Linux Audio

Check our new training course

Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Generic hugetlb support.
   4 * (C) Nadia Yvette Chambers, April 2004
   5 */
   6#include <linux/list.h>
   7#include <linux/init.h>
 
   8#include <linux/mm.h>
   9#include <linux/seq_file.h>
  10#include <linux/sysctl.h>
  11#include <linux/highmem.h>
  12#include <linux/mmu_notifier.h>
  13#include <linux/nodemask.h>
  14#include <linux/pagemap.h>
  15#include <linux/mempolicy.h>
  16#include <linux/compiler.h>
  17#include <linux/cpuset.h>
  18#include <linux/mutex.h>
  19#include <linux/memblock.h>
  20#include <linux/sysfs.h>
  21#include <linux/slab.h>
  22#include <linux/mmdebug.h>
  23#include <linux/sched/signal.h>
  24#include <linux/rmap.h>
  25#include <linux/string_helpers.h>
  26#include <linux/swap.h>
  27#include <linux/swapops.h>
  28#include <linux/jhash.h>
  29#include <linux/numa.h>
  30
  31#include <asm/page.h>
  32#include <asm/pgtable.h>
  33#include <asm/tlb.h>
  34
  35#include <linux/io.h>
 
  36#include <linux/hugetlb.h>
  37#include <linux/hugetlb_cgroup.h>
  38#include <linux/node.h>
  39#include <linux/userfaultfd_k.h>
  40#include <linux/page_owner.h>
  41#include "internal.h"
  42
  43int hugetlb_max_hstate __read_mostly;
 
 
 
 
  44unsigned int default_hstate_idx;
  45struct hstate hstates[HUGE_MAX_HSTATE];
  46/*
  47 * Minimum page order among possible hugepage sizes, set to a proper value
  48 * at boot time.
  49 */
  50static unsigned int minimum_order __read_mostly = UINT_MAX;
  51
  52__initdata LIST_HEAD(huge_boot_pages);
  53
  54/* for command line parsing */
  55static struct hstate * __initdata parsed_hstate;
  56static unsigned long __initdata default_hstate_max_huge_pages;
  57static unsigned long __initdata default_hstate_size;
  58static bool __initdata parsed_valid_hugepagesz = true;
  59
  60/*
  61 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
  62 * free_huge_pages, and surplus_huge_pages.
  63 */
  64DEFINE_SPINLOCK(hugetlb_lock);
  65
  66/*
  67 * Serializes faults on the same logical page.  This is used to
  68 * prevent spurious OOMs when the hugepage pool is fully utilized.
  69 */
  70static int num_fault_mutexes;
  71struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
  72
  73/* Forward declaration */
  74static int hugetlb_acct_memory(struct hstate *h, long delta);
  75
  76static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
  77{
  78	bool free = (spool->count == 0) && (spool->used_hpages == 0);
  79
  80	spin_unlock(&spool->lock);
  81
  82	/* If no pages are used, and no other handles to the subpool
  83	 * remain, give up any reservations mased on minimum size and
  84	 * free the subpool */
  85	if (free) {
  86		if (spool->min_hpages != -1)
  87			hugetlb_acct_memory(spool->hstate,
  88						-spool->min_hpages);
  89		kfree(spool);
  90	}
  91}
  92
  93struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
  94						long min_hpages)
  95{
  96	struct hugepage_subpool *spool;
  97
  98	spool = kzalloc(sizeof(*spool), GFP_KERNEL);
  99	if (!spool)
 100		return NULL;
 101
 102	spin_lock_init(&spool->lock);
 103	spool->count = 1;
 104	spool->max_hpages = max_hpages;
 105	spool->hstate = h;
 106	spool->min_hpages = min_hpages;
 107
 108	if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
 109		kfree(spool);
 110		return NULL;
 111	}
 112	spool->rsv_hpages = min_hpages;
 113
 114	return spool;
 115}
 116
 117void hugepage_put_subpool(struct hugepage_subpool *spool)
 118{
 119	spin_lock(&spool->lock);
 120	BUG_ON(!spool->count);
 121	spool->count--;
 122	unlock_or_release_subpool(spool);
 123}
 124
 125/*
 126 * Subpool accounting for allocating and reserving pages.
 127 * Return -ENOMEM if there are not enough resources to satisfy the
 128 * the request.  Otherwise, return the number of pages by which the
 129 * global pools must be adjusted (upward).  The returned value may
 130 * only be different than the passed value (delta) in the case where
 131 * a subpool minimum size must be manitained.
 132 */
 133static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
 134				      long delta)
 135{
 136	long ret = delta;
 137
 138	if (!spool)
 139		return ret;
 140
 141	spin_lock(&spool->lock);
 142
 143	if (spool->max_hpages != -1) {		/* maximum size accounting */
 144		if ((spool->used_hpages + delta) <= spool->max_hpages)
 145			spool->used_hpages += delta;
 146		else {
 147			ret = -ENOMEM;
 148			goto unlock_ret;
 149		}
 150	}
 151
 152	/* minimum size accounting */
 153	if (spool->min_hpages != -1 && spool->rsv_hpages) {
 154		if (delta > spool->rsv_hpages) {
 155			/*
 156			 * Asking for more reserves than those already taken on
 157			 * behalf of subpool.  Return difference.
 158			 */
 159			ret = delta - spool->rsv_hpages;
 160			spool->rsv_hpages = 0;
 161		} else {
 162			ret = 0;	/* reserves already accounted for */
 163			spool->rsv_hpages -= delta;
 164		}
 165	}
 166
 167unlock_ret:
 168	spin_unlock(&spool->lock);
 
