAIN;
2023	}
2024	spin_unlock(&sg->guest_table_lock);
2025	return rc;
2026
2027}
2028EXPORT_SYMBOL_GPL(gmap_shadow_pgt_lookup);
2029
2030/**
2031 * gmap_shadow_pgt - instantiate a shadow page table
2032 * @sg: pointer to the shadow guest address space structure
2033 * @saddr: faulting address in the shadow gmap
2034 * @pgt: parent gmap address of the page table to get shadowed
2035 * @fake: pgt references contiguous guest memory block, not a pgtable
2036 *
2037 * Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
2038 * shadow table structure is incomplete, -ENOMEM if out of memory,
2039 * -EFAULT if an address in the parent gmap could not be resolved and
2040 *
2041 * Called with gmap->mm->mmap_lock in read
2042 */
2043int gmap_shadow_pgt(struct gmap *sg, unsigned long saddr, unsigned long pgt,
2044		    int fake)
2045{
2046	unsigned long raddr, origin;
2047	unsigned long *table;
2048	struct page *page;
2049	phys_addr_t s_pgt;
2050	int rc;
2051
2052	BUG_ON(!gmap_is_shadow(sg) || (pgt & _SEGMENT_ENTRY_LARGE));
2053	/* Allocate a shadow page table */
2054	page = page_table_alloc_pgste(sg->mm);
2055	if (!page)
2056		return -ENOMEM;
2057	page->index = pgt & _SEGMENT_ENTRY_ORIGIN;
2058	if (fake)
2059		page->index |= GMAP_SHADOW_FAKE_TABLE;
2060	s_pgt = page_to_phys(page);
2061	/* Install shadow page table */
2062	spin_lock(&sg->guest_table_lock);
2063	table = gmap_table_walk(sg, saddr, 1); /* get segment pointer */
2064	if (!table) {
2065		rc = -EAGAIN;		/* Race with unshadow */
2066		goto out_free;
2067	}
2068	if (!(*table & _SEGMENT_ENTRY_INVALID)) {
2069		rc = 0;			/* Already established */
2070		goto out_free;
2071	} else if (*table & _SEGMENT_ENTRY_ORIGIN) {
2072		rc = -EAGAIN;		/* Race with shadow */
2073		goto out_free;
2074	}
2075	/* mark as invalid as long as the parent table is not protected */
2076	*table = (unsigned long) s_pgt | _SEGMENT_ENTRY |
2077		 (pgt & _SEGMENT_ENTRY_PROTECT) | _SEGMENT_ENTRY_INVALID;
2078	list_add(&page->lru, &sg->pt_list);
2079	if (fake) {
2080		/* nothing to protect for fake tables */
2081		*table &= ~_SEGMENT_ENTRY_INVALID;
2082		spin_unlock(&sg->guest_table_lock);
2083		return 0;
2084	}
2085	spin_unlock(&sg->guest_table_lock);
2086	/* Make pgt read-only in parent gmap page table (not the pgste) */
2087	raddr = (saddr & _SEGMENT_MASK) | _SHADOW_RMAP_SEGMENT;
2088	origin = pgt & _SEGMENT_ENTRY_ORIGIN & PAGE_MASK;
2089	rc = gmap_protect_rmap(sg, raddr, origin, PAGE_SIZE);
2090	spin_lock(&sg->guest_table_lock);
2091	if (!rc) {
2092		table = gmap_table_walk(sg, saddr, 1);
2093		if (!table || (*table & _SEGMENT_ENTRY_ORIGIN) != s_pgt)
2094			rc = -EAGAIN;		/* Race with unshadow */
2095		else
2096			*table &= ~_SEGMENT_ENTRY_INVALID;
2097	} else {
2098		gmap_unshadow_pgt(sg, raddr);
2099	}
2100	spin_unlock(&sg->guest_table_lock);
2101	return rc;
2102out_free:
2103	spin_unlock(&sg->guest_table_lock);
2104	page_table_free_pgste(page);
2105	return rc;
2106
2107}
2108EXPORT_SYMBOL_GPL(gmap_shadow_pgt);
2109
2110/**
2111 * gmap_shadow_page - create a shadow page mapping
2112 * @sg: pointer to the shadow guest address space structure
2113 * @saddr: faulting address in the shadow gmap
2114 * @pte: pte in parent gmap address space to get shadowed
2115 *
2116 * Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
2117 * shadow table structure is incomplete, -ENOMEM if out of memory and
2118 * -EFAULT if an address in the parent gmap could not be resolved.
2119 *
2120 * Called with sg->mm->mmap_lock in read.
2121 */
2122int gmap_shadow_page(struct gmap *sg, unsigned long saddr, pte_t pte)
2123{
2124	struct gmap *parent;
2125	struct gmap_rmap *rmap;
2126	unsigned long vmaddr, paddr;
2127	spinlock_t *ptl;
2128	pte_t *sptep, *tptep;
2129	int prot;
2130	int rc;
2131
2132	BUG_ON(!gmap_is_shadow(sg));
2133	parent = sg->parent;
2134	prot = (pte_val(pte) & _PAGE_PROTECT) ? PROT_READ : PROT_WRITE;
2135
2136	rmap = kzalloc(sizeof(*rmap), GFP_KERNEL_ACCOUNT);
2137	if (!rmap)
2138		return -ENOMEM;
2139	rmap->raddr = (saddr & PAGE_MASK) | _SHADOW_RMAP_PGTABLE;
2140
2141	while (1) {
2142		paddr = pte_val(pte) & PAGE_MASK;
2143		vmaddr = __gmap_translate(parent, paddr);
2144		if (IS_ERR_VALUE(vmaddr)) {
2145			rc = vmaddr;
2146			break;
2147		}
2148		rc = radix_tree_preload(GFP_KERNEL_ACCOUNT);
2149		if (rc)
2150			break;
2151		rc = -EAGAIN;
2152		sptep = gmap_pte_op_walk(parent, paddr, &ptl);
2153		if (sptep) {
2154			spin_lock(&sg->guest_table_lock);
2155			/* Get page table pointer */
2156			tptep = (pte_t *) gmap_table_walk(sg, saddr, 0);
2157			if (!tptep) {
2158				spin_unlock(&sg->guest_table_lock);
2159				gmap_pte_op_end(ptl);
2160				radix_tree_preload_end();
2161				break;
2162			}
2163			rc = ptep_shadow_pte(sg->mm, saddr, sptep, tptep, pte);
2164			if (rc > 0) {
2165				/* Success and a new mapping */
2166				gmap_insert_rmap(sg, vmaddr, rmap);
2167				rmap = NULL;
2168				rc = 0;
2169			}
2170			gmap_pte_op_end(ptl);
2171			spin_unlock(&sg->guest_table_lock);
2172		}
2173		radix_tree_preload_end();
2174		if (!rc)
2175			break;
2176		rc = gmap_pte_op_fixup(parent, paddr, vmaddr, prot);
2177		if (rc)
2178			break;
2179	}
2180	kfree(rmap);
2181	return rc;
2182}
2183EXPORT_SYMBOL_GPL(gmap_shadow_page);
2184
2185/*
2186 * gmap_shadow_notify - handle notifications for shadow gmap
2187 *
2188 * Called with sg->parent->shadow_lock.
2189 */
2190static void gmap_shadow_notify(struct gmap *sg, unsigned long vmaddr,
2191			       unsigned long gaddr)
2192{
2193	struct gmap_rmap *rmap, *rnext, *head;
2194	unsigned long start, end, bits, raddr;
2195
2196	BUG_ON(!gmap_is_shadow(sg));
2197
2198	spin_lock(&sg->guest_table_lock);
2199	if (sg->removed) {
2200		spin_unlock(&sg->guest_table_lock);
2201		return;
2202	}
2203	/* Check for top level table */
2204	start = sg->orig_asce & _ASCE_ORIGIN;
2205	end = start + ((sg->orig_asce & _ASCE_TABLE_LENGTH) + 1) * PAGE_SIZE;
2206	if (!(sg->orig_asce & _ASCE_REAL_SPACE) && gaddr >= start &&
2207	    gaddr < end) {
2208		/* The complete shadow table has to go */
2209		gmap_unshadow(sg);
2210		spin_unlock(&sg->guest_table_lock);
2211		list_del(&sg->list);
2212		gmap_put(sg);
2213		return;
2214	}
2215	/* Remove the page table tree from on specific entry */
2216	head = radix_tree_delete(&sg->host_to_rmap, vmaddr >> PAGE_SHIFT);
2217	gmap_for_each_rmap_safe(rmap, rnext, head) {
2218		bits = rmap->raddr & _SHADOW_RMAP_MASK;
2219		raddr = rmap->raddr ^ bits;
2220		switch (bits) {
2221		case _SHADOW_RMAP_REGION1:
2222			gmap_unshadow_r2t(sg, raddr);
2223			break;
2224		case _SHADOW_RMAP_REGION2:
2225			gmap_unshadow_r3t(sg, raddr);
2226			break;
2227		case _SHADOW_RMAP_REGION3:
2228			gmap_unshadow_sgt(sg, raddr);
2229			break;
2230		case _SHADOW_RMAP_SEGMENT:
2231			gmap_unshadow_pgt(sg, raddr);
2232			break;
2233		case _SHADOW_RMAP_PGTABLE:
2234			gmap_unshadow_page(sg, raddr);
2235			break;
2236		}
2237		kfree(rmap);
2238	}
2239	spin_unlock(&sg->guest_table_lock);
2240}
2241
2242/**
2243 * ptep_notify - call all invalidation callbacks for a specific pte.
2244 * @mm: pointer to the process mm_struct
2245 * @vmaddr: virtual address in the process address space
2246 * @pte: pointer to the page table entry
2247 * @bits: bits from the pgste that caused the notify call
2248 *
2249 * This function is assumed to be called with the page table lock held
2250 * for the pte to notify.
2251 */
2252void ptep_notify(struct mm_struct *mm, unsigned long vmaddr,
2253		 pte_t *pte, unsigned long bits)
2254{
2255	unsigned long offset, gaddr = 0;
2256	unsigned long *table;
2257	struct gmap *gmap, *sg, *next;
2258
2259	offset = ((unsigned long) pte) & (255 * sizeof(pte_t));
2260	offset = offset * (PAGE_SIZE / sizeof(pte_t));
2261	rcu_read_lock();
2262	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
2263		spin_lock(&gmap->guest_table_lock);
2264		table = radix_tree_lookup(&gmap->host_to_guest,
2265					  vmaddr >> PMD_SHIFT);
2266		if (table)
2267			gaddr = __gmap_segment_gaddr(table) + offset;
2268		spin_unlock(&gmap->guest_table_lock);
2269		if (!table)
2270			continue;
2271
2272		if (!list_empty(&gmap->children) && (bits & PGSTE_VSIE_BIT)) {
2273			spin_lock(&gmap->shadow_lock);
2274			list_for_each_entry_safe(sg, next,
2275						 &gmap->children, list)
2276				gmap_shadow_notify(sg, vmaddr, gaddr);
2277			spin_unlock(&gmap->shadow_lock);
2278		}
2279		if (bits & PGSTE_IN_BIT)
2280			gmap_call_notifier(gmap, gaddr, gaddr + PAGE_SIZE - 1);
2281	}
2282	rcu_read_unlock();
2283}
2284EXPORT_SYMBOL_GPL(ptep_notify);
2285
2286static void pmdp_notify_gmap(struct gmap *gmap, pmd_t *pmdp,
2287			     unsigned long gaddr)
2288{
2289	set_pmd(pmdp, clear_pmd_bit(*pmdp, __pgprot(_SEGMENT_ENTRY_GMAP_IN)));
2290	gmap_call_notifier(gmap, gaddr, gaddr + HPAGE_SIZE - 1);
2291}
2292
2293/**
2294 * gmap_pmdp_xchg - exchange a gmap pmd with another
2295 * @gmap: pointer to the guest address space structure
2296 * @pmdp: pointer to the pmd entry
2297 * @new: replacement entry
2298 * @gaddr: the affected guest address
2299 *
2300 * This function is assumed to be called with the guest_table_lock
2301 * held.
2302 */
2303static void gmap_pmdp_xchg(struct gmap *gmap, pmd_t *pmdp, pmd_t new,
2304			   unsigned long gaddr)
2305{
2306	gaddr &= HPAGE_MASK;
2307	pmdp_notify_gmap(gmap, pmdp, gaddr);
2308	new = clear_pmd_bit(new, __pgprot(_SEGMENT_ENTRY_GMAP_IN));
2309	if (MACHINE_HAS_TLB_GUEST)
2310		__pmdp_idte(gaddr, (pmd_t *)pmdp, IDTE_GUEST_ASCE, gmap->asce,
2311			    IDTE_GLOBAL);
2312	else if (MACHINE_HAS_IDTE)
2313		__pmdp_idte(gaddr, (pmd_t *)pmdp, 0, 0, IDTE_GLOBAL);
2314	else
2315		__pmdp_csp(pmdp);
2316	set_pmd(pmdp, new);
2317}
2318
2319static void gmap_pmdp_clear(struct mm_struct *mm, unsigned long vmaddr,
2320			    int purge)
2321{
2322	pmd_t *pmdp;
2323	struct gmap *gmap;
2324	unsigned long gaddr;
2325
2326	rcu_read_lock();
2327	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
2328		spin_lock(&gmap->guest_table_lock);
2329		pmdp = (pmd_t *)radix_tree_delete(&gmap->host_to_guest,
2330						  vmaddr >> PMD_SHIFT);
2331		if (pmdp) {
2332			gaddr = __gmap_segment_gaddr((unsigned long *)pmdp);
2333			pmdp_notify_gmap(gmap, pmdp, gaddr);
2334			WARN_ON(pmd_val(*pmdp) & ~(_SEGMENT_ENTRY_HARDWARE_BITS_LARGE |
2335						   _SEGMENT_ENTRY_GMAP_UC));
2336			if (purge)
2337				__pmdp_csp(pmdp);
2338			set_pmd(pmdp, __pmd(_SEGMENT_ENTRY_EMPTY));
2339		}
2340		spin_unlock(&gmap->guest_table_lock);
2341	}
2342	rcu_read_unlock();
2343}
2344
2345/**
2346 * gmap_pmdp_invalidate - invalidate all affected guest pmd entries without
2347 *                        flushing
2348 * @mm: pointer to the process mm_struct
2349 * @vmaddr: virtual address in the process address space
2350 */
2351void gmap_pmdp_invalidate(struct mm_struct *mm, unsigned long vmaddr)
2352{
2353	gmap_pmdp_clear(mm, vmaddr, 0);
2354}
2355EXPORT_SYMBOL_GPL(gmap_pmdp_invalidate);
2356
2357/**
2358 * gmap_pmdp_csp - csp all affected guest pmd entries
2359 * @mm: pointer to the process mm_struct
2360 * @vmaddr: virtual address in the process address space
2361 */
2362void gmap_pmdp_csp(struct mm_struct *mm, unsigned long vmaddr)
2363{
2364	gmap_pmdp_clear(mm, vmaddr, 1);
2365}
2366EXPORT_SYMBOL_GPL(gmap_pmdp_csp);
2367
2368/**
2369 * gmap_pmdp_idte_local - invalidate and clear a guest pmd entry
2370 * @mm: pointer to the process mm_struct
2371 * @vmaddr: virtual address in the process address space
2372 */
2373void gmap_pmdp_idte_local(struct mm_struct *mm, unsigned long vmaddr)
2374{
2375	unsigned long *entry, gaddr;
2376	struct gmap *gmap;
2377	pmd_t *pmdp;
2378
2379	rcu_read_lock();
2380	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
2381		spin_lock(&gmap->guest_table_lock);
2382		entry = radix_tree_delete(&gmap->host_to_guest,
2383					  vmaddr >> PMD_SHIFT);
2384		if (entry) {
2385			pmdp = (pmd_t *)entry;
2386			gaddr = __gmap_segment_gaddr(entry);
2387			pmdp_notify_gmap(gmap, pmdp, gaddr);
2388			WARN_ON(*entry & ~(_SEGMENT_ENTRY_HARDWARE_BITS_LARGE |
2389					   _SEGMENT_ENTRY_GMAP_UC));
2390			if (MACHINE_HAS_TLB_GUEST)
2391				__pmdp_idte(gaddr, pmdp, IDTE_GUEST_ASCE,
2392					    gmap->asce, IDTE_LOCAL);
2393			else if (MACHINE_HAS_IDTE)
2394				__pmdp_idte(gaddr, pmdp, 0, 0, IDTE_LOCAL);
2395			*entry = _SEGMENT_ENTRY_EMPTY;
2396		}
2397		spin_unlock(&gmap->guest_table_lock);
2398	}
2399	rcu_read_unlock();
2400}
2401EXPORT_SYMBOL_GPL(gmap_pmdp_idte_local);
2402
2403/**
2404 * gmap_pmdp_idte_global - invalidate and clear a guest pmd entry
2405 * @mm: pointer to the process mm_struct
2406 * @vmaddr: virtual address in the process address space
2407 */
2408void gmap_pmdp_idte_global(struct mm_struct *mm, unsigned long vmaddr)
2409{
2410	unsigned long *entry, gaddr;
2411	struct gmap *gmap;
2412	pmd_t *pmdp;
2413
2414	rcu_read_lock();
2415	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
2416		spin_lock(&gmap->guest_table_lock);
2417		entry = radix_tree_delete(&gmap->host_to_guest,
2418					  vmaddr >> PMD_SHIFT);
2419		if (entry) {
2420			pmdp = (pmd_t *)entry;
2421			gaddr = __gmap_segment_gaddr(entry);
2422			pmdp_notify_gmap(gmap, pmdp, gaddr);
2423			WARN_ON(*entry & ~(_SEGMENT_ENTRY_HARDWARE_BITS_LARGE |
2424					   _SEGMENT_ENTRY_GMAP_UC));
2425			if (MACHINE_HAS_TLB_GUEST)
2426				__pmdp_idte(gaddr, pmdp, IDTE_GUEST_ASCE,
2427					    gmap->asce, IDTE_GLOBAL);
2428			else if (MACHINE_HAS_IDTE)
2429				__pmdp_idte(gaddr, pmdp, 0, 0, IDTE_GLOBAL);
2430			else
2431				__pmdp_csp(pmdp);
2432			*entry = _SEGMENT_ENTRY_EMPTY;
2433		}
2434		spin_unlock(&gmap->guest_table_lock);
2435	}
2436	rcu_read_unlock();
2437}
2438EXPORT_SYMBOL_GPL(gmap_pmdp_idte_global);
2439
2440/**
2441 * gmap_test_and_clear_dirty_pmd - test and reset segment dirty status
2442 * @gmap: pointer to guest address space
2443 * @pmdp: pointer to the pmd to be tested
2444 * @gaddr: virtual address in the guest address space
2445 *
2446 * This function is assumed to be called with the guest_table_lock
2447 * held.
2448 */
2449static bool gmap_test_and_clear_dirty_pmd(struct gmap *gmap, pmd_t *pmdp,
2450					  unsigned long gaddr)
2451{
2452	if (pmd_val(*pmdp) & _SEGMENT_ENTRY_INVALID)
2453		return false;
2454
2455	/* Already protected memory, which did not change is clean */
2456	if (pmd_val(*pmdp) & _SEGMENT_ENTRY_PROTECT &&
2457	    !(pmd_val(*pmdp) & _SEGMENT_ENTRY_GMAP_UC))
2458		return false;
2459
2460	/* Clear UC indication and reset protection */
2461	set_pmd(pmdp, clear_pmd_bit(*pmdp, __pgprot(_SEGMENT_ENTRY_GMAP_UC)));
2462	gmap_protect_pmd(gmap, gaddr, pmdp, PROT_READ, 0);
2463	return true;
2464}
2465
2466/**
2467 * gmap_sync_dirty_log_pmd - set bitmap based on dirty status of segment
2468 * @gmap: pointer to guest address space
2469 * @bitmap: dirty bitmap for this pmd
2470 * @gaddr: virtual address in the guest address space
2471 * @vmaddr: virtual address in the host address space
2472 *
2473 * This function is assumed to be called with the guest_table_lock
2474 * held.
2475 */
2476void gmap_sync_dirty_log_pmd(struct gmap *gmap, unsigned long bitmap[4],
2477			     unsigned long gaddr, unsigned long vmaddr)
2478{
2479	int i;
2480	pmd_t *pmdp;
2481	pte_t *ptep;
2482	spinlock_t *ptl;
2483
2484	pmdp = gmap_pmd_op_walk(gmap, gaddr);
2485	if (!pmdp)
2486		return;
2487
2488	if (pmd_large(*pmdp)) {
2489		if (gmap_test_and_clear_dirty_pmd(gmap, pmdp, gaddr))
2490			bitmap_fill(bitmap, _PAGE_ENTRIES);
2491	} else {
2492		for (i = 0; i < _PAGE_ENTRIES; i++, vmaddr += PAGE_SIZE) {
2493			ptep = pte_alloc_map_lock(gmap->mm, pmdp, vmaddr, &ptl);
2494			if (!ptep)
2495				continue;
2496			if (ptep_test_and_clear_uc(gmap->mm, vmaddr, ptep))
2497				set_bit(i, bitmap);
2498			spin_unlock(ptl);
2499		}
2500	}
2501	gmap_pmd_op_end(gmap, pmdp);
2502}
2503EXPORT_SYMBOL_GPL(gmap_sync_dirty_log_pmd);
2504
2505#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2506static int thp_split_walk_pmd_entry(pmd_t *pmd, unsigned long addr,
2507				    unsigned long end, struct mm_walk *walk)
2508{
2509	struct vm_area_struct *vma = walk->vma;
2510
2511	split_huge_pmd(vma, pmd, addr);
2512	return 0;
2513}
2514
2515static const struct mm_walk_ops thp_split_walk_ops = {
2516	.pmd_entry	= thp_split_walk_pmd_entry,
 
