Linux Audio

Check our new training course

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