 169	return ret;
 170}
 171
 172/*
 173 * Subpool accounting for freeing and unreserving pages.
 174 * Return the number of global page reservations that must be dropped.
 175 * The return value may only be different than the passed value (delta)
 176 * in the case where a subpool minimum size must be maintained.
 177 */
 178static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
 179				       long delta)
 180{
 181	long ret = delta;
 182
 183	if (!spool)
 184		return delta;
 185
 186	spin_lock(&spool->lock);
 187
 188	if (spool->max_hpages != -1)		/* maximum size accounting */
 189		spool->used_hpages -= delta;
 190
 191	 /* minimum size accounting */
 192	if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
 193		if (spool->rsv_hpages + delta <= spool->min_hpages)
 194			ret = 0;
 195		else
 196			ret = spool->rsv_hpages + delta - spool->min_hpages;
 197
 198		spool->rsv_hpages += delta;
 199		if (spool->rsv_hpages > spool->min_hpages)
 200			spool->rsv_hpages = spool->min_hpages;
 201	}
 202
 203	/*
 204	 * If hugetlbfs_put_super couldn't free spool due to an outstanding
 205	 * quota reference, free it now.
 206	 */
 207	unlock_or_release_subpool(spool);
 208
 209	return ret;
 210}
 211
 212static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
 213{
 214	return HUGETLBFS_SB(inode->i_sb)->spool;
 215}
 216
 217static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
 218{
 219	return subpool_inode(file_inode(vma->vm_file));
 220}
 221
 222/*
 223 * Region tracking -- allows tracking of reservations and instantiated pages
 224 *                    across the pages in a mapping.
 225 *
 226 * The region data structures are embedded into a resv_map and protected
 227 * by a resv_map's lock.  The set of regions within the resv_map represent
 228 * reservations for huge pages, or huge pages that have already been
 229 * instantiated within the map.  The from and to elements are huge page
 230 * indicies into the associated mapping.  from indicates the starting index
 231 * of the region.  to represents the first index past the end of  the region.
 232 *
 233 * For example, a file region structure with from == 0 and to == 4 represents
 234 * four huge pages in a mapping.  It is important to note that the to element
 235 * represents the first element past the end of the region. This is used in
 236 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
 237 *
 238 * Interval notation of the form [from, to) will be used to indicate that
 239 * the endpoint from is inclusive and to is exclusive.
 240 */
 241struct file_region {
 242	struct list_head link;
 243	long from;
 244	long to;
 245};
 246
 247/*
 248 * Add the huge page range represented by [f, t) to the reserve
 249 * map.  In the normal case, existing regions will be expanded
 250 * to accommodate the specified range.  Sufficient regions should
 251 * exist for expansion due to the previous call to region_chg
 252 * with the same range.  However, it is possible that region_del
 253 * could have been called after region_chg and modifed the map
 254 * in such a way that no region exists to be expanded.  In this
 255 * case, pull a region descriptor from the cache associated with
 256 * the map and use that for the new range.
 257 *
 258 * Return the number of new huge pages added to the map.  This
 259 * number is greater than or equal to zero.
 260 */
 261static long region_add(struct resv_map *resv, long f, long t)
 262{
 263	struct list_head *head = &resv->regions;
 264	struct file_region *rg, *nrg, *trg;
 265	long add = 0;
 266
 267	spin_lock(&resv->lock);
 268	/* Locate the region we are either in or before. */
 269	list_for_each_entry(rg, head, link)
 270		if (f <= rg->to)
 271			break;
 272
 273	/*
 274	 * If no region exists which can be expanded to include the
 275	 * specified range, the list must have been modified by an
 276	 * interleving call to region_del().  Pull a region descriptor
 277	 * from the cache and use it for this range.
 278	 */
 279	if (&rg->link == head || t < rg->from) {
 280		VM_BUG_ON(resv->region_cache_count <= 0);
 281
 282		resv->region_cache_count--;
 283		nrg = list_first_entry(&resv->region_cache, struct file_region,
 284					link);
 285		list_del(&nrg->link);
 286
 287		nrg->from = f;
 288		nrg->to = t;
 289		list_add(&nrg->link, rg->link.prev);
 290
 291		add += t - f;
 292		goto out_locked;
 293	}
 294
 295	/* Round our left edge to the current segment if it encloses us. */
 296	if (f > rg->from)
 297		f = rg->from;
 298
 299	/* Check for and consume any regions we now overlap with. */
 300	nrg = rg;
 301	list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 302		if (&rg->link == head)
 303			break;
 304		if (rg->from > t)
 305			break;
 306
 307		/* If this area reaches higher then extend our area to
 308		 * include it completely.  If this is not the first area
 309		 * which we intend to reuse, free it. */
 310		if (rg->to > t)
 311			t = rg->to;
 312		if (rg != nrg) {
 313			/* Decrement return value by the deleted range.
 314			 * Another range will span this area so that by
 315			 * end of routine add will be >= zero
 316			 */
 317			add -= (rg->to - rg->from);
 318			list_del(&rg->link);
 319			kfree(rg);
 320		}
 321	}
 322
 323	add += (nrg->from - f);		/* Added to beginning of region */
 324	nrg->from = f;
 325	add += t - nrg->to;		/* Added to end of region */
 326	nrg->to = t;
 327
 328out_locked:
 329	resv->adds_in_progress--;
 330	spin_unlock(&resv->lock);
 331	VM_BUG_ON(add < 0);
 332	return add;
 333}
 334
 335/*
 336 * Examine the existing reserve map and determine how many
 337 * huge pages in the specified range [f, t) are NOT currently
 338 * represented.  This routine is called before a subsequent
 339 * call to region_add that will actually modify the reserve
 340 * map to add the specified range [f, t).  region_chg does
 341 * not change the number of huge pages represented by the
 342 * map.  However, if the existing regions in the map can not
 343 * be expanded to represent the new range, a new file_region
 344 * structure is added to the map as a placeholder.  This is
 345 * so that the subsequent region_add call will have all the
 346 * regions it needs and will not fail.
 347 *
 348 * Upon entry, region_chg will also examine the cache of region descriptors