2517};
2518
2519static inline void thp_split_mm(struct mm_struct *mm)
2520{
2521	struct vm_area_struct *vma;
2522	VMA_ITERATOR(vmi, mm, 0);
2523
2524	for_each_vma(vmi, vma) {
2525		vma->vm_flags &= ~VM_HUGEPAGE;
2526		vma->vm_flags |= VM_NOHUGEPAGE;
2527		walk_page_vma(vma, &thp_split_walk_ops, NULL);
2528	}
2529	mm->def_flags |= VM_NOHUGEPAGE;
2530}
2531#else
2532static inline void thp_split_mm(struct mm_struct *mm)
2533{
2534}
2535#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2536
2537/*
2538 * Remove all empty zero pages from the mapping for lazy refaulting
2539 * - This must be called after mm->context.has_pgste is set, to avoid
2540 *   future creation of zero pages
2541 * - This must be called after THP was enabled
2542 */
2543static int __zap_zero_pages(pmd_t *pmd, unsigned long start,
2544			   unsigned long end, struct mm_walk *walk)
2545{
2546	unsigned long addr;
2547
2548	for (addr = start; addr != end; addr += PAGE_SIZE) {
2549		pte_t *ptep;
2550		spinlock_t *ptl;
2551
2552		ptep = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
2553		if (is_zero_pfn(pte_pfn(*ptep)))
2554			ptep_xchg_direct(walk->mm, addr, ptep, __pte(_PAGE_INVALID));
2555		pte_unmap_unlock(ptep, ptl);
2556	}
2557	return 0;
2558}
2559
2560static const struct mm_walk_ops zap_zero_walk_ops = {
2561	.pmd_entry	= __zap_zero_pages,
2562};
2563
2564/*
2565 * switch on pgstes for its userspace process (for kvm)
2566 */
2567int s390_enable_sie(void)
2568{
2569	struct mm_struct *mm = current->mm;
2570
2571	/* Do we have pgstes? if yes, we are done */
2572	if (mm_has_pgste(mm))
2573		return 0;
2574	/* Fail if the page tables are 2K */
2575	if (!mm_alloc_pgste(mm))
2576		return -EINVAL;
2577	mmap_write_lock(mm);
2578	mm->context.has_pgste = 1;
2579	/* split thp mappings and disable thp for future mappings */
2580	thp_split_mm(mm);
2581	walk_page_range(mm, 0, TASK_SIZE, &zap_zero_walk_ops, NULL);
2582	mmap_write_unlock(mm);
2583	return 0;
2584}
2585EXPORT_SYMBOL_GPL(s390_enable_sie);
2586
2587int gmap_mark_unmergeable(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2588{
2589	struct mm_struct *mm = current->mm;
2590	struct vm_area_struct *vma;
2591	int ret;
2592	VMA_ITERATOR(vmi, mm, 0);
 
 
 
2593
2594	for_each_vma(vmi, vma) {
2595		ret = ksm_madvise(vma, vma->vm_start, vma->vm_end,
2596				  MADV_UNMERGEABLE, &vma->vm_flags);
2597		if (ret)
2598			return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2599	}
2600	mm->def_flags &= ~VM_MERGEABLE;
2601	return 0;
2602}
2603EXPORT_SYMBOL_GPL(gmap_mark_unmergeable);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2604
2605/*
2606 * Enable storage key handling from now on and initialize the storage
2607 * keys with the default key.
2608 */
2609static int __s390_enable_skey_pte(pte_t *pte, unsigned long addr,
2610				  unsigned long next, struct mm_walk *walk)
2611{
2612	/* Clear storage key */
2613	ptep_zap_key(walk->mm, addr, pte);
2614	return 0;
2615}
2616
2617/*
2618 * Give a chance to schedule after setting a key to 256 pages.
2619 * We only hold the mm lock, which is a rwsem and the kvm srcu.
2620 * Both can sleep.
2621 */
2622static int __s390_enable_skey_pmd(pmd_t *pmd, unsigned long addr,
2623				  unsigned long next, struct mm_walk *walk)
2624{
2625	cond_resched();
2626	return 0;
2627}
2628
2629static int __s390_enable_skey_hugetlb(pte_t *pte, unsigned long addr,
2630				      unsigned long hmask, unsigned long next,
2631				      struct mm_walk *walk)
2632{
2633	pmd_t *pmd = (pmd_t *)pte;
2634	unsigned long start, end;
2635	struct page *page = pmd_page(*pmd);
2636
2637	/*
2638	 * The write check makes sure we do not set a key on shared
2639	 * memory. This is needed as the walker does not differentiate
2640	 * between actual guest memory and the process executable or
2641	 * shared libraries.
2642	 */
2643	if (pmd_val(*pmd) & _SEGMENT_ENTRY_INVALID ||
2644	    !(pmd_val(*pmd) & _SEGMENT_ENTRY_WRITE))
2645		return 0;
2646
2647	start = pmd_val(*pmd) & HPAGE_MASK;
2648	end = start + HPAGE_SIZE - 1;
2649	__storage_key_init_range(start, end);
2650	set_bit(PG_arch_1, &page->flags);
2651	cond_resched();
2652	return 0;
2653}
2654
2655static const struct mm_walk_ops enable_skey_walk_ops = {
2656	.hugetlb_entry		= __s390_enable_skey_hugetlb,
2657	.pte_entry		= __s390_enable_skey_pte,
2658	.pmd_entry		= __s390_enable_skey_pmd,
 
2659};
2660
2661int s390_enable_skey(void)
2662{
2663	struct mm_struct *mm = current->mm;
2664	int rc = 0;
2665
2666	mmap_write_lock(mm);
2667	if (mm_uses_skeys(mm))
2668		goto out_up;
2669
2670	mm->context.uses_skeys = 1;
2671	rc = gmap_mark_unmergeable();
2672	if (rc) {
2673		mm->context.uses_skeys = 0;
2674		goto out_up;
2675	}
2676	walk_page_range(mm, 0, TASK_SIZE, &enable_skey_walk_ops, NULL);
2677
2678out_up:
2679	mmap_write_unlock(mm);
2680	return rc;
2681}
2682EXPORT_SYMBOL_GPL(s390_enable_skey);
2683
2684/*
2685 * Reset CMMA state, make all pages stable again.
2686 */
2687static int __s390_reset_cmma(pte_t *pte, unsigned long addr,
2688			     unsigned long next, struct mm_walk *walk)
2689{
2690	ptep_zap_unused(walk->mm, addr, pte, 1);
2691	return 0;
2692}
2693
2694static const struct mm_walk_ops reset_cmma_walk_ops = {
2695	.pte_entry		= __s390_reset_cmma,
 
2696};
2697
2698void s390_reset_cmma(struct mm_struct *mm)
2699{
2700	mmap_write_lock(mm);
2701	walk_page_range(mm, 0, TASK_SIZE, &reset_cmma_walk_ops, NULL);
2702	mmap_write_unlock(mm);
2703}
2704EXPORT_SYMBOL_GPL(s390_reset_cmma);
2705
2706#define GATHER_GET_PAGES 32
2707
2708struct reset_walk_state {
2709	unsigned long next;
2710	unsigned long count;
2711	unsigned long pfns[GATHER_GET_PAGES];
2712};
2713
2714static int s390_gather_pages(pte_t *ptep, unsigned long addr,
2715			     unsigned long next, struct mm_walk *walk)
2716{
2717	struct reset_walk_state *p = walk->private;
2718	pte_t pte = READ_ONCE(*ptep);
2719
2720	if (pte_present(pte)) {
2721		/* we have a reference from the mapping, take an extra one */
2722		get_page(phys_to_page(pte_val(pte)));
2723		p->pfns[p->count] = phys_to_pfn(pte_val(pte));
2724		p->next = next;
2725		p->count++;
2726	}
2727	return p->count >= GATHER_GET_PAGES;
2728}
2729
2730static const struct mm_walk_ops gather_pages_ops = {
2731	.pte_entry = s390_gather_pages,
 
2732};
2733
2734/*
2735 * Call the Destroy secure page UVC on each page in the given array of PFNs.
2736 * Each page needs to have an extra reference, which will be released here.
2737 */
2738void s390_uv_destroy_pfns(unsigned long count, unsigned long *pfns)
2739{
2740	unsigned long i;
2741
2742	for (i = 0; i < count; i++) {
2743		/* we always have an extra reference */
2744		uv_destroy_owned_page(pfn_to_phys(pfns[i]));
2745		/* get rid of the extra reference */
2746		put_page(pfn_to_page(pfns[i]));
2747		cond_resched();
2748	}
2749}
2750EXPORT_SYMBOL_GPL(s390_uv_destroy_pfns);
2751
2752/**
2753 * __s390_uv_destroy_range - Call the destroy secure page UVC on each page
2754 * in the given range of the given address space.
2755 * @mm: the mm to operate on
2756 * @start: the start of the range
2757 * @end: the end of the range
2758 * @interruptible: if not 0, stop when a fatal signal is received
2759 *
2760 * Walk the given range of the given address space and call the destroy
2761 * secure page UVC on each page. Optionally exit early if a fatal signal is
2762 * pending.
2763 *
2764 * Return: 0 on success, -EINTR if the function stopped before completing
2765 */
2766int __s390_uv_destroy_range(struct mm_struct *mm, unsigned long start,
2767			    unsigned long end, bool interruptible)
2768{
2769	struct reset_walk_state state = { .next = start };
2770	int r = 1;
2771
2772	while (r > 0) {
2773		state.count = 0;
2774		mmap_read_lock(mm);
2775		r = walk_page_range(mm, state.next, end, &gather_pages_ops, &state);
2776		mmap_read_unlock(mm);
2777		cond_resched();
2778		s390_uv_destroy_pfns(state.count, state.pfns);
2779		if (interruptible && fatal_signal_pending(current))
2780			return -EINTR;
2781	}
2782	return 0;
2783}
2784EXPORT_SYMBOL_GPL(__s390_uv_destroy_range);
2785
2786/**
2787 * s390_unlist_old_asce - Remove the topmost level of page tables from the
2788 * list of page tables of the gmap.
2789 * @gmap: the gmap whose table is to be removed
2790 *
2791 * On s390x, KVM keeps a list of all pages containing the page tables of the
2792 * gmap (the CRST list). This list is used at tear down time to free all
2793 * pages that are now not needed anymore.
2794 *
2795 * This function removes the topmost page of the tree (the one pointed to by
2796 * the ASCE) from the CRST list.
2797 *
2798 * This means that it will not be freed when the VM is torn down, and needs
2799 * to be handled separately by the caller, unless a leak is actually
2800 * intended. Notice that this function will only remove the page from the
2801 * list, the page will still be used as a top level page table (and ASCE).
2802 */
2803void s390_unlist_old_asce(struct gmap *gmap)
2804{
2805	struct page *old;
2806
2807	old = virt_to_page(gmap->table);
2808	spin_lock(&gmap->guest_table_lock);
2809	list_del(&old->lru);
2810	/*
2811	 * Sometimes the topmost page might need to be "removed" multiple
2812	 * times, for example if the VM is rebooted into secure mode several
2813	 * times concurrently, or if s390_replace_asce fails after calling
2814	 * s390_remove_old_asce and is attempted again later. In that case
2815	 * the old asce has been removed from the list, and therefore it
2816	 * will not be freed when the VM terminates, but the ASCE is still
2817	 * in use and still pointed to.
2818	 * A subsequent call to replace_asce will follow the pointer and try
2819	 * to remove the same page from the list again.
2820	 * Therefore it's necessary that the page of the ASCE has valid
2821	 * pointers, so list_del can work (and do nothing) without
2822	 * dereferencing stale or invalid pointers.
2823	 */
2824	INIT_LIST_HEAD(&old->lru);
2825	spin_unlock(&gmap->guest_table_lock);
2826}
2827EXPORT_SYMBOL_GPL(s390_unlist_old_asce);
2828
2829/**
2830 * s390_replace_asce - Try to replace the current ASCE of a gmap with a copy
2831 * @gmap: the gmap whose ASCE needs to be replaced
2832 *
 
 
 
2833 * If the allocation of the new top level page table fails, the ASCE is not
2834 * replaced.
2835 * In any case, the old ASCE is always removed from the gmap CRST list.
2836 * Therefore the caller has to make sure to save a pointer to it
2837 * beforehand, unless a leak is actually intended.
2838 */
2839int s390_replace_asce(struct gmap *gmap)
2840{
2841	unsigned long asce;
2842	struct page *page;
2843	void *table;
2844
2845	s390_unlist_old_asce(gmap);
2846
2847	page = alloc_pages(GFP_KERNEL_ACCOUNT, CRST_ALLOC_ORDER);
 
 
 
 
2848	if (!page)
2849		return -ENOMEM;
 
2850	table = page_to_virt(page);
2851	memcpy(table, gmap->table, 1UL << (CRST_ALLOC_ORDER + PAGE_SHIFT));
2852
2853	/*
2854	 * The caller has to deal with the old ASCE, but here we make sure
2855	 * the new one is properly added to the CRST list, so that
2856	 * it will be freed when the VM is torn down.
2857	 */
2858	spin_lock(&gmap->guest_table_lock);
2859	list_add(&page->lru, &gmap->crst_list);
2860	spin_unlock(&gmap->guest_table_lock);
2861
2862	/* Set new table origin while preserving existing ASCE control bits */
2863	asce = (gmap->asce & ~_ASCE_ORIGIN) | __pa(table);
2864	WRITE_ONCE(gmap->asce, asce);
2865	WRITE_ONCE(gmap->mm->context.gmap_asce, asce);
2866	WRITE_ONCE(gmap->table, table);
2867
2868	return 0;
2869}
2870EXPORT_SYMBOL_GPL(s390_replace_asce);
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  KVM guest address space mapping code
   4 *
   5 *    Copyright IBM Corp. 2007, 2020
   6 *    Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
   7 *		 David Hildenbrand <david@redhat.com>
   8 *		 Janosch Frank <frankja@linux.vnet.ibm.com>
   9 */
  10
  11#include <linux/kernel.h>
  12#include <linux/pagewalk.h>
  13#include <linux/swap.h>
  14#include <linux/smp.h>
  15#include <linux/spinlock.h>
  16#include <linux/slab.h>
  17#include <linux/swapops.h>
  18#include <linux/ksm.h>
  19#include <linux/mman.h>
  20#include <linux/pgtable.h>
  21#include <asm/page-states.h>
  22#include <asm/pgalloc.h>
  23#include <asm/gmap.h>
  24#include <asm/page.h>
  25#include <asm/tlb.h>
  26
  27#define GMAP_SHADOW_FAKE_TABLE 1ULL
  28
  29static struct page *gmap_alloc_crst(void)
  30{
  31	struct page *page;
  32
  33	page = alloc_pages(GFP_KERNEL_ACCOUNT, CRST_ALLOC_ORDER);
  34	if (!page)
  35		return NULL;
  36	__arch_set_page_dat(page_to_virt(page), 1UL << CRST_ALLOC_ORDER);
  37	return page;
  38}
  39
  40/**
  41 * gmap_alloc - allocate and initialize a guest address space
  42 * @limit: maximum address of the gmap address space
  43 *
  44 * Returns a guest address space structure.
  45 */
  46static struct gmap *gmap_alloc(unsigned long limit)
  47{
  48	struct gmap *gmap;
  49	struct page *page;
  50	unsigned long *table;
  51	unsigned long etype, atype;
  52
  53	if (limit < _REGION3_SIZE) {
  54		limit = _REGION3_SIZE - 1;
  55		atype = _ASCE_TYPE_SEGMENT;
  56		etype = _SEGMENT_ENTRY_EMPTY;
  57	} else if (limit < _REGION2_SIZE) {
  58		limit = _REGION2_SIZE - 1;
  59		atype = _ASCE_TYPE_REGION3;
  60		etype = _REGION3_ENTRY_EMPTY;
  61	} else if (limit < _REGION1_SIZE) {
  62		limit = _REGION1_SIZE - 1;
  63		atype = _ASCE_TYPE_REGION2;
  64		etype = _REGION2_ENTRY_EMPTY;
  65	} else {
  66		limit = -1UL;
  67		atype = _ASCE_TYPE_REGION1;
  68		etype = _REGION1_ENTRY_EMPTY;
  69	}
  70	gmap = kzalloc(sizeof(struct gmap), GFP_KERNEL_ACCOUNT);
  71	if (!gmap)
  72		goto out;
  73	INIT_LIST_HEAD(&gmap->crst_list);
  74	INIT_LIST_HEAD(&gmap->children);
  75	INIT_LIST_HEAD(&gmap->pt_list);
  76	INIT_RADIX_TREE(&gmap->guest_to_host, GFP_KERNEL_ACCOUNT);
  77	INIT_RADIX_TREE(&gmap->host_to_guest, GFP_ATOMIC | __GFP_ACCOUNT);
  78	INIT_RADIX_TREE(&gmap->host_to_rmap, GFP_ATOMIC | __GFP_ACCOUNT);
  79	spin_lock_init(&gmap->guest_table_lock);
  80	spin_lock_init(&gmap->shadow_lock);
  81	refcount_set(&gmap->ref_count, 1);
  82	page = gmap_alloc_crst();
  83	if (!page)
  84		goto out_free;
  85	page->index = 0;
  86	list_add(&page->lru, &gmap->crst_list);
  87	table = page_to_virt(page);
  88	crst_table_init(table, etype);
  89	gmap->table = table;
  90	gmap->asce = atype | _ASCE_TABLE_LENGTH |
  91		_ASCE_USER_BITS | __pa(table);
  92	gmap->asce_end = limit;
  93	return gmap;
  94
  95out_free:
  96	kfree(gmap);
  97out:
  98	return NULL;
  99}
 100
 101/**
 102 * gmap_create - create a guest address space
 103 * @mm: pointer to the parent mm_struct
 104 * @limit: maximum size of the gmap address space
 105 *
 106 * Returns a guest address space structure.
 107 */
 108struct gmap *gmap_create(struct mm_struct *mm, unsigned long limit)
 109{
 110	struct gmap *gmap;
 111	unsigned long gmap_asce;
 112
 113	gmap = gmap_alloc(limit);
 114	if (!gmap)
 115		return NULL;
 116	gmap->mm = mm;
 117	spin_lock(&mm->context.lock);
 118	list_add_rcu(&gmap->list, &mm->context.gmap_list);
 119	if (list_is_singular(&mm->context.gmap_list))
 120		gmap_asce = gmap->asce;
 121	else
 122		gmap_asce = -1UL;
 123	WRITE_ONCE(mm->context.gmap_asce, gmap_asce);
 124	spin_unlock(&mm->context.lock);
 125	return gmap;
 126}
 127EXPORT_SYMBOL_GPL(gmap_create);
 128
 129static void gmap_flush_tlb(struct gmap *gmap)
 130{
 131	if (MACHINE_HAS_IDTE)
 132		__tlb_flush_idte(gmap->asce);
 133	else
 134		__tlb_flush_global();
 135}
 136
 137static void gmap_radix_tree_free(struct radix_tree_root *root)
 138{
 139	struct radix_tree_iter iter;
 140	unsigned long indices[16];
 141	unsigned long index;
 142	void __rcu **slot;
 143	int i, nr;
 144
 145	/* A radix tree is freed by deleting all of its entries */
 146	index = 0;
 147	do {
 148		nr = 0;
 149		radix_tree_for_each_slot(slot, root, &iter, index) {
 150			indices[nr] = iter.index;
 151			if (++nr == 16)
 152				break;
 153		}
 154		for (i = 0; i < nr; i++) {
 155			index = indices[i];
 156			radix_tree_delete(root, index);
 157		}
 158	} while (nr > 0);
 159}
 160
 161static void gmap_rmap_radix_tree_free(struct radix_tree_root *root)
 162{
 163	struct gmap_rmap *rmap, *rnext, *head;
 164	struct radix_tree_iter iter;
 165	unsigned long indices[16];
 166	unsigned long index;
 167	void __rcu **slot;
 168	int i, nr;
 169
 170	/* A radix tree is freed by deleting all of its entries */
 171	index = 0;
 172	do {
 173		nr = 0;
 174		radix_tree_for_each_slot(slot, root, &iter, index) {
 175			indices[nr] = iter.index;
 176			if (++nr == 16)
 177				break;
 178		}
 179		for (i = 0; i < nr; i++) {
 180			index = indices[i];
 181			head = radix_tree_delete(root, index);
 182			gmap_for_each_rmap_safe(rmap, rnext, head)
 183				kfree(rmap);
 184		}
 185	} while (nr > 0);
 186}
 187
 188/**
 189 * gmap_free - free a guest address space
 190 * @gmap: pointer to the guest address space structure
 191 *
 192 * No locks required. There are no references to this gmap anymore.
 193 */
 194static void gmap_free(struct gmap *gmap)
 195{
 196	struct page *page, *next;
 197
 198	/* Flush tlb of all gmaps (if not already done for shadows) */
 199	if (!(gmap_is_shadow(gmap) && gmap->removed))
 200		gmap_flush_tlb(gmap);
 201	/* Free all segment & region tables. */
 202	list_for_each_entry_safe(page, next, &gmap->crst_list, lru)
 203		__free_pages(page, CRST_ALLOC_ORDER);
 204	gmap_radix_tree_free(&gmap->guest_to_host);
 205	gmap_radix_tree_free(&gmap->host_to_guest);
 206
 207	/* Free additional data for a shadow gmap */
 208	if (gmap_is_shadow(gmap)) {
 209		struct ptdesc *ptdesc, *n;
 210
 211		/* Free all page tables. */
 212		list_for_each_entry_safe(ptdesc, n, &gmap->pt_list, pt_list)
 213			page_table_free_pgste(ptdesc);
 214		gmap_rmap_radix_tree_free(&gmap->host_to_rmap);
 215		/* Release reference to the parent */
 216		gmap_put(gmap->parent);
 217	}
 218
 219	kfree(gmap);
 220}
 221
 222/**
 223 * gmap_get - increase reference counter for guest address space
 224 * @gmap: pointer to the guest address space structure
 225 *
 226 * Returns the gmap pointer
 227 */
 228struct gmap *gmap_get(struct gmap *gmap)
 229{
 230	refcount_inc(&gmap->ref_count);
 231	return gmap;
 232}
 233EXPORT_SYMBOL_GPL(gmap_get);
 234
 235/**
 236 * gmap_put - decrease reference counter for guest address space
 237 * @gmap: pointer to the guest address space structure
 238 *
 239 * If the reference counter reaches zero the guest address space is freed.
 240 */
 241void gmap_put(struct gmap *gmap)
 242{
 243	if (refcount_dec_and_test(&gmap->ref_count))
 244		gmap_free(gmap);
 245}
 246EXPORT_SYMBOL_GPL(gmap_put);
 247
 248/**
 249 * gmap_remove - remove a guest address space but do not free it yet
 250 * @gmap: pointer to the guest address space structure
 251 */
 252void gmap_remove(struct gmap *gmap)
 253{
 254	struct gmap *sg, *next;
 255	unsigned long gmap_asce;
 256
 257	/* Remove all shadow gmaps linked to this gmap */
 258	if (!list_empty(&gmap->children)) {
 259		spin_lock(&gmap->shadow_lock);
 260		list_for_each_entry_safe(sg, next, &gmap->children, list) {
 261			list_del(&sg->list);
 262			gmap_put(sg);
 263		}
 264		spin_unlock(&gmap->shadow_lock);
 265	}
 266	/* Remove gmap from the pre-mm list */
 267	spin_lock(&gmap->mm->context.lock);
 268	list_del_rcu(&gmap->list);
 269	if (list_empty(&gmap->mm->context.gmap_list))
 270		gmap_asce = 0;
 271	else if (list_is_singular(&gmap->mm->context.gmap_list))
 272		gmap_asce = list_first_entry(&gmap->mm->context.gmap_list,
 273					     struct gmap, list)->asce;
 274	else
 275		gmap_asce = -1UL;
 276	WRITE_ONCE(gmap->mm->context.gmap_asce, gmap_asce);
 277	spin_unlock(&gmap->mm->context.lock);
 278	synchronize_rcu();
 279	/* Put reference */
 280	gmap_put(gmap);
 281}
 282EXPORT_SYMBOL_GPL(gmap_remove);
 283
 284/**
 285 * gmap_enable - switch primary space to the guest address space
 286 * @gmap: pointer to the guest address space structure
 287 */
 288void gmap_enable(struct gmap *gmap)
 289{
 290	S390_lowcore.gmap = (unsigned long) gmap;
 291}
 292EXPORT_SYMBOL_GPL(gmap_enable);
 293
 294/**
 295 * gmap_disable - switch back to the standard primary address space
 296 * @gmap: pointer to the guest address space structure
 297 */
 298void gmap_disable(struct gmap *gmap)
 299{
 300	S390_lowcore.gmap = 0UL;
 301}
 302EXPORT_SYMBOL_GPL(gmap_disable);
 303
 304/**
 305 * gmap_get_enabled - get a pointer to the currently enabled gmap
 306 *
 307 * Returns a pointer to the currently enabled gmap. 0 if none is enabled.
 308 */
 309struct gmap *gmap_get_enabled(void)
 310{
 311	return (struct gmap *) S390_lowcore.gmap;
 312}
 313EXPORT_SYMBOL_GPL(gmap_get_enabled);
 314
 315/*
 316 * gmap_alloc_table is assumed to be called with mmap_lock held
 317 */
 318static int gmap_alloc_table(struct gmap *gmap, unsigned long *table,
 319			    unsigned long init, unsigned long gaddr)
 320{
 321	struct page *page;
 322	unsigned long *new;
 323
 324	/* since we dont free the gmap table until gmap_free we can unlock */
 325	page = gmap_alloc_crst();
 326	if (!page)
 327		return -ENOMEM;
 328	new = page_to_virt(page);
 329	crst_table_init(new, init);
 330	spin_lock(&gmap->guest_table_lock);
 331	if (*table & _REGION_ENTRY_INVALID) {
 332		list_add(&page->lru, &gmap->crst_list);
 333		*table = __pa(new) | _REGION_ENTRY_LENGTH |
 334			(*table & _REGION_ENTRY_TYPE_MASK);
 335		page->index = gaddr;
 336		page = NULL;
 337	}
 338	spin_unlock(&gmap->guest_table_lock);
 339	if (page)
 340		__free_pages(page, CRST_ALLOC_ORDER);
 341	return 0;
 342}
 343
 344/**
 345 * __gmap_segment_gaddr - find virtual address from segment pointer
 346 * @entry: pointer to a segment table entry in the guest address space
 347 *
 348 * Returns the virtual address in the guest address space for the segment
 349 */
 350static unsigned long __gmap_segment_gaddr(unsigned long *entry)
 351{
 352	struct page *page;
 353	unsigned long offset;
 354
 355	offset = (unsigned long) entry / sizeof(unsigned long);
 356	offset = (offset & (PTRS_PER_PMD - 1)) * PMD_SIZE;
 357	page = pmd_pgtable_page((pmd_t *) entry);
 358	return page->index + offset;
 359}
 360
 361/**
 362 * __gmap_unlink_by_vmaddr - unlink a single segment via a host address
 363 * @gmap: pointer to the guest address space structure
 364 * @vmaddr: address in the host process address space
 365 *
 366 * Returns 1 if a TLB flush is required
 367 */
 368static int __gmap_unlink_by_vmaddr(struct gmap *gmap, unsigned long vmaddr)
 369{
 370	unsigned long *entry;
 371	int flush = 0;
 372
 373	BUG_ON(gmap_is_shadow(gmap));
 374	spin_lock(&gmap->guest_table_lock);
 375	entry = radix_tree_delete(&gmap->host_to_guest, vmaddr >> PMD_SHIFT);
 376	if (entry) {
 377		flush = (*entry != _SEGMENT_ENTRY_EMPTY);
 378		*entry = _SEGMENT_ENTRY_EMPTY;
 379	}
 380	spin_unlock(&gmap->guest_table_lock);
 381	return flush;
 382}
 383
 384/**
 385 * __gmap_unmap_by_gaddr - unmap a single segment via a guest address
 386 * @gmap: pointer to the guest address space structure
 387 * @gaddr: address in the guest address space
 388 *
 389 * Returns 1 if a TLB flush is required
 390 */
 391static int __gmap_unmap_by_gaddr(struct gmap *gmap, unsigned long gaddr)
 392{
 393	unsigned long vmaddr;
 394
 395	vmaddr = (unsigned long) radix_tree_delete(&gmap->guest_to_host,
 396						   gaddr >> PMD_SHIFT);
 397	return vmaddr ? __gmap_unlink_by_vmaddr(gmap, vmaddr) : 0;
 398}
 399
 400/**
 401 * gmap_unmap_segment - unmap segment from the guest address space
 402 * @gmap: pointer to the guest address space structure
 403 * @to: address in the guest address space
 404 * @len: length of the memory area to unmap
 405 *
 406 * Returns 0 if the unmap succeeded, -EINVAL if not.
 407 */
 408int gmap_unmap_segment(struct gmap *gmap, unsigned long to, unsigned long len)
 409{
 410	unsigned long off;
 411	int flush;
 412
 413	BUG_ON(gmap_is_shadow(gmap));
 414	if ((to | len) & (PMD_SIZE - 1))
 415		return -EINVAL;
 416	if (len == 0 || to + len < to)
 417		return -EINVAL;
 418
 419	flush = 0;
 420	mmap_write_lock(gmap->mm);
 421	for (off = 0; off < len; off += PMD_SIZE)
 422		flush |= __gmap_unmap_by_gaddr(gmap, to + off);
 423	mmap_write_unlock(gmap->mm);
 424	if (flush)
 425		gmap_flush_tlb(gmap);
 426	return 0;
 427}
 428EXPORT_SYMBOL_GPL(gmap_unmap_segment);
 429
 430/**
 431 * gmap_map_segment - map a segment to the guest address space
 432 * @gmap: pointer to the guest address space structure
 433 * @from: source address in the parent address space
 434 * @to: target address in the guest address space
 435 * @len: length of the memory area to map
 436 *
 437 * Returns 0 if the mmap succeeded, -EINVAL or -ENOMEM if not.
 438 */
 439int gmap_map_segment(struct gmap *gmap, unsigned long from,
 440		     unsigned long to, unsigned long len)
 441{
 442	unsigned long off;
 443	int flush;
 444
 445	BUG_ON(gmap_is_shadow(gmap));
 446	if ((from | to | len) & (PMD_SIZE - 1))
 447		return -EINVAL;
 448	if (len == 0 || from + len < from || to + len < to ||
 449	    from + len - 1 > TASK_SIZE_MAX || to + len - 1 > gmap->asce_end)
 450		return -EINVAL;
 451
 452	flush = 0;
 453	mmap_write_lock(gmap->mm);
 454	for (off = 0; off < len; off += PMD_SIZE) {
 455		/* Remove old translation */
 456		flush |= __gmap_unmap_by_gaddr(gmap, to + off);
 457		/* Store new translation */
 458		if (radix_tree_insert(&gmap->guest_to_host,
 459				      (to + off) >> PMD_SHIFT,
 460				      (void *) from + off))
 461			break;
 462	}
 463	mmap_write_unlock(gmap->mm);
 464	if (flush)
 465		gmap_flush_tlb(gmap);
 466	if (off >= len)
 467		return 0;
 468	gmap_unmap_segment(gmap, to, len);
 469	return -ENOMEM;
 470}
 471EXPORT_SYMBOL_GPL(gmap_map_segment);
 472
 473/**
 474 * __gmap_translate - translate a guest address to a user space address
 475 * @gmap: pointer to guest mapping meta data structure
 476 * @gaddr: guest address
 477 *
 478 * Returns user space address which corresponds to the guest address or
 479 * -EFAULT if no such mapping exists.
 480 * This function does not establish potentially missing page table entries.
 481 * The mmap_lock of the mm that belongs to the address space must be held
 482 * when this function gets called.
 483 *
 484 * Note: Can also be called for shadow gmaps.
 485 */
 486unsigned long __gmap_translate(struct gmap *gmap, unsigned long gaddr)
 487{
 488	unsigned long vmaddr;
 489
 490	vmaddr = (unsigned long)
 491		radix_tree_lookup(&gmap->guest_to_host, gaddr >> PMD_SHIFT);
 492	/* Note: guest_to_host is empty for a shadow gmap */
 493	return vmaddr ? (vmaddr | (gaddr & ~PMD_MASK)) : -EFAULT;
 494}
 495EXPORT_SYMBOL_GPL(__gmap_translate);
 496
 497/**
 498 * gmap_translate - translate a guest address to a user space address
 499 * @gmap: pointer to guest mapping meta data structure
 500 * @gaddr: guest address
 501 *
 502 * Returns user space address which corresponds to the guest address or
 503 * -EFAULT if no such mapping exists.
 504 * This function does not establish potentially missing page table entries.
 505 */
 506unsigned long gmap_translate(struct gmap *gmap, unsigned long gaddr)
 507{
 508	unsigned long rc;
 509
 510	mmap_read_lock(gmap->mm);
 511	rc = __gmap_translate(gmap, gaddr);
 512	mmap_read_unlock(gmap->mm);
 513	return rc;
 514}
 515EXPORT_SYMBOL_GPL(gmap_translate);
 516
 517/**
 518 * gmap_unlink - disconnect a page table from the gmap shadow tables
 519 * @mm: pointer to the parent mm_struct
 520 * @table: pointer to the host page table
 521 * @vmaddr: vm address associated with the host page table
 522 */
 523void gmap_unlink(struct mm_struct *mm, unsigned long *table,
 524		 unsigned long vmaddr)
 525{
 526	struct gmap *gmap;
 527	int flush;
 528
 529	rcu_read_lock();
 530	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
 531		flush = __gmap_unlink_by_vmaddr(gmap, vmaddr);
 532		if (flush)
 533			gmap_flush_tlb(gmap);
 534	}
 535	rcu_read_unlock();
 536}
 537
 538static void gmap_pmdp_xchg(struct gmap *gmap, pmd_t *old, pmd_t new,
 539			   unsigned long gaddr);
 540
 541/**
 542 * __gmap_link - set up shadow page tables to connect a host to a guest address
 543 * @gmap: pointer to guest mapping meta data structure
 544 * @gaddr: guest address
 545 * @vmaddr: vm address
 546 *
 547 * Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
 548 * if the vm address is already mapped to a different guest segment.
 549 * The mmap_lock of the mm that belongs to the address space must be held
 550 * when this function gets called.
 551 */
 552int __gmap_link(struct gmap *gmap, unsigned long gaddr, unsigned long vmaddr)
 553{
 554	struct mm_struct *mm;
 555	unsigned long *table;
 556	spinlock_t *ptl;
 557	pgd_t *pgd;
 558	p4d_t *p4d;
 559	pud_t *pud;
 560	pmd_t *pmd;
 561	u64 unprot;
 562	int rc;
 563
 564	BUG_ON(gmap_is_shadow(gmap));
 565	/* Create higher level tables in the gmap page table */
 566	table = gmap->table;
 567	if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION1) {
 568		table += (gaddr & _REGION1_INDEX) >> _REGION1_SHIFT;
 569		if ((*table & _REGION_ENTRY_INVALID) &&
 570		    gmap_alloc_table(gmap, table, _REGION2_ENTRY_EMPTY,
 571				     gaddr & _REGION1_MASK))
 572			return -ENOMEM;
 573		table = __va(*table & _REGION_ENTRY_ORIGIN);
 574	}
 575	if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION2) {
 576		table += (gaddr & _REGION2_INDEX) >> _REGION2_SHIFT;
 577		if ((*table & _REGION_ENTRY_INVALID) &&
 578		    gmap_alloc_table(gmap, table, _REGION3_ENTRY_EMPTY,
 579				     gaddr & _REGION2_MASK))
 580			return -ENOMEM;
 581		table = __va(*table & _REGION_ENTRY_ORIGIN);
 582	}
 583	if ((gmap->asce & _ASCE_TYPE_MASK) >= _ASCE_TYPE_REGION3) {
 584		table += (gaddr & _REGION3_INDEX) >> _REGION3_SHIFT;
 585		if ((*table & _REGION_ENTRY_INVALID) &&
 586		    gmap_alloc_table(gmap, table, _SEGMENT_ENTRY_EMPTY,
 587				     gaddr & _REGION3_MASK))
 588			return -ENOMEM;
 589		table = __va(*table & _REGION_ENTRY_ORIGIN);
 590	}
 591	table += (gaddr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
 592	/* Walk the parent mm page table */
 593	mm = gmap->mm;
 594	pgd = pgd_offset(mm, vmaddr);
 595	VM_BUG_ON(pgd_none(*pgd));
 596	p4d = p4d_offset(pgd, vmaddr);
 597	VM_BUG_ON(p4d_none(*p4d));
 598	pud = pud_offset(p4d, vmaddr);
 599	VM_BUG_ON(pud_none(*pud));
 600	/* large puds cannot yet be handled */
 601	if (pud_leaf(*pud))
 602		return -EFAULT;
 603	pmd = pmd_offset(pud, vmaddr);
 604	VM_BUG_ON(pmd_none(*pmd));
 605	/* Are we allowed to use huge pages? */
 606	if (pmd_leaf(*pmd) && !gmap->mm->context.allow_gmap_hpage_1m)
 607		return -EFAULT;
 608	/* Link gmap segment table entry location to page table. */
 609	rc = radix_tree_preload(GFP_KERNEL_ACCOUNT);
 610	if (rc)
 611		return rc;
 612	ptl = pmd_lock(mm, pmd);
 613	spin_lock(&gmap->guest_table_lock);
 614	if (*table == _SEGMENT_ENTRY_EMPTY) {
 615		rc = radix_tree_insert(&gmap->host_to_guest,
 616				       vmaddr >> PMD_SHIFT, table);
 617		if (!rc) {
 618			if (pmd_leaf(*pmd)) {
 619				*table = (pmd_val(*pmd) &
 620					  _SEGMENT_ENTRY_HARDWARE_BITS_LARGE)
 621					| _SEGMENT_ENTRY_GMAP_UC;
 622			} else
 623				*table = pmd_val(*pmd) &
 624					_SEGMENT_ENTRY_HARDWARE_BITS;
 625		}
 626	} else if (*table & _SEGMENT_ENTRY_PROTECT &&
 627		   !(pmd_val(*pmd) & _SEGMENT_ENTRY_PROTECT)) {
 628		unprot = (u64)*table;
 629		unprot &= ~_SEGMENT_ENTRY_PROTECT;
 630		unprot |= _SEGMENT_ENTRY_GMAP_UC;
 631		gmap_pmdp_xchg(gmap, (pmd_t *)table, __pmd(unprot), gaddr);
 632	}
 633	spin_unlock(&gmap->guest_table_lock);
 634	spin_unlock(ptl);
 635	radix_tree_preload_end();
 636	return rc;
 637}
 638
 639/**
 640 * gmap_fault - resolve a fault on a guest address
 641 * @gmap: pointer to guest mapping meta data structure
 642 * @gaddr: guest address
 643 * @fault_flags: flags to pass down to handle_mm_fault()
 644 *
 645 * Returns 0 on success, -ENOMEM for out of memory conditions, and -EFAULT
 646 * if the vm address is already mapped to a different guest segment.
 647 */
 648int gmap_fault(struct gmap *gmap, unsigned long gaddr,
 649	       unsigned int fault_flags)
 650{
 651	unsigned long vmaddr;
 652	int rc;
 653	bool unlocked;
 654
 655	mmap_read_lock(gmap->mm);
 656
 657retry:
 658	unlocked = false;
 659	vmaddr = __gmap_translate(gmap, gaddr);
 660	if (IS_ERR_VALUE(vmaddr)) {
 661		rc = vmaddr;
 662		goto out_up;
 663	}
 664	if (fixup_user_fault(gmap->mm, vmaddr, fault_flags,
 665			     &unlocked)) {
 666		rc = -EFAULT;
 667		goto out_up;
 668	}
 669	/*
 670	 * In the case that fixup_user_fault unlocked the mmap_lock during
 671	 * faultin redo __gmap_translate to not race with a map/unmap_segment.
 672	 */
 673	if (unlocked)
 674		goto retry;
 675
 676	rc = __gmap_link(gmap, gaddr, vmaddr);
 677out_up:
 678	mmap_read_unlock(gmap->mm);
 679	return rc;
 680}
 681EXPORT_SYMBOL_GPL(gmap_fault);
 682
 683/*
 684 * this function is assumed to be called with mmap_lock held
 685 */
 686void __gmap_zap(struct gmap *gmap, unsigned long gaddr)
 687{
 688	struct vm_area_struct *vma;
 689	unsigned long vmaddr;
 690	spinlock_t *ptl;
 691	pte_t *ptep;
 692
 693	/* Find the vm address for the guest address */
 694	vmaddr = (unsigned long) radix_tree_lookup(&gmap->guest_to_host,
 695						   gaddr >> PMD_SHIFT);
 696	if (vmaddr) {
 697		vmaddr |= gaddr & ~PMD_MASK;
 698
 699		vma = vma_lookup(gmap->mm, vmaddr);
 700		if (!vma || is_vm_hugetlb_page(vma))
 701			return;
 702
 703		/* Get pointer to the page table entry */
 704		ptep = get_locked_pte(gmap->mm, vmaddr, &ptl);
 705		if (likely(ptep)) {
 706			ptep_zap_unused(gmap->mm, vmaddr, ptep, 0);
 707			pte_unmap_unlock(ptep, ptl);
 708		}
 709	}
 710}
 711EXPORT_SYMBOL_GPL(__gmap_zap);
 712
 713void gmap_discard(struct gmap *gmap, unsigned long from, unsigned long to)
 714{
 715	unsigned long gaddr, vmaddr, size;
 716	struct vm_area_struct *vma;
 717
 718	mmap_read_lock(gmap->mm);
 719	for (gaddr = from; gaddr < to;
 720	     gaddr = (gaddr + PMD_SIZE) & PMD_MASK) {
 721		/* Find the vm address for the guest address */
 722		vmaddr = (unsigned long)
 723			radix_tree_lookup(&gmap->guest_to_host,
 724					  gaddr >> PMD_SHIFT);
 725		if (!vmaddr)
 726			continue;
 727		vmaddr |= gaddr & ~PMD_MASK;
 728		/* Find vma in the parent mm */
 729		vma = find_vma(gmap->mm, vmaddr);
 730		if (!vma)
 731			continue;
 732		/*
 733		 * We do not discard pages that are backed by
 734		 * hugetlbfs, so we don't have to refault them.
 735		 */
 736		if (is_vm_hugetlb_page(vma))
 737			continue;
 738		size = min(to - gaddr, PMD_SIZE - (gaddr & ~PMD_MASK));
 739		zap_page_range_single(vma, vmaddr, size, NULL);
 740	}
 741	mmap_read_unlock(gmap->mm);
 742}
 743EXPORT_SYMBOL_GPL(gmap_discard);
 744
 745static LIST_HEAD(gmap_notifier_list);
 746static DEFINE_SPINLOCK(gmap_notifier_lock);
 747
 748/**
 749 * gmap_register_pte_notifier - register a pte invalidation callback
 750 * @nb: pointer to the gmap notifier block
 751 */
 752void gmap_register_pte_notifier(struct gmap_notifier *nb)
 753{
 754	spin_lock(&gmap_notifier_lock);
 755	list_add_rcu(&nb->list, &gmap_notifier_list);
 756	spin_unlock(&gmap_notifier_lock);
 757}
 758EXPORT_SYMBOL_GPL(gmap_register_pte_notifier);
 759
 760/**
 761 * gmap_unregister_pte_notifier - remove a pte invalidation callback
 762 * @nb: pointer to the gmap notifier block
 763 */
 764void gmap_unregister_pte_notifier(struct gmap_notifier *nb)
 765{
 766	spin_lock(&gmap_notifier_lock);
 767	list_del_rcu(&nb->list);
 768	spin_unlock(&gmap_notifier_lock);
 769	synchronize_rcu();
 770}
 771EXPORT_SYMBOL_GPL(gmap_unregister_pte_notifier);
 772
 773/**
 774 * gmap_call_notifier - call all registered invalidation callbacks
 775 * @gmap: pointer to guest mapping meta data structure
 776 * @start: start virtual address in the guest address space
 777 * @end: end virtual address in the guest address space
 778 */
 779static void gmap_call_notifier(struct gmap *gmap, unsigned long start,
 780			       unsigned long end)
 781{
 782	struct gmap_notifier *nb;
 783
 784	list_for_each_entry(nb, &gmap_notifier_list, list)
 785		nb->notifier_call(gmap, start, end);
 786}
 787
 788/**
 789 * gmap_table_walk - walk the gmap page tables
 790 * @gmap: pointer to guest mapping meta data structure
 791 * @gaddr: virtual address in the guest address space
 792 * @level: page table level to stop at
 793 *
 794 * Returns a table entry pointer for the given guest address and @level
 795 * @level=0 : returns a pointer to a page table table entry (or NULL)
 796 * @level=1 : returns a pointer to a segment table entry (or NULL)
 797 * @level=2 : returns a pointer to a region-3 table entry (or NULL)
 798 * @level=3 : returns a pointer to a region-2 table entry (or NULL)
 799 * @level=4 : returns a pointer to a region-1 table entry (or NULL)
 800 *
 801 * Returns NULL if the gmap page tables could not be walked to the
 802 * requested level.
 803 *
 804 * Note: Can also be called for shadow gmaps.
 805 */
 806static inline unsigned long *gmap_table_walk(struct gmap *gmap,
 807					     unsigned long gaddr, int level)
 808{
 809	const int asce_type = gmap->asce & _ASCE_TYPE_MASK;
 810	unsigned long *table = gmap->table;
 811
 812	if (gmap_is_shadow(gmap) && gmap->removed)
 813		return NULL;
 814
 815	if (WARN_ON_ONCE(level > (asce_type >> 2) + 1))
 816		return NULL;
 817
 818	if (asce_type != _ASCE_TYPE_REGION1 &&
 819	    gaddr & (-1UL << (31 + (asce_type >> 2) * 11)))
 820		return NULL;
 821
 822	switch (asce_type) {
 823	case _ASCE_TYPE_REGION1:
 824		table += (gaddr & _REGION1_INDEX) >> _REGION1_SHIFT;
 825		if (level == 4)
 826			break;
 827		if (*table & _REGION_ENTRY_INVALID)
 828			return NULL;
 829		table = __va(*table & _REGION_ENTRY_ORIGIN);
 830		fallthrough;
 831	case _ASCE_TYPE_REGION2:
 832		table += (gaddr & _REGION2_INDEX) >> _REGION2_SHIFT;
 833		if (level == 3)
 834			break;
 835		if (*table & _REGION_ENTRY_INVALID)
 836			return NULL;
 837		table = __va(*table & _REGION_ENTRY_ORIGIN);
 838		fallthrough;
 839	case _ASCE_TYPE_REGION3:
 840		table += (gaddr & _REGION3_INDEX) >> _REGION3_SHIFT;
 841		if (level == 2)
 842			break;
 843		if (*table & _REGION_ENTRY_INVALID)
 844			return NULL;
 845		table = __va(*table & _REGION_ENTRY_ORIGIN);
 846		fallthrough;
 847	case _ASCE_TYPE_SEGMENT:
 848		table += (gaddr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
 849		if (level == 1)
 850			break;
 851		if (*table & _REGION_ENTRY_INVALID)
 852			return NULL;
 853		table = __va(*table & _SEGMENT_ENTRY_ORIGIN);
 854		table += (gaddr & _PAGE_INDEX) >> _PAGE_SHIFT;
 855	}
 856	return table;
 857}
 858
 859/**
 860 * gmap_pte_op_walk - walk the gmap page table, get the page table lock
 861 *		      and return the pte pointer
 862 * @gmap: pointer to guest mapping meta data structure
 863 * @gaddr: virtual address in the guest address space
 864 * @ptl: pointer to the spinlock pointer
 865 *
 866 * Returns a pointer to the locked pte for a guest address, or NULL
 867 */
 868static pte_t *gmap_pte_op_walk(struct gmap *gmap, unsigned long gaddr,
 869			       spinlock_t **ptl)
 870{
 871	unsigned long *table;
 872
 873	BUG_ON(gmap_is_shadow(gmap));
 874	/* Walk the gmap page table, lock and get pte pointer */
 875	table = gmap_table_walk(gmap, gaddr, 1); /* get segment pointer */
 876	if (!table || *table & _SEGMENT_ENTRY_INVALID)
 877		return NULL;
 878	return pte_alloc_map_lock(gmap->mm, (pmd_t *) table, gaddr, ptl);
 879}
 880
 881/**
 882 * gmap_pte_op_fixup - force a page in and connect the gmap page table
 883 * @gmap: pointer to guest mapping meta data structure
 884 * @gaddr: virtual address in the guest address space
 885 * @vmaddr: address in the host process address space
 886 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
 887 *
 888 * Returns 0 if the caller can retry __gmap_translate (might fail again),
 889 * -ENOMEM if out of memory and -EFAULT if anything goes wrong while fixing
 890 * up or connecting the gmap page table.
 891 */
 892static int gmap_pte_op_fixup(struct gmap *gmap, unsigned long gaddr,
 893			     unsigned long vmaddr, int prot)
 894{
 895	struct mm_struct *mm = gmap->mm;
 896	unsigned int fault_flags;
 897	bool unlocked = false;
 898
 899	BUG_ON(gmap_is_shadow(gmap));
 900	fault_flags = (prot == PROT_WRITE) ? FAULT_FLAG_WRITE : 0;
 901	if (fixup_user_fault(mm, vmaddr, fault_flags, &unlocked))
 902		return -EFAULT;
 903	if (unlocked)
 904		/* lost mmap_lock, caller has to retry __gmap_translate */
 905		return 0;
 906	/* Connect the page tables */
 907	return __gmap_link(gmap, gaddr, vmaddr);
 908}
 909
 910/**
 911 * gmap_pte_op_end - release the page table lock
 912 * @ptep: pointer to the locked pte
 913 * @ptl: pointer to the page table spinlock
 914 */
 915static void gmap_pte_op_end(pte_t *ptep, spinlock_t *ptl)
 916{
 917	pte_unmap_unlock(ptep, ptl);
 