 349 * associated with the map.  If there are not enough descriptors cached, one
 350 * will be allocated for the in progress add operation.
 351 *
 352 * Returns the number of huge pages that need to be added to the existing
 353 * reservation map for the range [f, t).  This number is greater or equal to
 354 * zero.  -ENOMEM is returned if a new file_region structure or cache entry
 355 * is needed and can not be allocated.
 356 */
 357static long region_chg(struct resv_map *resv, long f, long t)
 358{
 359	struct list_head *head = &resv->regions;
 360	struct file_region *rg, *nrg = NULL;
 361	long chg = 0;
 362
 363retry:
 364	spin_lock(&resv->lock);
 365retry_locked:
 366	resv->adds_in_progress++;
 367
 368	/*
 369	 * Check for sufficient descriptors in the cache to accommodate
 370	 * the number of in progress add operations.
 371	 */
 372	if (resv->adds_in_progress > resv->region_cache_count) {
 373		struct file_region *trg;
 374
 375		VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
 376		/* Must drop lock to allocate a new descriptor. */
 377		resv->adds_in_progress--;
 378		spin_unlock(&resv->lock);
 379
 380		trg = kmalloc(sizeof(*trg), GFP_KERNEL);
 381		if (!trg) {
 382			kfree(nrg);
 383			return -ENOMEM;
 384		}
 385
 386		spin_lock(&resv->lock);
 387		list_add(&trg->link, &resv->region_cache);
 388		resv->region_cache_count++;
 389		goto retry_locked;
 390	}
 391
 392	/* Locate the region we are before or in. */
 393	list_for_each_entry(rg, head, link)
 394		if (f <= rg->to)
 395			break;
 396
 397	/* If we are below the current region then a new region is required.
 398	 * Subtle, allocate a new region at the position but make it zero
 399	 * size such that we can guarantee to record the reservation. */
 400	if (&rg->link == head || t < rg->from) {
 401		if (!nrg) {
 402			resv->adds_in_progress--;
 403			spin_unlock(&resv->lock);
 404			nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 405			if (!nrg)
 406				return -ENOMEM;
 407
 408			nrg->from = f;
 409			nrg->to   = f;
 410			INIT_LIST_HEAD(&nrg->link);
 411			goto retry;
 412		}
 413
 414		list_add(&nrg->link, rg->link.prev);
 415		chg = t - f;
 416		goto out_nrg;
 417	}
 418
 419	/* Round our left edge to the current segment if it encloses us. */
 420	if (f > rg->from)
 421		f = rg->from;
 422	chg = t - f;
 423
 424	/* Check for and consume any regions we now overlap with. */
 425	list_for_each_entry(rg, rg->link.prev, link) {
 426		if (&rg->link == head)
 427			break;
 428		if (rg->from > t)
 429			goto out;
 430
 431		/* We overlap with this area, if it extends further than
 432		 * us then we must extend ourselves.  Account for its
 433		 * existing reservation. */
 434		if (rg->to > t) {
 435			chg += rg->to - t;
 436			t = rg->to;
 437		}
 438		chg -= rg->to - rg->from;
 439	}
 440
 441out:
 442	spin_unlock(&resv->lock);
 443	/*  We already know we raced and no longer need the new region */
 444	kfree(nrg);
 445	return chg;
 446out_nrg:
 447	spin_unlock(&resv->lock);
 448	return chg;
 449}
 450
 451/*
 452 * Abort the in progress add operation.  The adds_in_progress field
 453 * of the resv_map keeps track of the operations in progress between
 454 * calls to region_chg and region_add.  Operations are sometimes
 455 * aborted after the call to region_chg.  In such cases, region_abort
 456 * is called to decrement the adds_in_progress counter.
 457 *
 458 * NOTE: The range arguments [f, t) are not needed or used in this
 459 * routine.  They are kept to make reading the calling code easier as
 460 * arguments will match the associated region_chg call.
 461 */
 462static void region_abort(struct resv_map *resv, long f, long t)
 463{
 464	spin_lock(&resv->lock);
 465	VM_BUG_ON(!resv->region_cache_count);
 466	resv->adds_in_progress--;
 467	spin_unlock(&resv->lock);
 468}
 469
 470/*
 471 * Delete the specified range [f, t) from the reserve map.  If the
 472 * t parameter is LONG_MAX, this indicates that ALL regions after f
 473 * should be deleted.  Locate the regions which intersect [f, t)
 474 * and either trim, delete or split the existing regions.
 475 *
 476 * Returns the number of huge pages deleted from the reserve map.
 477 * In the normal case, the return value is zero or more.  In the
 478 * case where a region must be split, a new region descriptor must
 479 * be allocated.  If the allocation fails, -ENOMEM will be returned.
 480 * NOTE: If the parameter t == LONG_MAX, then we will never split
 481 * a region and possibly return -ENOMEM.  Callers specifying
 482 * t == LONG_MAX do not need to check for -ENOMEM error.
 483 */
 484static long region_del(struct resv_map *resv, long f, long t)
 485{
 486	struct list_head *head = &resv->regions;
 487	struct file_region *rg, *trg;
 488	struct file_region *nrg = NULL;
 489	long del = 0;
 490
 491retry:
 492	spin_lock(&resv->lock);
 493	list_for_each_entry_safe(rg, trg, head, link) {
 494		/*
 495		 * Skip regions before the range to be deleted.  file_region
 496		 * ranges are normally of the form [from, to).  However, there
 497		 * may be a "placeholder" entry in the map which is of the form
 498		 * (from, to) with from == to.  Check for placeholder entries
 499		 * at the beginning of the range to be deleted.
 500		 */
 501		if (rg->to <= f && (rg->to != rg->from || rg->to != f))
 502			continue;
 503
 504		if (rg->from >= t)
 505			break;
 506
 507		if (f > rg->from && t < rg->to) { /* Must split region */
 508			/*
 509			 * Check for an entry in the cache before dropping
 510			 * lock and attempting allocation.
 511			 */
 512			if (!nrg &&
 513			    resv->region_cache_count > resv->adds_in_progress) {
 514				nrg = list_first_entry(&resv->region_cache,
 515							struct file_region,
 516							link);
 517				list_del(&nrg->link);
 518				resv->region_cache_count--;
 519			}
 520
 521			if (!nrg) {
 522				spin_unlock(&resv->lock);
 523				nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 524				if (!nrg)
 525					return -ENOMEM;
 526				goto retry;
 527			}
 528
 529			del += t - f;
 530
 531			/* New entry for end of split region */
 532			nrg->from = t;
 533			nrg->to = rg->to;
 534			INIT_LIST_HEAD(&nrg->link);
 535
 536			/* Original entry is trimmed */
 537			rg->to = f;
 538
 539			list_add(&nrg->link, &rg->link);
 540			nrg = NULL;
 