 918}
 919
 920/**
 921 * gmap_pmd_op_walk - walk the gmap tables, get the guest table lock
 922 *		      and return the pmd pointer
 923 * @gmap: pointer to guest mapping meta data structure
 924 * @gaddr: virtual address in the guest address space
 925 *
 926 * Returns a pointer to the pmd for a guest address, or NULL
 927 */
 928static inline pmd_t *gmap_pmd_op_walk(struct gmap *gmap, unsigned long gaddr)
 929{
 930	pmd_t *pmdp;
 931
 932	BUG_ON(gmap_is_shadow(gmap));
 933	pmdp = (pmd_t *) gmap_table_walk(gmap, gaddr, 1);
 934	if (!pmdp)
 935		return NULL;
 936
 937	/* without huge pages, there is no need to take the table lock */
 938	if (!gmap->mm->context.allow_gmap_hpage_1m)
 939		return pmd_none(*pmdp) ? NULL : pmdp;
 940
 941	spin_lock(&gmap->guest_table_lock);
 942	if (pmd_none(*pmdp)) {
 943		spin_unlock(&gmap->guest_table_lock);
 944		return NULL;
 945	}
 946
 947	/* 4k page table entries are locked via the pte (pte_alloc_map_lock). */
 948	if (!pmd_leaf(*pmdp))
 949		spin_unlock(&gmap->guest_table_lock);
 950	return pmdp;
 951}
 952
 953/**
 954 * gmap_pmd_op_end - release the guest_table_lock if needed
 955 * @gmap: pointer to the guest mapping meta data structure
 956 * @pmdp: pointer to the pmd
 957 */
 958static inline void gmap_pmd_op_end(struct gmap *gmap, pmd_t *pmdp)
 959{
 960	if (pmd_leaf(*pmdp))
 961		spin_unlock(&gmap->guest_table_lock);
 962}
 963
 964/*
 965 * gmap_protect_pmd - remove access rights to memory and set pmd notification bits
 966 * @pmdp: pointer to the pmd to be protected
 967 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
 968 * @bits: notification bits to set
 969 *
 970 * Returns:
 971 * 0 if successfully protected
 972 * -EAGAIN if a fixup is needed
 973 * -EINVAL if unsupported notifier bits have been specified
 974 *
 975 * Expected to be called with sg->mm->mmap_lock in read and
 976 * guest_table_lock held.
 977 */
 978static int gmap_protect_pmd(struct gmap *gmap, unsigned long gaddr,
 979			    pmd_t *pmdp, int prot, unsigned long bits)
 980{
 981	int pmd_i = pmd_val(*pmdp) & _SEGMENT_ENTRY_INVALID;
 982	int pmd_p = pmd_val(*pmdp) & _SEGMENT_ENTRY_PROTECT;
 983	pmd_t new = *pmdp;
 984
 985	/* Fixup needed */
 986	if ((pmd_i && (prot != PROT_NONE)) || (pmd_p && (prot == PROT_WRITE)))
 987		return -EAGAIN;
 988
 989	if (prot == PROT_NONE && !pmd_i) {
 990		new = set_pmd_bit(new, __pgprot(_SEGMENT_ENTRY_INVALID));
 991		gmap_pmdp_xchg(gmap, pmdp, new, gaddr);
 992	}
 993
 994	if (prot == PROT_READ && !pmd_p) {
 995		new = clear_pmd_bit(new, __pgprot(_SEGMENT_ENTRY_INVALID));
 996		new = set_pmd_bit(new, __pgprot(_SEGMENT_ENTRY_PROTECT));
 997		gmap_pmdp_xchg(gmap, pmdp, new, gaddr);
 998	}
 999
1000	if (bits & GMAP_NOTIFY_MPROT)
1001		set_pmd(pmdp, set_pmd_bit(*pmdp, __pgprot(_SEGMENT_ENTRY_GMAP_IN)));
1002
1003	/* Shadow GMAP protection needs split PMDs */
1004	if (bits & GMAP_NOTIFY_SHADOW)
1005		return -EINVAL;
1006
1007	return 0;
1008}
1009
1010/*
1011 * gmap_protect_pte - remove access rights to memory and set pgste bits
1012 * @gmap: pointer to guest mapping meta data structure
1013 * @gaddr: virtual address in the guest address space
1014 * @pmdp: pointer to the pmd associated with the pte
1015 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
1016 * @bits: notification bits to set
1017 *
1018 * Returns 0 if successfully protected, -ENOMEM if out of memory and
1019 * -EAGAIN if a fixup is needed.
1020 *
1021 * Expected to be called with sg->mm->mmap_lock in read
1022 */
1023static int gmap_protect_pte(struct gmap *gmap, unsigned long gaddr,
1024			    pmd_t *pmdp, int prot, unsigned long bits)
1025{
1026	int rc;
1027	pte_t *ptep;
1028	spinlock_t *ptl;
1029	unsigned long pbits = 0;
1030
1031	if (pmd_val(*pmdp) & _SEGMENT_ENTRY_INVALID)
1032		return -EAGAIN;
1033
1034	ptep = pte_alloc_map_lock(gmap->mm, pmdp, gaddr, &ptl);
1035	if (!ptep)
1036		return -ENOMEM;
1037
1038	pbits |= (bits & GMAP_NOTIFY_MPROT) ? PGSTE_IN_BIT : 0;
1039	pbits |= (bits & GMAP_NOTIFY_SHADOW) ? PGSTE_VSIE_BIT : 0;
1040	/* Protect and unlock. */
1041	rc = ptep_force_prot(gmap->mm, gaddr, ptep, prot, pbits);
1042	gmap_pte_op_end(ptep, ptl);
1043	return rc;
1044}
1045
1046/*
1047 * gmap_protect_range - remove access rights to memory and set pgste bits
1048 * @gmap: pointer to guest mapping meta data structure
1049 * @gaddr: virtual address in the guest address space
1050 * @len: size of area
1051 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
1052 * @bits: pgste notification bits to set
1053 *
1054 * Returns 0 if successfully protected, -ENOMEM if out of memory and
1055 * -EFAULT if gaddr is invalid (or mapping for shadows is missing).
1056 *
1057 * Called with sg->mm->mmap_lock in read.
1058 */
1059static int gmap_protect_range(struct gmap *gmap, unsigned long gaddr,
1060			      unsigned long len, int prot, unsigned long bits)
1061{
1062	unsigned long vmaddr, dist;
1063	pmd_t *pmdp;
1064	int rc;
1065
1066	BUG_ON(gmap_is_shadow(gmap));
1067	while (len) {
1068		rc = -EAGAIN;
1069		pmdp = gmap_pmd_op_walk(gmap, gaddr);
1070		if (pmdp) {
1071			if (!pmd_leaf(*pmdp)) {
1072				rc = gmap_protect_pte(gmap, gaddr, pmdp, prot,
1073						      bits);
1074				if (!rc) {
1075					len -= PAGE_SIZE;
1076					gaddr += PAGE_SIZE;
1077				}
1078			} else {
1079				rc = gmap_protect_pmd(gmap, gaddr, pmdp, prot,
1080						      bits);
1081				if (!rc) {
1082					dist = HPAGE_SIZE - (gaddr & ~HPAGE_MASK);
1083					len = len < dist ? 0 : len - dist;
1084					gaddr = (gaddr & HPAGE_MASK) + HPAGE_SIZE;
1085				}
1086			}
1087			gmap_pmd_op_end(gmap, pmdp);
1088		}
1089		if (rc) {
1090			if (rc == -EINVAL)
1091				return rc;
1092
1093			/* -EAGAIN, fixup of userspace mm and gmap */
1094			vmaddr = __gmap_translate(gmap, gaddr);
1095			if (IS_ERR_VALUE(vmaddr))
1096				return vmaddr;
1097			rc = gmap_pte_op_fixup(gmap, gaddr, vmaddr, prot);
1098			if (rc)
1099				return rc;
1100		}
1101	}
1102	return 0;
1103}
1104
1105/**
1106 * gmap_mprotect_notify - change access rights for a range of ptes and
1107 *                        call the notifier if any pte changes again
1108 * @gmap: pointer to guest mapping meta data structure
1109 * @gaddr: virtual address in the guest address space
1110 * @len: size of area
1111 * @prot: indicates access rights: PROT_NONE, PROT_READ or PROT_WRITE
1112 *
1113 * Returns 0 if for each page in the given range a gmap mapping exists,
1114 * the new access rights could be set and the notifier could be armed.
1115 * If the gmap mapping is missing for one or more pages -EFAULT is
1116 * returned. If no memory could be allocated -ENOMEM is returned.
1117 * This function establishes missing page table entries.
1118 */
1119int gmap_mprotect_notify(struct gmap *gmap, unsigned long gaddr,
1120			 unsigned long len, int prot)
1121{
1122	int rc;
1123
1124	if ((gaddr & ~PAGE_MASK) || (len & ~PAGE_MASK) || gmap_is_shadow(gmap))
1125		return -EINVAL;
1126	if (!MACHINE_HAS_ESOP && prot == PROT_READ)
1127		return -EINVAL;
1128	mmap_read_lock(gmap->mm);
1129	rc = gmap_protect_range(gmap, gaddr, len, prot, GMAP_NOTIFY_MPROT);
1130	mmap_read_unlock(gmap->mm);
1131	return rc;
1132}
1133EXPORT_SYMBOL_GPL(gmap_mprotect_notify);
1134
1135/**
1136 * gmap_read_table - get an unsigned long value from a guest page table using
1137 *                   absolute addressing, without marking the page referenced.
1138 * @gmap: pointer to guest mapping meta data structure
1139 * @gaddr: virtual address in the guest address space
1140 * @val: pointer to the unsigned long value to return
1141 *
1142 * Returns 0 if the value was read, -ENOMEM if out of memory and -EFAULT
1143 * if reading using the virtual address failed. -EINVAL if called on a gmap
1144 * shadow.
1145 *
1146 * Called with gmap->mm->mmap_lock in read.
1147 */
1148int gmap_read_table(struct gmap *gmap, unsigned long gaddr, unsigned long *val)
1149{
1150	unsigned long address, vmaddr;
1151	spinlock_t *ptl;
1152	pte_t *ptep, pte;
1153	int rc;
1154
1155	if (gmap_is_shadow(gmap))
1156		return -EINVAL;
1157
1158	while (1) {
1159		rc = -EAGAIN;
1160		ptep = gmap_pte_op_walk(gmap, gaddr, &ptl);
1161		if (ptep) {
1162			pte = *ptep;
1163			if (pte_present(pte) && (pte_val(pte) & _PAGE_READ)) {
1164				address = pte_val(pte) & PAGE_MASK;
1165				address += gaddr & ~PAGE_MASK;
1166				*val = *(unsigned long *)__va(address);
1167				set_pte(ptep, set_pte_bit(*ptep, __pgprot(_PAGE_YOUNG)));
1168				/* Do *NOT* clear the _PAGE_INVALID bit! */
1169				rc = 0;
1170			}
1171			gmap_pte_op_end(ptep, ptl);
1172		}
1173		if (!rc)
1174			break;
1175		vmaddr = __gmap_translate(gmap, gaddr);
1176		if (IS_ERR_VALUE(vmaddr)) {
1177			rc = vmaddr;
1178			break;
1179		}
1180		rc = gmap_pte_op_fixup(gmap, gaddr, vmaddr, PROT_READ);
1181		if (rc)
1182			break;
1183	}
1184	return rc;
1185}
1186EXPORT_SYMBOL_GPL(gmap_read_table);
1187
1188/**
1189 * gmap_insert_rmap - add a rmap to the host_to_rmap radix tree
1190 * @sg: pointer to the shadow guest address space structure
1191 * @vmaddr: vm address associated with the rmap
1192 * @rmap: pointer to the rmap structure
1193 *
1194 * Called with the sg->guest_table_lock
1195 */
1196static inline void gmap_insert_rmap(struct gmap *sg, unsigned long vmaddr,
1197				    struct gmap_rmap *rmap)
1198{
1199	struct gmap_rmap *temp;
1200	void __rcu **slot;
1201
1202	BUG_ON(!gmap_is_shadow(sg));
1203	slot = radix_tree_lookup_slot(&sg->host_to_rmap, vmaddr >> PAGE_SHIFT);
1204	if (slot) {
1205		rmap->next = radix_tree_deref_slot_protected(slot,
1206							&sg->guest_table_lock);
1207		for (temp = rmap->next; temp; temp = temp->next) {
1208			if (temp->raddr == rmap->raddr) {
1209				kfree(rmap);
1210				return;
1211			}
1212		}
1213		radix_tree_replace_slot(&sg->host_to_rmap, slot, rmap);
1214	} else {
1215		rmap->next = NULL;
1216		radix_tree_insert(&sg->host_to_rmap, vmaddr >> PAGE_SHIFT,
1217				  rmap);
1218	}
1219}
1220
1221/**
1222 * gmap_protect_rmap - restrict access rights to memory (RO) and create an rmap
1223 * @sg: pointer to the shadow guest address space structure
1224 * @raddr: rmap address in the shadow gmap
1225 * @paddr: address in the parent guest address space
1226 * @len: length of the memory area to protect
1227 *
1228 * Returns 0 if successfully protected and the rmap was created, -ENOMEM
1229 * if out of memory and -EFAULT if paddr is invalid.
1230 */
1231static int gmap_protect_rmap(struct gmap *sg, unsigned long raddr,
1232			     unsigned long paddr, unsigned long len)
1233{
1234	struct gmap *parent;
1235	struct gmap_rmap *rmap;
1236	unsigned long vmaddr;
1237	spinlock_t *ptl;
1238	pte_t *ptep;
1239	int rc;
1240
1241	BUG_ON(!gmap_is_shadow(sg));
1242	parent = sg->parent;
1243	while (len) {
1244		vmaddr = __gmap_translate(parent, paddr);
1245		if (IS_ERR_VALUE(vmaddr))
1246			return vmaddr;
1247		rmap = kzalloc(sizeof(*rmap), GFP_KERNEL_ACCOUNT);
1248		if (!rmap)
1249			return -ENOMEM;
1250		rmap->raddr = raddr;
1251		rc = radix_tree_preload(GFP_KERNEL_ACCOUNT);
1252		if (rc) {
1253			kfree(rmap);
1254			return rc;
1255		}
1256		rc = -EAGAIN;
1257		ptep = gmap_pte_op_walk(parent, paddr, &ptl);
1258		if (ptep) {
1259			spin_lock(&sg->guest_table_lock);
1260			rc = ptep_force_prot(parent->mm, paddr, ptep, PROT_READ,
1261					     PGSTE_VSIE_BIT);
1262			if (!rc)
1263				gmap_insert_rmap(sg, vmaddr, rmap);
1264			spin_unlock(&sg->guest_table_lock);
1265			gmap_pte_op_end(ptep, ptl);
1266		}
1267		radix_tree_preload_end();
1268		if (rc) {
1269			kfree(rmap);
1270			rc = gmap_pte_op_fixup(parent, paddr, vmaddr, PROT_READ);
1271			if (rc)
1272				return rc;
1273			continue;
1274		}
1275		paddr += PAGE_SIZE;
1276		len -= PAGE_SIZE;
1277	}
1278	return 0;
1279}
1280
1281#define _SHADOW_RMAP_MASK	0x7
1282#define _SHADOW_RMAP_REGION1	0x5
1283#define _SHADOW_RMAP_REGION2	0x4
1284#define _SHADOW_RMAP_REGION3	0x3
1285#define _SHADOW_RMAP_SEGMENT	0x2
1286#define _SHADOW_RMAP_PGTABLE	0x1
1287
1288/**
1289 * gmap_idte_one - invalidate a single region or segment table entry
1290 * @asce: region or segment table *origin* + table-type bits
1291 * @vaddr: virtual address to identify the table entry to flush
1292 *
1293 * The invalid bit of a single region or segment table entry is set
1294 * and the associated TLB entries depending on the entry are flushed.
1295 * The table-type of the @asce identifies the portion of the @vaddr
1296 * that is used as the invalidation index.
1297 */
1298static inline void gmap_idte_one(unsigned long asce, unsigned long vaddr)
1299{
1300	asm volatile(
1301		"	idte	%0,0,%1"
1302		: : "a" (asce), "a" (vaddr) : "cc", "memory");
1303}
1304
1305/**
1306 * gmap_unshadow_page - remove a page from a shadow page table
1307 * @sg: pointer to the shadow guest address space structure
1308 * @raddr: rmap address in the shadow guest address space
1309 *
1310 * Called with the sg->guest_table_lock
1311 */
1312static void gmap_unshadow_page(struct gmap *sg, unsigned long raddr)
1313{
1314	unsigned long *table;
1315
1316	BUG_ON(!gmap_is_shadow(sg));
1317	table = gmap_table_walk(sg, raddr, 0); /* get page table pointer */
1318	if (!table || *table & _PAGE_INVALID)
1319		return;
1320	gmap_call_notifier(sg, raddr, raddr + _PAGE_SIZE - 1);
1321	ptep_unshadow_pte(sg->mm, raddr, (pte_t *) table);
1322}
1323
1324/**
1325 * __gmap_unshadow_pgt - remove all entries from a shadow page table
1326 * @sg: pointer to the shadow guest address space structure
1327 * @raddr: rmap address in the shadow guest address space
1328 * @pgt: pointer to the start of a shadow page table
1329 *
1330 * Called with the sg->guest_table_lock
1331 */
1332static void __gmap_unshadow_pgt(struct gmap *sg, unsigned long raddr,
1333				unsigned long *pgt)
1334{
1335	int i;
1336
1337	BUG_ON(!gmap_is_shadow(sg));
1338	for (i = 0; i < _PAGE_ENTRIES; i++, raddr += _PAGE_SIZE)
1339		pgt[i] = _PAGE_INVALID;
1340}
1341
1342/**
1343 * gmap_unshadow_pgt - remove a shadow page table from a segment entry
1344 * @sg: pointer to the shadow guest address space structure
1345 * @raddr: address in the shadow guest address space
1346 *
1347 * Called with the sg->guest_table_lock
1348 */
1349static void gmap_unshadow_pgt(struct gmap *sg, unsigned long raddr)
1350{
1351	unsigned long *ste;
1352	phys_addr_t sto, pgt;
1353	struct ptdesc *ptdesc;
1354
1355	BUG_ON(!gmap_is_shadow(sg));
1356	ste = gmap_table_walk(sg, raddr, 1); /* get segment pointer */
1357	if (!ste || !(*ste & _SEGMENT_ENTRY_ORIGIN))
1358		return;
1359	gmap_call_notifier(sg, raddr, raddr + _SEGMENT_SIZE - 1);
1360	sto = __pa(ste - ((raddr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT));
1361	gmap_idte_one(sto | _ASCE_TYPE_SEGMENT, raddr);
1362	pgt = *ste & _SEGMENT_ENTRY_ORIGIN;
1363	*ste = _SEGMENT_ENTRY_EMPTY;
1364	__gmap_unshadow_pgt(sg, raddr, __va(pgt));
1365	/* Free page table */
1366	ptdesc = page_ptdesc(phys_to_page(pgt));
1367	list_del(&ptdesc->pt_list);
1368	page_table_free_pgste(ptdesc);
1369}
1370
1371/**
1372 * __gmap_unshadow_sgt - remove all entries from a shadow segment table
1373 * @sg: pointer to the shadow guest address space structure
1374 * @raddr: rmap address in the shadow guest address space
1375 * @sgt: pointer to the start of a shadow segment table
1376 *
1377 * Called with the sg->guest_table_lock
1378 */
1379static void __gmap_unshadow_sgt(struct gmap *sg, unsigned long raddr,
1380				unsigned long *sgt)
1381{
1382	struct ptdesc *ptdesc;
1383	phys_addr_t pgt;
1384	int i;
1385
1386	BUG_ON(!gmap_is_shadow(sg));
1387	for (i = 0; i < _CRST_ENTRIES; i++, raddr += _SEGMENT_SIZE) {
1388		if (!(sgt[i] & _SEGMENT_ENTRY_ORIGIN))
1389			continue;
1390		pgt = sgt[i] & _REGION_ENTRY_ORIGIN;
1391		sgt[i] = _SEGMENT_ENTRY_EMPTY;
1392		__gmap_unshadow_pgt(sg, raddr, __va(pgt));
1393		/* Free page table */
1394		ptdesc = page_ptdesc(phys_to_page(pgt));
1395		list_del(&ptdesc->pt_list);
1396		page_table_free_pgste(ptdesc);
1397	}
1398}
1399
1400/**
1401 * gmap_unshadow_sgt - remove a shadow segment table from a region-3 entry
1402 * @sg: pointer to the shadow guest address space structure
1403 * @raddr: rmap address in the shadow guest address space
1404 *
1405 * Called with the shadow->guest_table_lock
1406 */
1407static void gmap_unshadow_sgt(struct gmap *sg, unsigned long raddr)
1408{
1409	unsigned long r3o, *r3e;
1410	phys_addr_t sgt;
1411	struct page *page;
1412
1413	BUG_ON(!gmap_is_shadow(sg));
1414	r3e = gmap_table_walk(sg, raddr, 2); /* get region-3 pointer */
1415	if (!r3e || !(*r3e & _REGION_ENTRY_ORIGIN))
1416		return;
1417	gmap_call_notifier(sg, raddr, raddr + _REGION3_SIZE - 1);
1418	r3o = (unsigned long) (r3e - ((raddr & _REGION3_INDEX) >> _REGION3_SHIFT));
1419	gmap_idte_one(__pa(r3o) | _ASCE_TYPE_REGION3, raddr);
1420	sgt = *r3e & _REGION_ENTRY_ORIGIN;
1421	*r3e = _REGION3_ENTRY_EMPTY;
1422	__gmap_unshadow_sgt(sg, raddr, __va(sgt));
1423	/* Free segment table */
1424	page = phys_to_page(sgt);
1425	list_del(&page->lru);
1426	__free_pages(page, CRST_ALLOC_ORDER);
1427}
1428
1429/**
1430 * __gmap_unshadow_r3t - remove all entries from a shadow region-3 table
1431 * @sg: pointer to the shadow guest address space structure
1432 * @raddr: address in the shadow guest address space
1433 * @r3t: pointer to the start of a shadow region-3 table
1434 *
1435 * Called with the sg->guest_table_lock
1436 */
1437static void __gmap_unshadow_r3t(struct gmap *sg, unsigned long raddr,
1438				unsigned long *r3t)
1439{
1440	struct page *page;
1441	phys_addr_t sgt;
1442	int i;
1443
1444	BUG_ON(!gmap_is_shadow(sg));
1445	for (i = 0; i < _CRST_ENTRIES; i++, raddr += _REGION3_SIZE) {
1446		if (!(r3t[i] & _REGION_ENTRY_ORIGIN))
1447			continue;
1448		sgt = r3t[i] & _REGION_ENTRY_ORIGIN;
1449		r3t[i] = _REGION3_ENTRY_EMPTY;
1450		__gmap_unshadow_sgt(sg, raddr, __va(sgt));
1451		/* Free segment table */
1452		page = phys_to_page(sgt);
1453		list_del(&page->lru);
1454		__free_pages(page, CRST_ALLOC_ORDER);
1455	}
1456}
1457
1458/**
1459 * gmap_unshadow_r3t - remove a shadow region-3 table from a region-2 entry
1460 * @sg: pointer to the shadow guest address space structure
1461 * @raddr: rmap address in the shadow guest address space
1462 *
1463 * Called with the sg->guest_table_lock
1464 */
1465static void gmap_unshadow_r3t(struct gmap *sg, unsigned long raddr)
1466{
1467	unsigned long r2o, *r2e;
1468	phys_addr_t r3t;
1469	struct page *page;
1470
1471	BUG_ON(!gmap_is_shadow(sg));
1472	r2e = gmap_table_walk(sg, raddr, 3); /* get region-2 pointer */
1473	if (!r2e || !(*r2e & _REGION_ENTRY_ORIGIN))
1474		return;
1475	gmap_call_notifier(sg, raddr, raddr + _REGION2_SIZE - 1);
1476	r2o = (unsigned long) (r2e - ((raddr & _REGION2_INDEX) >> _REGION2_SHIFT));
1477	gmap_idte_one(__pa(r2o) | _ASCE_TYPE_REGION2, raddr);
1478	r3t = *r2e & _REGION_ENTRY_ORIGIN;
1479	*r2e = _REGION2_ENTRY_EMPTY;
1480	__gmap_unshadow_r3t(sg, raddr, __va(r3t));
1481	/* Free region 3 table */
1482	page = phys_to_page(r3t);
1483	list_del(&page->lru);
1484	__free_pages(page, CRST_ALLOC_ORDER);
1485}
1486
1487/**
1488 * __gmap_unshadow_r2t - remove all entries from a shadow region-2 table
1489 * @sg: pointer to the shadow guest address space structure
1490 * @raddr: rmap address in the shadow guest address space
1491 * @r2t: pointer to the start of a shadow region-2 table
1492 *
1493 * Called with the sg->guest_table_lock
1494 */
1495static void __gmap_unshadow_r2t(struct gmap *sg, unsigned long raddr,
1496				unsigned long *r2t)
1497{
1498	phys_addr_t r3t;
1499	struct page *page;
1500	int i;
1501
1502	BUG_ON(!gmap_is_shadow(sg));
1503	for (i = 0; i < _CRST_ENTRIES; i++, raddr += _REGION2_SIZE) {
1504		if (!(r2t[i] & _REGION_ENTRY_ORIGIN))
1505			continue;
1506		r3t = r2t[i] & _REGION_ENTRY_ORIGIN;
1507		r2t[i] = _REGION2_ENTRY_EMPTY;
1508		__gmap_unshadow_r3t(sg, raddr, __va(r3t));
1509		/* Free region 3 table */
1510		page = phys_to_page(r3t);
1511		list_del(&page->lru);
1512		__free_pages(page, CRST_ALLOC_ORDER);
1513	}
1514}
1515
1516/**
1517 * gmap_unshadow_r2t - remove a shadow region-2 table from a region-1 entry
1518 * @sg: pointer to the shadow guest address space structure
1519 * @raddr: rmap address in the shadow guest address space
1520 *
1521 * Called with the sg->guest_table_lock
1522 */
1523static void gmap_unshadow_r2t(struct gmap *sg, unsigned long raddr)
1524{
1525	unsigned long r1o, *r1e;
1526	struct page *page;
1527	phys_addr_t r2t;
1528
1529	BUG_ON(!gmap_is_shadow(sg));
1530	r1e = gmap_table_walk(sg, raddr, 4); /* get region-1 pointer */
1531	if (!r1e || !(*r1e & _REGION_ENTRY_ORIGIN))
1532		return;
1533	gmap_call_notifier(sg, raddr, raddr + _REGION1_SIZE - 1);
1534	r1o = (unsigned long) (r1e - ((raddr & _REGION1_INDEX) >> _REGION1_SHIFT));
1535	gmap_idte_one(__pa(r1o) | _ASCE_TYPE_REGION1, raddr);
1536	r2t = *r1e & _REGION_ENTRY_ORIGIN;
1537	*r1e = _REGION1_ENTRY_EMPTY;
1538	__gmap_unshadow_r2t(sg, raddr, __va(r2t));
1539	/* Free region 2 table */
1540	page = phys_to_page(r2t);
1541	list_del(&page->lru);
1542	__free_pages(page, CRST_ALLOC_ORDER);
1543}
1544
1545/**
1546 * __gmap_unshadow_r1t - remove all entries from a shadow region-1 table
1547 * @sg: pointer to the shadow guest address space structure
1548 * @raddr: rmap address in the shadow guest address space
1549 * @r1t: pointer to the start of a shadow region-1 table
1550 *
1551 * Called with the shadow->guest_table_lock
1552 */
1553static void __gmap_unshadow_r1t(struct gmap *sg, unsigned long raddr,
1554				unsigned long *r1t)
1555{
1556	unsigned long asce;
1557	struct page *page;
1558	phys_addr_t r2t;
1559	int i;
1560
1561	BUG_ON(!gmap_is_shadow(sg));
1562	asce = __pa(r1t) | _ASCE_TYPE_REGION1;
1563	for (i = 0; i < _CRST_ENTRIES; i++, raddr += _REGION1_SIZE) {
1564		if (!(r1t[i] & _REGION_ENTRY_ORIGIN))
1565			continue;
1566		r2t = r1t[i] & _REGION_ENTRY_ORIGIN;
1567		__gmap_unshadow_r2t(sg, raddr, __va(r2t));
1568		/* Clear entry and flush translation r1t -> r2t */
1569		gmap_idte_one(asce, raddr);
1570		r1t[i] = _REGION1_ENTRY_EMPTY;
1571		/* Free region 2 table */
1572		page = phys_to_page(r2t);
1573		list_del(&page->lru);
1574		__free_pages(page, CRST_ALLOC_ORDER);
1575	}
1576}
1577
1578/**
1579 * gmap_unshadow - remove a shadow page table completely
1580 * @sg: pointer to the shadow guest address space structure
1581 *
1582 * Called with sg->guest_table_lock
1583 */
1584static void gmap_unshadow(struct gmap *sg)
1585{
1586	unsigned long *table;
1587
1588	BUG_ON(!gmap_is_shadow(sg));
1589	if (sg->removed)
1590		return;
1591	sg->removed = 1;
1592	gmap_call_notifier(sg, 0, -1UL);
1593	gmap_flush_tlb(sg);
1594	table = __va(sg->asce & _ASCE_ORIGIN);
1595	switch (sg->asce & _ASCE_TYPE_MASK) {
1596	case _ASCE_TYPE_REGION1:
1597		__gmap_unshadow_r1t(sg, 0, table);
1598		break;
1599	case _ASCE_TYPE_REGION2:
1600		__gmap_unshadow_r2t(sg, 0, table);
1601		break;
1602	case _ASCE_TYPE_REGION3:
1603		__gmap_unshadow_r3t(sg, 0, table);
1604		break;
1605	case _ASCE_TYPE_SEGMENT:
1606		__gmap_unshadow_sgt(sg, 0, table);
1607		break;
1608	}
1609}
1610
1611/**
1612 * gmap_find_shadow - find a specific asce in the list of shadow tables
1613 * @parent: pointer to the parent gmap
1614 * @asce: ASCE for which the shadow table is created
1615 * @edat_level: edat level to be used for the shadow translation
1616 *
1617 * Returns the pointer to a gmap if a shadow table with the given asce is
1618 * already available, ERR_PTR(-EAGAIN) if another one is just being created,
1619 * otherwise NULL
1620 */
1621static struct gmap *gmap_find_shadow(struct gmap *parent, unsigned long asce,
1622				     int edat_level)
1623{
1624	struct gmap *sg;
1625
1626	list_for_each_entry(sg, &parent->children, list) {
1627		if (sg->orig_asce != asce || sg->edat_level != edat_level ||
1628		    sg->removed)
1629			continue;
1630		if (!sg->initialized)
1631			return ERR_PTR(-EAGAIN);
1632		refcount_inc(&sg->ref_count);
1633		return sg;
1634	}
1635	return NULL;
1636}
1637
1638/**
1639 * gmap_shadow_valid - check if a shadow guest address space matches the
1640 *                     given properties and is still valid
1641 * @sg: pointer to the shadow guest address space structure
1642 * @asce: ASCE for which the shadow table is requested
1643 * @edat_level: edat level to be used for the shadow translation
1644 *
1645 * Returns 1 if the gmap shadow is still valid and matches the given
1646 * properties, the caller can continue using it. Returns 0 otherwise, the
1647 * caller has to request a new shadow gmap in this case.
1648 *
1649 */
1650int gmap_shadow_valid(struct gmap *sg, unsigned long asce, int edat_level)
1651{
1652	if (sg->removed)
1653		return 0;
1654	return sg->orig_asce == asce && sg->edat_level == edat_level;
1655}
1656EXPORT_SYMBOL_GPL(gmap_shadow_valid);
1657
1658/**
1659 * gmap_shadow - create/find a shadow guest address space
1660 * @parent: pointer to the parent gmap
1661 * @asce: ASCE for which the shadow table is created
1662 * @edat_level: edat level to be used for the shadow translation
1663 *
1664 * The pages of the top level page table referred by the asce parameter
1665 * will be set to read-only and marked in the PGSTEs of the kvm process.
1666 * The shadow table will be removed automatically on any change to the
1667 * PTE mapping for the source table.
1668 *
1669 * Returns a guest address space structure, ERR_PTR(-ENOMEM) if out of memory,
1670 * ERR_PTR(-EAGAIN) if the caller has to retry and ERR_PTR(-EFAULT) if the
1671 * parent gmap table could not be protected.
1672 */
1673struct gmap *gmap_shadow(struct gmap *parent, unsigned long asce,
1674			 int edat_level)
1675{
1676	struct gmap *sg, *new;
1677	unsigned long limit;
1678	int rc;
1679
1680	BUG_ON(parent->mm->context.allow_gmap_hpage_1m);
1681	BUG_ON(gmap_is_shadow(parent));
1682	spin_lock(&parent->shadow_lock);
1683	sg = gmap_find_shadow(parent, asce, edat_level);
1684	spin_unlock(&parent->shadow_lock);
1685	if (sg)
1686		return sg;
1687	/* Create a new shadow gmap */
1688	limit = -1UL >> (33 - (((asce & _ASCE_TYPE_MASK) >> 2) * 11));
1689	if (asce & _ASCE_REAL_SPACE)
1690		limit = -1UL;
1691	new = gmap_alloc(limit);
1692	if (!new)
1693		return ERR_PTR(-ENOMEM);
1694	new->mm = parent->mm;
1695	new->parent = gmap_get(parent);
1696	new->private = parent->private;
1697	new->orig_asce = asce;
1698	new->edat_level = edat_level;
1699	new->initialized = false;
1700	spin_lock(&parent->shadow_lock);
1701	/* Recheck if another CPU created the same shadow */
1702	sg = gmap_find_shadow(parent, asce, edat_level);
1703	if (sg) {
1704		spin_unlock(&parent->shadow_lock);
1705		gmap_free(new);
1706		return sg;
1707	}
1708	if (asce & _ASCE_REAL_SPACE) {
1709		/* only allow one real-space gmap shadow */
1710		list_for_each_entry(sg, &parent->children, list) {
1711			if (sg->orig_asce & _ASCE_REAL_SPACE) {
1712				spin_lock(&sg->guest_table_lock);
1713				gmap_unshadow(sg);
1714				spin_unlock(&sg->guest_table_lock);
1715				list_del(&sg->list);
1716				gmap_put(sg);
1717				break;
1718			}
1719		}
1720	}
1721	refcount_set(&new->ref_count, 2);
1722	list_add(&new->list, &parent->children);
1723	if (asce & _ASCE_REAL_SPACE) {
1724		/* nothing to protect, return right away */
1725		new->initialized = true;
1726		spin_unlock(&parent->shadow_lock);
1727		return new;
1728	}
1729	spin_unlock(&parent->shadow_lock);
1730	/* protect after insertion, so it will get properly invalidated */
1731	mmap_read_lock(parent->mm);
1732	rc = gmap_protect_range(parent, asce & _ASCE_ORIGIN,
1733				((asce & _ASCE_TABLE_LENGTH) + 1) * PAGE_SIZE,
1734				PROT_READ, GMAP_NOTIFY_SHADOW);
1735	mmap_read_unlock(parent->mm);
1736	spin_lock(&parent->shadow_lock);
1737	new->initialized = true;
1738	if (rc) {
1739		list_del(&new->list);
1740		gmap_free(new);
1741		new = ERR_PTR(rc);
1742	}
1743	spin_unlock(&parent->shadow_lock);
1744	return new;
1745}
1746EXPORT_SYMBOL_GPL(gmap_shadow);
1747
1748/**
1749 * gmap_shadow_r2t - create an empty shadow region 2 table
1750 * @sg: pointer to the shadow guest address space structure
1751 * @saddr: faulting address in the shadow gmap
1752 * @r2t: parent gmap address of the region 2 table to get shadowed
1753 * @fake: r2t references contiguous guest memory block, not a r2t
1754 *
1755 * The r2t parameter specifies the address of the source table. The
1756 * four pages of the source table are made read-only in the parent gmap
1757 * address space. A write to the source table area @r2t will automatically
1758 * remove the shadow r2 table and all of its descendants.
1759 *
1760 * Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
1761 * shadow table structure is incomplete, -ENOMEM if out of memory and
1762 * -EFAULT if an address in the parent gmap could not be resolved.
1763 *
1764 * Called with sg->mm->mmap_lock in read.
1765 */
1766int gmap_shadow_r2t(struct gmap *sg, unsigned long saddr, unsigned long r2t,
1767		    int fake)
1768{
1769	unsigned long raddr, origin, offset, len;
1770	unsigned long *table;
1771	phys_addr_t s_r2t;
1772	struct page *page;
1773	int rc;
1774
1775	BUG_ON(!gmap_is_shadow(sg));
1776	/* Allocate a shadow region second table */
1777	page = gmap_alloc_crst();
1778	if (!page)
1779		return -ENOMEM;
1780	page->index = r2t & _REGION_ENTRY_ORIGIN;
1781	if (fake)
1782		page->index |= GMAP_SHADOW_FAKE_TABLE;
1783	s_r2t = page_to_phys(page);
1784	/* Install shadow region second table */
1785	spin_lock(&sg->guest_table_lock);
1786	table = gmap_table_walk(sg, saddr, 4); /* get region-1 pointer */
1787	if (!table) {
1788		rc = -EAGAIN;		/* Race with unshadow */
1789		goto out_free;
1790	}
1791	if (!(*table & _REGION_ENTRY_INVALID)) {
1792		rc = 0;			/* Already established */
1793		goto out_free;
1794	} else if (*table & _REGION_ENTRY_ORIGIN) {
1795		rc = -EAGAIN;		/* Race with shadow */
1796		goto out_free;
1797	}
1798	crst_table_init(__va(s_r2t), _REGION2_ENTRY_EMPTY);
1799	/* mark as invalid as long as the parent table is not protected */
1800	*table = s_r2t | _REGION_ENTRY_LENGTH |
1801		 _REGION_ENTRY_TYPE_R1 | _REGION_ENTRY_INVALID;
1802	if (sg->edat_level >= 1)
1803		*table |= (r2t & _REGION_ENTRY_PROTECT);
1804	list_add(&page->lru, &sg->crst_list);
1805	if (fake) {
1806		/* nothing to protect for fake tables */
1807		*table &= ~_REGION_ENTRY_INVALID;
1808		spin_unlock(&sg->guest_table_lock);
1809		return 0;
1810	}
1811	spin_unlock(&sg->guest_table_lock);
1812	/* Make r2t read-only in parent gmap page table */
1813	raddr = (saddr & _REGION1_MASK) | _SHADOW_RMAP_REGION1;
1814	origin = r2t & _REGION_ENTRY_ORIGIN;
1815	offset = ((r2t & _REGION_ENTRY_OFFSET) >> 6) * PAGE_SIZE;
1816	len = ((r2t & _REGION_ENTRY_LENGTH) + 1) * PAGE_SIZE - offset;
1817	rc = gmap_protect_rmap(sg, raddr, origin + offset, len);
1818	spin_lock(&sg->guest_table_lock);
1819	if (!rc) {
1820		table = gmap_table_walk(sg, saddr, 4);
1821		if (!table || (*table & _REGION_ENTRY_ORIGIN) != s_r2t)
1822			rc = -EAGAIN;		/* Race with unshadow */
1823		else
1824			*table &= ~_REGION_ENTRY_INVALID;
1825	} else {
1826		gmap_unshadow_r2t(sg, raddr);
1827	}
1828	spin_unlock(&sg->guest_table_lock);
1829	return rc;
1830out_free:
1831	spin_unlock(&sg->guest_table_lock);
1832	__free_pages(page, CRST_ALLOC_ORDER);
1833	return rc;
1834}
1835EXPORT_SYMBOL_GPL(gmap_shadow_r2t);
1836
1837/**
1838 * gmap_shadow_r3t - create a shadow region 3 table
1839 * @sg: pointer to the shadow guest address space structure
1840 * @saddr: faulting address in the shadow gmap
1841 * @r3t: parent gmap address of the region 3 table to get shadowed
1842 * @fake: r3t references contiguous guest memory block, not a r3t
1843 *
1844 * Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
1845 * shadow table structure is incomplete, -ENOMEM if out of memory and
1846 * -EFAULT if an address in the parent gmap could not be resolved.
1847 *
1848 * Called with sg->mm->mmap_lock in read.
1849 */
1850int gmap_shadow_r3t(struct gmap *sg, unsigned long saddr, unsigned long r3t,
1851		    int fake)
1852{
1853	unsigned long raddr, origin, offset, len;
1854	unsigned long *table;
1855	phys_addr_t s_r3t;
1856	struct page *page;
1857	int rc;
1858
1859	BUG_ON(!gmap_is_shadow(sg));
1860	/* Allocate a shadow region second table */
1861	page = gmap_alloc_crst();
1862	if (!page)
1863		return -ENOMEM;
1864	page->index = r3t & _REGION_ENTRY_ORIGIN;
1865	if (fake)
1866		page->index |= GMAP_SHADOW_FAKE_TABLE;
1867	s_r3t = page_to_phys(page);
1868	/* Install shadow region second table */
1869	spin_lock(&sg->guest_table_lock);
1870	table = gmap_table_walk(sg, saddr, 3); /* get region-2 pointer */
1871	if (!table) {
1872		rc = -EAGAIN;		/* Race with unshadow */
1873		goto out_free;
1874	}
1875	if (!(*table & _REGION_ENTRY_INVALID)) {
1876		rc = 0;			/* Already established */
1877		goto out_free;
1878	} else if (*table & _REGION_ENTRY_ORIGIN) {
1879		rc = -EAGAIN;		/* Race with shadow */
1880		goto out_free;
1881	}
1882	crst_table_init(__va(s_r3t), _REGION3_ENTRY_EMPTY);
1883	/* mark as invalid as long as the parent table is not protected */
1884	*table = s_r3t | _REGION_ENTRY_LENGTH |
1885		 _REGION_ENTRY_TYPE_R2 | _REGION_ENTRY_INVALID;
1886	if (sg->edat_level >= 1)
1887		*table |= (r3t & _REGION_ENTRY_PROTECT);
1888	list_add(&page->lru, &sg->crst_list);
1889	if (fake) {
1890		/* nothing to protect for fake tables */
1891		*table &= ~_REGION_ENTRY_INVALID;
1892		spin_unlock(&sg->guest_table_lock);
1893		return 0;
1894	}
1895	spin_unlock(&sg->guest_table_lock);
1896	/* Make r3t read-only in parent gmap page table */
1897	raddr = (saddr & _REGION2_MASK) | _SHADOW_RMAP_REGION2;
1898	origin = r3t & _REGION_ENTRY_ORIGIN;
1899	offset = ((r3t & _REGION_ENTRY_OFFSET) >> 6) * PAGE_SIZE;
1900	len = ((r3t & _REGION_ENTRY_LENGTH) + 1) * PAGE_SIZE - offset;
1901	rc = gmap_protect_rmap(sg, raddr, origin + offset, len);
1902	spin_lock(&sg->guest_table_lock);
1903	if (!rc) {
1904		table = gmap_table_walk(sg, saddr, 3);
1905		if (!table || (*table & _REGION_ENTRY_ORIGIN) != s_r3t)
1906			rc = -EAGAIN;		/* Race with unshadow */
1907		else
1908			*table &= ~_REGION_ENTRY_INVALID;
1909	} else {
1910		gmap_unshadow_r3t(sg, raddr);
1911	}
1912	spin_unlock(&sg->guest_table_lock);
1913	return rc;
1914out_free:
1915	spin_unlock(&sg->guest_table_lock);
1916	__free_pages(page, CRST_ALLOC_ORDER);
1917	return rc;
1918}
1919EXPORT_SYMBOL_GPL(gmap_shadow_r3t);
1920
1921/**
1922 * gmap_shadow_sgt - create a shadow segment table
1923 * @sg: pointer to the shadow guest address space structure
1924 * @saddr: faulting address in the shadow gmap
1925 * @sgt: parent gmap address of the segment table to get shadowed
1926 * @fake: sgt references contiguous guest memory block, not a sgt
1927 *
1928 * Returns: 0 if successfully shadowed or already shadowed, -EAGAIN if the
1929 * shadow table structure is incomplete, -ENOMEM if out of memory and
1930 * -EFAULT if an address in the parent gmap could not be resolved.
1931 *
1932 * Called with sg->mm->mmap_lock in read.
1933 */
1934int gmap_shadow_sgt(struct gmap *sg, unsigned long saddr, unsigned long sgt,
1935		    int fake)
1936{
1937	unsigned long raddr, origin, offset, len;
1938	unsigned long *table;
1939	phys_addr_t s_sgt;
1940	struct page *page;
1941	int rc;
1942
1943	BUG_ON(!gmap_is_shadow(sg) || (sgt & _REGION3_ENTRY_LARGE));
1944	/* Allocate a shadow segment table */
1945	page = gmap_alloc_crst();
1946	if (!page)
1947		return -ENOMEM;
1948	page->index = sgt & _REGION_ENTRY_ORIGIN;
1949	if (fake)
1950		page->index |= GMAP_SHADOW_FAKE_TABLE;
1951	s_sgt = page_to_phys(page);
1952	/* Install shadow region second table */
1953	spin_lock(&sg->guest_table_lock);
1954	table = gmap_table_walk(sg, saddr, 2); /* get region-3 pointer */
1955	if (!table) {
1956		rc = -EAGAIN;		/* Race with unshadow */
1957		goto out_free;
1958	}
1959	if (!(*table & _REGION_ENTRY_INVALID)) {
1960		rc = 0;			/* Already established */
1961		goto out_free;
1962	} else if (*table & _REGION_ENTRY_ORIGIN) {
1963		rc = -EAGAIN;		/* Race with shadow */
1964		goto out_free;
1965	}
1966	crst_table_init(__va(s_sgt), _SEGMENT_ENTRY_EMPTY);
1967	/* mark as invalid as long as the parent table is not protected */
1968	*table = s_sgt | _REGION_ENTRY_LENGTH |
1969		 _REGION_ENTRY_TYPE_R3 | _REGION_ENTRY_INVALID;
1970	if (sg->edat_level >= 1)
1971		*table |= sgt & _REGION_ENTRY_PROTECT;
1972	list_add(&page->lru, &sg->crst_list);
1973	if (fake) {
1974		/* nothing to protect for fake tables */
1975		*table &= ~_REGION_ENTRY_INVALID;
1976		spin_unlock(&sg->guest_table_lock);
1977		return 0;
1978	}
1979	spin_unlock(&sg->guest_table_lock);
1980	/* Make sgt read-only in parent gmap page table */
1981	raddr = (saddr & _REGION3_MASK) | _SHADOW_RMAP_REGION3;
1982	origin = sgt & _REGION_ENTRY_ORIGIN;
1983	offset = ((sgt & _REGION_ENTRY_OFFSET) >> 6) * PAGE_SIZE;
1984	len = ((sgt & _REGION_ENTRY_LENGTH) + 1) * PAGE_SIZE - offset;
1985	rc = gmap_protect_rmap(sg, raddr, origin + offset, len);