 541			break;
 542		}
 543
 544		if (f <= rg->from && t >= rg->to) { /* Remove entire region */
 545			del += rg->to - rg->from;
 546			list_del(&rg->link);
 547			kfree(rg);
 548			continue;
 549		}
 550
 551		if (f <= rg->from) {	/* Trim beginning of region */
 552			del += t - rg->from;
 553			rg->from = t;
 554		} else {		/* Trim end of region */
 555			del += rg->to - f;
 556			rg->to = f;
 557		}
 558	}
 559
 560	spin_unlock(&resv->lock);
 561	kfree(nrg);
 562	return del;
 563}
 564
 565/*
 566 * A rare out of memory error was encountered which prevented removal of
 567 * the reserve map region for a page.  The huge page itself was free'ed
 568 * and removed from the page cache.  This routine will adjust the subpool
 569 * usage count, and the global reserve count if needed.  By incrementing
 570 * these counts, the reserve map entry which could not be deleted will
 571 * appear as a "reserved" entry instead of simply dangling with incorrect
 572 * counts.
 573 */
 574void hugetlb_fix_reserve_counts(struct inode *inode)
 575{
 576	struct hugepage_subpool *spool = subpool_inode(inode);
 577	long rsv_adjust;
 578
 579	rsv_adjust = hugepage_subpool_get_pages(spool, 1);
 580	if (rsv_adjust) {
 581		struct hstate *h = hstate_inode(inode);
 582
 583		hugetlb_acct_memory(h, 1);
 584	}
 