1986	spin_lock(&sg->guest_table_lock);
1987	if (!rc) {
1988		table = gmap_table_walk(sg, saddr, 2);
1989		if (!table || (*table & _REGION_ENTRY_ORIGIN) != s_sgt)
1990			rc = -EAGAIN;		/* Race with unshadow */
1991		else
1992			*table &= ~_REGION_ENTRY_INVALID;
1993	} else {
1994		gmap_unshadow_sgt(sg, raddr);
1995	}
1996	spin_unlock(&sg->guest_table_lock);
1997	return rc;
1998out_free:
1999	spin_unlock(&sg->guest_table_lock);
2000	__free_pages(page, CRST_ALLOC_ORDER);
2001	return rc;
2002}
2003EXPORT_SYMBOL_GPL(gmap_shadow_sgt);
2004
2005/**
2006 * gmap_shadow_pgt_lookup - find a shadow page table
2007 * @sg: pointer to the shadow guest address space structure
2008 * @saddr: the address in the shadow aguest address space
2009 * @pgt: parent gmap address of the page table to get shadowed
2010 * @dat_protection: if the pgtable is marked as protected by dat
2011 * @fake: pgt references contiguous guest memory block, not a pgtable
2012 *
2013 * Returns 0 if the shadow page table was found and -EAGAIN if the page
2014 * table was not found.
2015 *
2016 * Called with sg->mm->mmap_lock in read.
2017 */
2018int gmap_shadow_pgt_lookup(struct gmap *sg, unsigned long saddr,
2019			   unsigned long *pgt, int *dat_protection,
2020			   int *fake)
2021{
2022	unsigned long *table;
2023	struct page *page;
2024	int rc;
2025
2026	BUG_ON(!gmap_is_shadow(sg));
2027	spin_lock(&sg->guest_table_lock);
2028	table = gmap_table_walk(sg, saddr, 1); /* get segment pointer */
2029	if (table && !(*table & _SEGMENT_ENTRY_INVALID)) {
2030		/* Shadow page tables are full pages (pte+pgste) */
2031		page = pfn_to_page(*table >> PAGE_SHIFT);
2032		*pgt = page->index & ~GMAP_SHADOW_FAKE_TABLE;
2033		*dat_protection = !!(*table & _SEGMENT_ENTRY_PROTECT);
2034		*fake = !!(page->index & GMAP_SHADOW_FAKE_TABLE);
2035		rc = 0;
2036	} else  {
2037		rc = -EAGAIN;
2038	}
2039	spin_unlock(&sg->guest_table_lock);
2040	return rc;
2041
2042}
2043EXPORT_SYMBOL_GPL(gmap_shadow_pgt_lookup);
2044
2045/**
2046 * gmap_shadow_pgt - instantiate a shadow page table
2047 * @sg: pointer to the shadow guest address space structure
2048 * @saddr: faulting address in the shadow gmap
2049 * @pgt: parent gmap address of the page table to get shadowed
2050 * @fake: pgt references contiguous guest memory block, not a pgtable
2051 *
2052 * Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
2053 * shadow table structure is incomplete, -ENOMEM if out of memory,
2054 * -EFAULT if an address in the parent gmap could not be resolved and
2055 *
2056 * Called with gmap->mm->mmap_lock in read
2057 */
2058int gmap_shadow_pgt(struct gmap *sg, unsigned long saddr, unsigned long pgt,
2059		    int fake)
2060{
2061	unsigned long raddr, origin;
2062	unsigned long *table;
2063	struct ptdesc *ptdesc;
2064	phys_addr_t s_pgt;
2065	int rc;
2066
2067	BUG_ON(!gmap_is_shadow(sg) || (pgt & _SEGMENT_ENTRY_LARGE));
2068	/* Allocate a shadow page table */
2069	ptdesc = page_table_alloc_pgste(sg->mm);
2070	if (!ptdesc)
2071		return -ENOMEM;
2072	ptdesc->pt_index = pgt & _SEGMENT_ENTRY_ORIGIN;
2073	if (fake)
2074		ptdesc->pt_index |= GMAP_SHADOW_FAKE_TABLE;
2075	s_pgt = page_to_phys(ptdesc_page(ptdesc));
2076	/* Install shadow page table */
2077	spin_lock(&sg->guest_table_lock);
2078	table = gmap_table_walk(sg, saddr, 1); /* get segment pointer */
2079	if (!table) {
2080		rc = -EAGAIN;		/* Race with unshadow */
2081		goto out_free;
2082	}
2083	if (!(*table & _SEGMENT_ENTRY_INVALID)) {
2084		rc = 0;			/* Already established */
2085		goto out_free;
2086	} else if (*table & _SEGMENT_ENTRY_ORIGIN) {
2087		rc = -EAGAIN;		/* Race with shadow */
2088		goto out_free;
2089	}
2090	/* mark as invalid as long as the parent table is not protected */
2091	*table = (unsigned long) s_pgt | _SEGMENT_ENTRY |
2092		 (pgt & _SEGMENT_ENTRY_PROTECT) | _SEGMENT_ENTRY_INVALID;
2093	list_add(&ptdesc->pt_list, &sg->pt_list);
2094	if (fake) {
2095		/* nothing to protect for fake tables */
2096		*table &= ~_SEGMENT_ENTRY_INVALID;
2097		spin_unlock(&sg->guest_table_lock);
2098		return 0;
2099	}
2100	spin_unlock(&sg->guest_table_lock);
2101	/* Make pgt read-only in parent gmap page table (not the pgste) */
2102	raddr = (saddr & _SEGMENT_MASK) | _SHADOW_RMAP_SEGMENT;
2103	origin = pgt & _SEGMENT_ENTRY_ORIGIN & PAGE_MASK;
2104	rc = gmap_protect_rmap(sg, raddr, origin, PAGE_SIZE);
2105	spin_lock(&sg->guest_table_lock);
2106	if (!rc) {
2107		table = gmap_table_walk(sg, saddr, 1);
2108		if (!table || (*table & _SEGMENT_ENTRY_ORIGIN) != s_pgt)
2109			rc = -EAGAIN;		/* Race with unshadow */
2110		else
2111			*table &= ~_SEGMENT_ENTRY_INVALID;
2112	} else {
2113		gmap_unshadow_pgt(sg, raddr);
2114	}
2115	spin_unlock(&sg->guest_table_lock);
2116	return rc;
2117out_free:
2118	spin_unlock(&sg->guest_table_lock);
2119	page_table_free_pgste(ptdesc);
2120	return rc;
2121
2122}
2123EXPORT_SYMBOL_GPL(gmap_shadow_pgt);
2124
2125/**
2126 * gmap_shadow_page - create a shadow page mapping
2127 * @sg: pointer to the shadow guest address space structure
2128 * @saddr: faulting address in the shadow gmap
2129 * @pte: pte in parent gmap address space to get shadowed
2130 *
2131 * Returns 0 if successfully shadowed or already shadowed, -EAGAIN if the
2132 * shadow table structure is incomplete, -ENOMEM if out of memory and
2133 * -EFAULT if an address in the parent gmap could not be resolved.
2134 *
2135 * Called with sg->mm->mmap_lock in read.
2136 */
2137int gmap_shadow_page(struct gmap *sg, unsigned long saddr, pte_t pte)
2138{
2139	struct gmap *parent;
2140	struct gmap_rmap *rmap;
2141	unsigned long vmaddr, paddr;
2142	spinlock_t *ptl;
2143	pte_t *sptep, *tptep;
2144	int prot;
2145	int rc;
2146
2147	BUG_ON(!gmap_is_shadow(sg));
2148	parent = sg->parent;
2149	prot = (pte_val(pte) & _PAGE_PROTECT) ? PROT_READ : PROT_WRITE;
2150
2151	rmap = kzalloc(sizeof(*rmap), GFP_KERNEL_ACCOUNT);
2152	if (!rmap)
2153		return -ENOMEM;
2154	rmap->raddr = (saddr & PAGE_MASK) | _SHADOW_RMAP_PGTABLE;
2155
2156	while (1) {
2157		paddr = pte_val(pte) & PAGE_MASK;
2158		vmaddr = __gmap_translate(parent, paddr);
2159		if (IS_ERR_VALUE(vmaddr)) {
2160			rc = vmaddr;
2161			break;
2162		}
2163		rc = radix_tree_preload(GFP_KERNEL_ACCOUNT);
2164		if (rc)
2165			break;
2166		rc = -EAGAIN;
2167		sptep = gmap_pte_op_walk(parent, paddr, &ptl);
2168		if (sptep) {
2169			spin_lock(&sg->guest_table_lock);
2170			/* Get page table pointer */
2171			tptep = (pte_t *) gmap_table_walk(sg, saddr, 0);
2172			if (!tptep) {
2173				spin_unlock(&sg->guest_table_lock);
2174				gmap_pte_op_end(sptep, ptl);
2175				radix_tree_preload_end();
2176				break;
2177			}
2178			rc = ptep_shadow_pte(sg->mm, saddr, sptep, tptep, pte);
2179			if (rc > 0) {
2180				/* Success and a new mapping */
2181				gmap_insert_rmap(sg, vmaddr, rmap);
2182				rmap = NULL;
2183				rc = 0;
2184			}
2185			gmap_pte_op_end(sptep, ptl);
2186			spin_unlock(&sg->guest_table_lock);
2187		}
2188		radix_tree_preload_end();
2189		if (!rc)
2190			break;
2191		rc = gmap_pte_op_fixup(parent, paddr, vmaddr, prot);
2192		if (rc)
2193			break;
2194	}
2195	kfree(rmap);
2196	return rc;
2197}
2198EXPORT_SYMBOL_GPL(gmap_shadow_page);
2199
2200/*
2201 * gmap_shadow_notify - handle notifications for shadow gmap
2202 *
2203 * Called with sg->parent->shadow_lock.
2204 */
2205static void gmap_shadow_notify(struct gmap *sg, unsigned long vmaddr,
2206			       unsigned long gaddr)
2207{
2208	struct gmap_rmap *rmap, *rnext, *head;
2209	unsigned long start, end, bits, raddr;
2210
2211	BUG_ON(!gmap_is_shadow(sg));
2212
2213	spin_lock(&sg->guest_table_lock);
2214	if (sg->removed) {
2215		spin_unlock(&sg->guest_table_lock);
2216		return;
2217	}
2218	/* Check for top level table */
2219	start = sg->orig_asce & _ASCE_ORIGIN;
2220	end = start + ((sg->orig_asce & _ASCE_TABLE_LENGTH) + 1) * PAGE_SIZE;
2221	if (!(sg->orig_asce & _ASCE_REAL_SPACE) && gaddr >= start &&
2222	    gaddr < end) {
2223		/* The complete shadow table has to go */
2224		gmap_unshadow(sg);
2225		spin_unlock(&sg->guest_table_lock);
2226		list_del(&sg->list);
2227		gmap_put(sg);
2228		return;
2229	}
2230	/* Remove the page table tree from on specific entry */
2231	head = radix_tree_delete(&sg->host_to_rmap, vmaddr >> PAGE_SHIFT);
2232	gmap_for_each_rmap_safe(rmap, rnext, head) {
2233		bits = rmap->raddr & _SHADOW_RMAP_MASK;
2234		raddr = rmap->raddr ^ bits;
2235		switch (bits) {
2236		case _SHADOW_RMAP_REGION1:
2237			gmap_unshadow_r2t(sg, raddr);
2238			break;
2239		case _SHADOW_RMAP_REGION2:
2240			gmap_unshadow_r3t(sg, raddr);
2241			break;
2242		case _SHADOW_RMAP_REGION3:
2243			gmap_unshadow_sgt(sg, raddr);
2244			break;
2245		case _SHADOW_RMAP_SEGMENT:
2246			gmap_unshadow_pgt(sg, raddr);
2247			break;
2248		case _SHADOW_RMAP_PGTABLE:
2249			gmap_unshadow_page(sg, raddr);
2250			break;
2251		}
2252		kfree(rmap);
2253	}
2254	spin_unlock(&sg->guest_table_lock);
2255}
2256
2257/**
2258 * ptep_notify - call all invalidation callbacks for a specific pte.
2259 * @mm: pointer to the process mm_struct
2260 * @vmaddr: virtual address in the process address space
2261 * @pte: pointer to the page table entry
2262 * @bits: bits from the pgste that caused the notify call
2263 *
2264 * This function is assumed to be called with the page table lock held
2265 * for the pte to notify.
2266 */
2267void ptep_notify(struct mm_struct *mm, unsigned long vmaddr,
2268		 pte_t *pte, unsigned long bits)
2269{
2270	unsigned long offset, gaddr = 0;
2271	unsigned long *table;
2272	struct gmap *gmap, *sg, *next;
2273
2274	offset = ((unsigned long) pte) & (255 * sizeof(pte_t));
2275	offset = offset * (PAGE_SIZE / sizeof(pte_t));
2276	rcu_read_lock();
2277	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
2278		spin_lock(&gmap->guest_table_lock);
2279		table = radix_tree_lookup(&gmap->host_to_guest,
2280					  vmaddr >> PMD_SHIFT);
2281		if (table)
2282			gaddr = __gmap_segment_gaddr(table) + offset;
2283		spin_unlock(&gmap->guest_table_lock);
2284		if (!table)
2285			continue;
2286
2287		if (!list_empty(&gmap->children) && (bits & PGSTE_VSIE_BIT)) {
2288			spin_lock(&gmap->shadow_lock);
2289			list_for_each_entry_safe(sg, next,
2290						 &gmap->children, list)
2291				gmap_shadow_notify(sg, vmaddr, gaddr);
2292			spin_unlock(&gmap->shadow_lock);
2293		}
2294		if (bits & PGSTE_IN_BIT)
2295			gmap_call_notifier(gmap, gaddr, gaddr + PAGE_SIZE - 1);
2296	}
2297	rcu_read_unlock();
2298}
2299EXPORT_SYMBOL_GPL(ptep_notify);
2300
2301static void pmdp_notify_gmap(struct gmap *gmap, pmd_t *pmdp,
2302			     unsigned long gaddr)
2303{
2304	set_pmd(pmdp, clear_pmd_bit(*pmdp, __pgprot(_SEGMENT_ENTRY_GMAP_IN)));
2305	gmap_call_notifier(gmap, gaddr, gaddr + HPAGE_SIZE - 1);
2306}
2307
2308/**
2309 * gmap_pmdp_xchg - exchange a gmap pmd with another
2310 * @gmap: pointer to the guest address space structure
2311 * @pmdp: pointer to the pmd entry
2312 * @new: replacement entry
2313 * @gaddr: the affected guest address
2314 *
2315 * This function is assumed to be called with the guest_table_lock
2316 * held.
2317 */
2318static void gmap_pmdp_xchg(struct gmap *gmap, pmd_t *pmdp, pmd_t new,
2319			   unsigned long gaddr)
2320{
2321	gaddr &= HPAGE_MASK;
2322	pmdp_notify_gmap(gmap, pmdp, gaddr);
2323	new = clear_pmd_bit(new, __pgprot(_SEGMENT_ENTRY_GMAP_IN));
2324	if (MACHINE_HAS_TLB_GUEST)
2325		__pmdp_idte(gaddr, (pmd_t *)pmdp, IDTE_GUEST_ASCE, gmap->asce,
2326			    IDTE_GLOBAL);
2327	else if (MACHINE_HAS_IDTE)
2328		__pmdp_idte(gaddr, (pmd_t *)pmdp, 0, 0, IDTE_GLOBAL);
2329	else
2330		__pmdp_csp(pmdp);
2331	set_pmd(pmdp, new);
2332}
2333
2334static void gmap_pmdp_clear(struct mm_struct *mm, unsigned long vmaddr,
2335			    int purge)
2336{
2337	pmd_t *pmdp;
2338	struct gmap *gmap;
2339	unsigned long gaddr;
2340
2341	rcu_read_lock();
2342	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
2343		spin_lock(&gmap->guest_table_lock);
2344		pmdp = (pmd_t *)radix_tree_delete(&gmap->host_to_guest,
2345						  vmaddr >> PMD_SHIFT);
2346		if (pmdp) {
2347			gaddr = __gmap_segment_gaddr((unsigned long *)pmdp);
2348			pmdp_notify_gmap(gmap, pmdp, gaddr);
2349			WARN_ON(pmd_val(*pmdp) & ~(_SEGMENT_ENTRY_HARDWARE_BITS_LARGE |
2350						   _SEGMENT_ENTRY_GMAP_UC));
2351			if (purge)
2352				__pmdp_csp(pmdp);
2353			set_pmd(pmdp, __pmd(_SEGMENT_ENTRY_EMPTY));
2354		}
2355		spin_unlock(&gmap->guest_table_lock);
2356	}
2357	rcu_read_unlock();
2358}
2359
2360/**
2361 * gmap_pmdp_invalidate - invalidate all affected guest pmd entries without
2362 *                        flushing
2363 * @mm: pointer to the process mm_struct
2364 * @vmaddr: virtual address in the process address space
2365 */
2366void gmap_pmdp_invalidate(struct mm_struct *mm, unsigned long vmaddr)
2367{
2368	gmap_pmdp_clear(mm, vmaddr, 0);
2369}
2370EXPORT_SYMBOL_GPL(gmap_pmdp_invalidate);
2371
2372/**
2373 * gmap_pmdp_csp - csp all affected guest pmd entries
2374 * @mm: pointer to the process mm_struct
2375 * @vmaddr: virtual address in the process address space
2376 */
2377void gmap_pmdp_csp(struct mm_struct *mm, unsigned long vmaddr)
2378{
2379	gmap_pmdp_clear(mm, vmaddr, 1);
2380}
2381EXPORT_SYMBOL_GPL(gmap_pmdp_csp);
2382
2383/**
2384 * gmap_pmdp_idte_local - invalidate and clear a guest pmd entry
2385 * @mm: pointer to the process mm_struct
2386 * @vmaddr: virtual address in the process address space
2387 */
2388void gmap_pmdp_idte_local(struct mm_struct *mm, unsigned long vmaddr)
2389{
2390	unsigned long *entry, gaddr;
2391	struct gmap *gmap;
2392	pmd_t *pmdp;
2393
2394	rcu_read_lock();
2395	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
2396		spin_lock(&gmap->guest_table_lock);
2397		entry = radix_tree_delete(&gmap->host_to_guest,
2398					  vmaddr >> PMD_SHIFT);
2399		if (entry) {
2400			pmdp = (pmd_t *)entry;
2401			gaddr = __gmap_segment_gaddr(entry);
2402			pmdp_notify_gmap(gmap, pmdp, gaddr);
2403			WARN_ON(*entry & ~(_SEGMENT_ENTRY_HARDWARE_BITS_LARGE |
2404					   _SEGMENT_ENTRY_GMAP_UC));
2405			if (MACHINE_HAS_TLB_GUEST)
2406				__pmdp_idte(gaddr, pmdp, IDTE_GUEST_ASCE,
2407					    gmap->asce, IDTE_LOCAL);
2408			else if (MACHINE_HAS_IDTE)
2409				__pmdp_idte(gaddr, pmdp, 0, 0, IDTE_LOCAL);
2410			*entry = _SEGMENT_ENTRY_EMPTY;
2411		}
2412		spin_unlock(&gmap->guest_table_lock);
2413	}
2414	rcu_read_unlock();
2415}
2416EXPORT_SYMBOL_GPL(gmap_pmdp_idte_local);
2417
2418/**
2419 * gmap_pmdp_idte_global - invalidate and clear a guest pmd entry
2420 * @mm: pointer to the process mm_struct
2421 * @vmaddr: virtual address in the process address space
2422 */
2423void gmap_pmdp_idte_global(struct mm_struct *mm, unsigned long vmaddr)
2424{
2425	unsigned long *entry, gaddr;
2426	struct gmap *gmap;
2427	pmd_t *pmdp;
2428
2429	rcu_read_lock();
2430	list_for_each_entry_rcu(gmap, &mm->context.gmap_list, list) {
2431		spin_lock(&gmap->guest_table_lock);
2432		entry = radix_tree_delete(&gmap->host_to_guest,
2433					  vmaddr >> PMD_SHIFT);
2434		if (entry) {
2435			pmdp = (pmd_t *)entry;
2436			gaddr = __gmap_segment_gaddr(entry);
2437			pmdp_notify_gmap(gmap, pmdp, gaddr);
2438			WARN_ON(*entry & ~(_SEGMENT_ENTRY_HARDWARE_BITS_LARGE |
2439					   _SEGMENT_ENTRY_GMAP_UC));
2440			if (MACHINE_HAS_TLB_GUEST)
2441				__pmdp_idte(gaddr, pmdp, IDTE_GUEST_ASCE,
2442					    gmap->asce, IDTE_GLOBAL);
2443			else if (MACHINE_HAS_IDTE)
2444				__pmdp_idte(gaddr, pmdp, 0, 0, IDTE_GLOBAL);
2445			else
2446				__pmdp_csp(pmdp);
2447			*entry = _SEGMENT_ENTRY_EMPTY;
2448		}
2449		spin_unlock(&gmap->guest_table_lock);
2450	}
2451	rcu_read_unlock();
2452}
2453EXPORT_SYMBOL_GPL(gmap_pmdp_idte_global);
2454
2455/**
2456 * gmap_test_and_clear_dirty_pmd - test and reset segment dirty status
2457 * @gmap: pointer to guest address space
2458 * @pmdp: pointer to the pmd to be tested
2459 * @gaddr: virtual address in the guest address space
2460 *
2461 * This function is assumed to be called with the guest_table_lock
2462 * held.
2463 */
2464static bool gmap_test_and_clear_dirty_pmd(struct gmap *gmap, pmd_t *pmdp,
2465					  unsigned long gaddr)
2466{
2467	if (pmd_val(*pmdp) & _SEGMENT_ENTRY_INVALID)
2468		return false;
2469
2470	/* Already protected memory, which did not change is clean */
2471	if (pmd_val(*pmdp) & _SEGMENT_ENTRY_PROTECT &&
2472	    !(pmd_val(*pmdp) & _SEGMENT_ENTRY_GMAP_UC))
2473		return false;
2474
2475	/* Clear UC indication and reset protection */
2476	set_pmd(pmdp, clear_pmd_bit(*pmdp, __pgprot(_SEGMENT_ENTRY_GMAP_UC)));
2477	gmap_protect_pmd(gmap, gaddr, pmdp, PROT_READ, 0);
2478	return true;
2479}
2480
2481/**
2482 * gmap_sync_dirty_log_pmd - set bitmap based on dirty status of segment
2483 * @gmap: pointer to guest address space
2484 * @bitmap: dirty bitmap for this pmd
2485 * @gaddr: virtual address in the guest address space
2486 * @vmaddr: virtual address in the host address space
2487 *
2488 * This function is assumed to be called with the guest_table_lock
2489 * held.
2490 */
2491void gmap_sync_dirty_log_pmd(struct gmap *gmap, unsigned long bitmap[4],
2492			     unsigned long gaddr, unsigned long vmaddr)
2493{
2494	int i;
2495	pmd_t *pmdp;
2496	pte_t *ptep;
2497	spinlock_t *ptl;
2498
2499	pmdp = gmap_pmd_op_walk(gmap, gaddr);
2500	if (!pmdp)
2501		return;
2502
2503	if (pmd_leaf(*pmdp)) {
2504		if (gmap_test_and_clear_dirty_pmd(gmap, pmdp, gaddr))
2505			bitmap_fill(bitmap, _PAGE_ENTRIES);
2506	} else {
2507		for (i = 0; i < _PAGE_ENTRIES; i++, vmaddr += PAGE_SIZE) {
2508			ptep = pte_alloc_map_lock(gmap->mm, pmdp, vmaddr, &ptl);
2509			if (!ptep)
2510				continue;
2511			if (ptep_test_and_clear_uc(gmap->mm, vmaddr, ptep))
2512				set_bit(i, bitmap);
2513			pte_unmap_unlock(ptep, ptl);
2514		}
2515	}
2516	gmap_pmd_op_end(gmap, pmdp);
2517}
2518EXPORT_SYMBOL_GPL(gmap_sync_dirty_log_pmd);
2519
2520#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2521static int thp_split_walk_pmd_entry(pmd_t *pmd, unsigned long addr,
2522				    unsigned long end, struct mm_walk *walk)
2523{
2524	struct vm_area_struct *vma = walk->vma;
2525
2526	split_huge_pmd(vma, pmd, addr);
2527	return 0;
2528}
2529
2530static const struct mm_walk_ops thp_split_walk_ops = {
2531	.pmd_entry	= thp_split_walk_pmd_entry,
2532	.walk_lock	= PGWALK_WRLOCK_VERIFY,
2533};
2534
2535static inline void thp_split_mm(struct mm_struct *mm)
2536{
2537	struct vm_area_struct *vma;
2538	VMA_ITERATOR(vmi, mm, 0);
2539
2540	for_each_vma(vmi, vma) {
2541		vm_flags_mod(vma, VM_NOHUGEPAGE, VM_HUGEPAGE);
 
2542		walk_page_vma(vma, &thp_split_walk_ops, NULL);
2543	}
2544	mm->def_flags |= VM_NOHUGEPAGE;
2545}
2546#else
2547static inline void thp_split_mm(struct mm_struct *mm)
2548{
2549}
2550#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2551
2552/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2553 * switch on pgstes for its userspace process (for kvm)
2554 */
2555int s390_enable_sie(void)
2556{
2557	struct mm_struct *mm = current->mm;
2558
2559	/* Do we have pgstes? if yes, we are done */
2560	if (mm_has_pgste(mm))
2561		return 0;
2562	/* Fail if the page tables are 2K */
2563	if (!mm_alloc_pgste(mm))
2564		return -EINVAL;
2565	mmap_write_lock(mm);
2566	mm->context.has_pgste = 1;
2567	/* split thp mappings and disable thp for future mappings */
2568	thp_split_mm(mm);
 
2569	mmap_write_unlock(mm);
2570	return 0;
2571}
2572EXPORT_SYMBOL_GPL(s390_enable_sie);
2573
2574static int find_zeropage_pte_entry(pte_t *pte, unsigned long addr,
2575				   unsigned long end, struct mm_walk *walk)
2576{
2577	unsigned long *found_addr = walk->private;
2578
2579	/* Return 1 of the page is a zeropage. */
2580	if (is_zero_pfn(pte_pfn(*pte))) {
2581		/*
2582		 * Shared zeropage in e.g., a FS DAX mapping? We cannot do the
2583		 * right thing and likely don't care: FAULT_FLAG_UNSHARE
2584		 * currently only works in COW mappings, which is also where
2585		 * mm_forbids_zeropage() is checked.
2586		 */
2587		if (!is_cow_mapping(walk->vma->vm_flags))
2588			return -EFAULT;
2589
2590		*found_addr = addr;
2591		return 1;
2592	}
2593	return 0;
2594}
2595
2596static const struct mm_walk_ops find_zeropage_ops = {
2597	.pte_entry	= find_zeropage_pte_entry,
2598	.walk_lock	= PGWALK_WRLOCK,
2599};
2600
2601/*
2602 * Unshare all shared zeropages, replacing them by anonymous pages. Note that
2603 * we cannot simply zap all shared zeropages, because this could later
2604 * trigger unexpected userfaultfd missing events.
2605 *
2606 * This must be called after mm->context.allow_cow_sharing was
2607 * set to 0, to avoid future mappings of shared zeropages.
2608 *
2609 * mm contracts with s390, that even if mm were to remove a page table,
2610 * and racing with walk_page_range_vma() calling pte_offset_map_lock()
2611 * would fail, it will never insert a page table containing empty zero
2612 * pages once mm_forbids_zeropage(mm) i.e.
2613 * mm->context.allow_cow_sharing is set to 0.
2614 */
2615static int __s390_unshare_zeropages(struct mm_struct *mm)
2616{
 
2617	struct vm_area_struct *vma;
 
2618	VMA_ITERATOR(vmi, mm, 0);
2619	unsigned long addr;
2620	vm_fault_t fault;
2621	int rc;
2622
2623	for_each_vma(vmi, vma) {
2624		/*
2625		 * We could only look at COW mappings, but it's more future
2626		 * proof to catch unexpected zeropages in other mappings and
2627		 * fail.
2628		 */
2629		if ((vma->vm_flags & VM_PFNMAP) || is_vm_hugetlb_page(vma))
2630			continue;
2631		addr = vma->vm_start;
2632
2633retry:
2634		rc = walk_page_range_vma(vma, addr, vma->vm_end,
2635					 &find_zeropage_ops, &addr);
2636		if (rc < 0)
2637			return rc;
2638		else if (!rc)
2639			continue;
2640
2641		/* addr was updated by find_zeropage_pte_entry() */
2642		fault = handle_mm_fault(vma, addr,
2643					FAULT_FLAG_UNSHARE | FAULT_FLAG_REMOTE,
2644					NULL);
2645		if (fault & VM_FAULT_OOM)
2646			return -ENOMEM;
2647		/*
2648		 * See break_ksm(): even after handle_mm_fault() returned 0, we
2649		 * must start the lookup from the current address, because
2650		 * handle_mm_fault() may back out if there's any difficulty.
2651		 *
2652		 * VM_FAULT_SIGBUS and VM_FAULT_SIGSEGV are unexpected but
2653		 * maybe they could trigger in the future on concurrent
2654		 * truncation. In that case, the shared zeropage would be gone
2655		 * and we can simply retry and make progress.
2656		 */
2657		cond_resched();
2658		goto retry;
2659	}
2660
2661	return 0;
2662}
2663
2664static int __s390_disable_cow_sharing(struct mm_struct *mm)
2665{
2666	int rc;
2667
2668	if (!mm->context.allow_cow_sharing)
2669		return 0;
2670
2671	mm->context.allow_cow_sharing = 0;
2672
2673	/* Replace all shared zeropages by anonymous pages. */
2674	rc = __s390_unshare_zeropages(mm);
2675	/*
2676	 * Make sure to disable KSM (if enabled for the whole process or
2677	 * individual VMAs). Note that nothing currently hinders user space
2678	 * from re-enabling it.
2679	 */
2680	if (!rc)
2681		rc = ksm_disable(mm);
2682	if (rc)
2683		mm->context.allow_cow_sharing = 1;
2684	return rc;
2685}
2686
2687/*
2688 * Disable most COW-sharing of memory pages for the whole process:
2689 * (1) Disable KSM and unmerge/unshare any KSM pages.
2690 * (2) Disallow shared zeropages and unshare any zerpages that are mapped.
2691 *
2692 * Not that we currently don't bother with COW-shared pages that are shared
2693 * with parent/child processes due to fork().
2694 */
2695int s390_disable_cow_sharing(void)
2696{
2697	int rc;
2698
2699	mmap_write_lock(current->mm);
2700	rc = __s390_disable_cow_sharing(current->mm);
2701	mmap_write_unlock(current->mm);
2702	return rc;
2703}
2704EXPORT_SYMBOL_GPL(s390_disable_cow_sharing);
2705
2706/*
2707 * Enable storage key handling from now on and initialize the storage
2708 * keys with the default key.
2709 */
2710static int __s390_enable_skey_pte(pte_t *pte, unsigned long addr,
2711				  unsigned long next, struct mm_walk *walk)
2712{
2713	/* Clear storage key */
2714	ptep_zap_key(walk->mm, addr, pte);
2715	return 0;
2716}
2717
2718/*
2719 * Give a chance to schedule after setting a key to 256 pages.
2720 * We only hold the mm lock, which is a rwsem and the kvm srcu.
2721 * Both can sleep.
2722 */
2723static int __s390_enable_skey_pmd(pmd_t *pmd, unsigned long addr,
2724				  unsigned long next, struct mm_walk *walk)
2725{
2726	cond_resched();
2727	return 0;
2728}
2729
2730static int __s390_enable_skey_hugetlb(pte_t *pte, unsigned long addr,
2731				      unsigned long hmask, unsigned long next,
2732				      struct mm_walk *walk)
2733{
2734	pmd_t *pmd = (pmd_t *)pte;
2735	unsigned long start, end;
2736	struct page *page = pmd_page(*pmd);
2737
2738	/*
2739	 * The write check makes sure we do not set a key on shared
2740	 * memory. This is needed as the walker does not differentiate
2741	 * between actual guest memory and the process executable or
2742	 * shared libraries.
2743	 */
2744	if (pmd_val(*pmd) & _SEGMENT_ENTRY_INVALID ||
2745	    !(pmd_val(*pmd) & _SEGMENT_ENTRY_WRITE))
2746		return 0;
2747
2748	start = pmd_val(*pmd) & HPAGE_MASK;
2749	end = start + HPAGE_SIZE;
2750	__storage_key_init_range(start, end);
2751	set_bit(PG_arch_1, &page->flags);
2752	cond_resched();
2753	return 0;
2754}
2755
2756static const struct mm_walk_ops enable_skey_walk_ops = {
2757	.hugetlb_entry		= __s390_enable_skey_hugetlb,
2758	.pte_entry		= __s390_enable_skey_pte,
2759	.pmd_entry		= __s390_enable_skey_pmd,
2760	.walk_lock		= PGWALK_WRLOCK,
2761};
2762
2763int s390_enable_skey(void)
2764{
2765	struct mm_struct *mm = current->mm;
2766	int rc = 0;
2767
2768	mmap_write_lock(mm);
2769	if (mm_uses_skeys(mm))
2770		goto out_up;
2771
2772	mm->context.uses_skeys = 1;
2773	rc = __s390_disable_cow_sharing(mm);
2774	if (rc) {
2775		mm->context.uses_skeys = 0;
2776		goto out_up;
2777	}
2778	walk_page_range(mm, 0, TASK_SIZE, &enable_skey_walk_ops, NULL);
2779
2780out_up:
2781	mmap_write_unlock(mm);
2782	return rc;
2783}
2784EXPORT_SYMBOL_GPL(s390_enable_skey);
2785
2786/*
2787 * Reset CMMA state, make all pages stable again.
2788 */
2789static int __s390_reset_cmma(pte_t *pte, unsigned long addr,
2790			     unsigned long next, struct mm_walk *walk)
2791{
2792	ptep_zap_unused(walk->mm, addr, pte, 1);
2793	return 0;
2794}
2795
2796static const struct mm_walk_ops reset_cmma_walk_ops = {
2797	.pte_entry		= __s390_reset_cmma,
2798	.walk_lock		= PGWALK_WRLOCK,
2799};
2800
2801void s390_reset_cmma(struct mm_struct *mm)
2802{
2803	mmap_write_lock(mm);
2804	walk_page_range(mm, 0, TASK_SIZE, &reset_cmma_walk_ops, NULL);
2805	mmap_write_unlock(mm);
2806}
2807EXPORT_SYMBOL_GPL(s390_reset_cmma);
2808
2809#define GATHER_GET_PAGES 32
2810
2811struct reset_walk_state {
2812	unsigned long next;
2813	unsigned long count;
2814	unsigned long pfns[GATHER_GET_PAGES];
2815};
2816
2817static int s390_gather_pages(pte_t *ptep, unsigned long addr,
2818			     unsigned long next, struct mm_walk *walk)
2819{
2820	struct reset_walk_state *p = walk->private;
2821	pte_t pte = READ_ONCE(*ptep);
2822
2823	if (pte_present(pte)) {
2824		/* we have a reference from the mapping, take an extra one */
2825		get_page(phys_to_page(pte_val(pte)));
2826		p->pfns[p->count] = phys_to_pfn(pte_val(pte));
2827		p->next = next;
2828		p->count++;
2829	}
2830	return p->count >= GATHER_GET_PAGES;
2831}
2832
2833static const struct mm_walk_ops gather_pages_ops = {
2834	.pte_entry = s390_gather_pages,
2835	.walk_lock = PGWALK_RDLOCK,
2836};
2837
2838/*
2839 * Call the Destroy secure page UVC on each page in the given array of PFNs.
2840 * Each page needs to have an extra reference, which will be released here.
2841 */
2842void s390_uv_destroy_pfns(unsigned long count, unsigned long *pfns)
2843{
2844	unsigned long i;
2845
2846	for (i = 0; i < count; i++) {
2847		/* we always have an extra reference */
2848		uv_destroy_owned_page(pfn_to_phys(pfns[i]));
2849		/* get rid of the extra reference */
2850		put_page(pfn_to_page(pfns[i]));
2851		cond_resched();
2852	}
2853}
2854EXPORT_SYMBOL_GPL(s390_uv_destroy_pfns);
2855
2856/**
2857 * __s390_uv_destroy_range - Call the destroy secure page UVC on each page
2858 * in the given range of the given address space.
2859 * @mm: the mm to operate on
2860 * @start: the start of the range
2861 * @end: the end of the range
2862 * @interruptible: if not 0, stop when a fatal signal is received
2863 *
2864 * Walk the given range of the given address space and call the destroy
2865 * secure page UVC on each page. Optionally exit early if a fatal signal is
2866 * pending.
2867 *
2868 * Return: 0 on success, -EINTR if the function stopped before completing
2869 */
2870int __s390_uv_destroy_range(struct mm_struct *mm, unsigned long start,
2871			    unsigned long end, bool interruptible)
2872{
2873	struct reset_walk_state state = { .next = start };
2874	int r = 1;
2875
2876	while (r > 0) {
2877		state.count = 0;
2878		mmap_read_lock(mm);
2879		r = walk_page_range(mm, state.next, end, &gather_pages_ops, &state);
2880		mmap_read_unlock(mm);
2881		cond_resched();
2882		s390_uv_destroy_pfns(state.count, state.pfns);
2883		if (interruptible && fatal_signal_pending(current))
2884			return -EINTR;
2885	}
2886	return 0;
2887}
2888EXPORT_SYMBOL_GPL(__s390_uv_destroy_range);
2889
2890/**
2891 * s390_unlist_old_asce - Remove the topmost level of page tables from the
2892 * list of page tables of the gmap.
2893 * @gmap: the gmap whose table is to be removed
2894 *
2895 * On s390x, KVM keeps a list of all pages containing the page tables of the
2896 * gmap (the CRST list). This list is used at tear down time to free all
2897 * pages that are now not needed anymore.
2898 *
2899 * This function removes the topmost page of the tree (the one pointed to by
2900 * the ASCE) from the CRST list.
2901 *
2902 * This means that it will not be freed when the VM is torn down, and needs
2903 * to be handled separately by the caller, unless a leak is actually
2904 * intended. Notice that this function will only remove the page from the
2905 * list, the page will still be used as a top level page table (and ASCE).
2906 */
2907void s390_unlist_old_asce(struct gmap *gmap)
2908{
2909	struct page *old;
2910
2911	old = virt_to_page(gmap->table);
2912	spin_lock(&gmap->guest_table_lock);
2913	list_del(&old->lru);
2914	/*
2915	 * Sometimes the topmost page might need to be "removed" multiple
2916	 * times, for example if the VM is rebooted into secure mode several
2917	 * times concurrently, or if s390_replace_asce fails after calling
2918	 * s390_remove_old_asce and is attempted again later. In that case
2919	 * the old asce has been removed from the list, and therefore it
2920	 * will not be freed when the VM terminates, but the ASCE is still
2921	 * in use and still pointed to.
2922	 * A subsequent call to replace_asce will follow the pointer and try
2923	 * to remove the same page from the list again.
2924	 * Therefore it's necessary that the page of the ASCE has valid
2925	 * pointers, so list_del can work (and do nothing) without
2926	 * dereferencing stale or invalid pointers.
2927	 */
2928	INIT_LIST_HEAD(&old->lru);
2929	spin_unlock(&gmap->guest_table_lock);
2930}
2931EXPORT_SYMBOL_GPL(s390_unlist_old_asce);
2932
2933/**
2934 * s390_replace_asce - Try to replace the current ASCE of a gmap with a copy
2935 * @gmap: the gmap whose ASCE needs to be replaced
2936 *
2937 * If the ASCE is a SEGMENT type then this function will return -EINVAL,
2938 * otherwise the pointers in the host_to_guest radix tree will keep pointing
2939 * to the wrong pages, causing use-after-free and memory corruption.
2940 * If the allocation of the new top level page table fails, the ASCE is not
2941 * replaced.
2942 * In any case, the old ASCE is always removed from the gmap CRST list.
2943 * Therefore the caller has to make sure to save a pointer to it
2944 * beforehand, unless a leak is actually intended.
2945 */
2946int s390_replace_asce(struct gmap *gmap)
2947{
2948	unsigned long asce;
2949	struct page *page;
2950	void *table;
2951
2952	s390_unlist_old_asce(gmap);
2953
2954	/* Replacing segment type ASCEs would cause serious issues */
2955	if ((gmap->asce & _ASCE_TYPE_MASK) == _ASCE_TYPE_SEGMENT)
2956		return -EINVAL;
2957
2958	page = gmap_alloc_crst();
2959	if (!page)
2960		return -ENOMEM;
2961	page->index = 0;
2962	table = page_to_virt(page);
2963	memcpy(table, gmap->table, 1UL << (CRST_ALLOC_ORDER + PAGE_SHIFT));
2964
2965	/*
2966	 * The caller has to deal with the old ASCE, but here we make sure
2967	 * the new one is properly added to the CRST list, so that
2968	 * it will be freed when the VM is torn down.
2969	 */
2970	spin_lock(&gmap->guest_table_lock);
2971	list_add(&page->lru, &gmap->crst_list);
2972	spin_unlock(&gmap->guest_table_lock);
2973
2974	/* Set new table origin while preserving existing ASCE control bits */
2975	asce = (gmap->asce & ~_ASCE_ORIGIN) | __pa(table);
2976	WRITE_ONCE(gmap->asce, asce);
2977	WRITE_ONCE(gmap->mm->context.gmap_asce, asce);
2978	WRITE_ONCE(gmap->table, table);
2979
2980	return 0;
2981}
2982EXPORT_SYMBOL_GPL(s390_replace_asce);