 585}
 586
 587/*
 588 * Count and return the number of huge pages in the reserve map
 589 * that intersect with the range [f, t).
 590 */
 591static long region_count(struct resv_map *resv, long f, long t)
 592{
 593	struct list_head *head = &resv->regions;
 594	struct file_region *rg;
 595	long chg = 0;
 596
 597	spin_lock(&resv->lock);
 598	/* Locate each segment we overlap with, and count that overlap. */
 599	list_for_each_entry(rg, head, link) {
 600		long seg_from;
 601		long seg_to;
 602
 603		if (rg->to <= f)
 604			continue;
 605		if (rg->from >= t)
 606			break;
 607
 608		seg_from = max(rg->from, f);
 609		seg_to = min(rg->to, t);
 610
 611		chg += seg_to - seg_from;
 612	}
 613	spin_unlock(&resv->lock);
 614
 615	return chg;
 616}
 617
 618/*
 619 * Convert the address within this vma to the page offset within
 620 * the mapping, in pagecache page units; huge pages here.
 621 */
 622static pgoff_t vma_hugecache_offset(struct hstate *h,
 623			struct vm_area_struct *vma, unsigned long address)
 624{
 625	return ((address - vma->vm_start) >> huge_page_shift(h)) +
 626			(vma->vm_pgoff >> huge_page_order(h));
 627}
 628
 629pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
 630				     unsigned long address)
 631{
 632	return vma_hugecache_offset(hstate_vma(vma), vma, address);
 633}
 634EXPORT_SYMBOL_GPL(linear_hugepage_index);
 635
 636/*
 637 * Return the size of the pages allocated when backing a VMA. In the majority
 638 * cases this will be same size as used by the page table entries.
 639 */
 640unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
 641{
 642	if (vma->vm_ops && vma->vm_ops->pagesize)
 643		return vma->vm_ops->pagesize(vma);
 644	return PAGE_SIZE;
 
 
 
 
 
 645}
 646EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
 647
 648/*
 649 * Return the page size being used by the MMU to back a VMA. In the majority
 650 * of cases, the page size used by the kernel matches the MMU size. On
 651 * architectures where it differs, an architecture-specific 'strong'
 652 * version of this symbol is required.
 653 */
 654__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 
 655{
 656	return vma_kernel_pagesize(vma);
 657}
 
 658
 659/*
 660 * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
 661 * bits of the reservation map pointer, which are always clear due to
 662 * alignment.
 663 */
 664#define HPAGE_RESV_OWNER    (1UL << 0)
 665#define HPAGE_RESV_UNMAPPED (1UL << 1)
 666#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
 667
 668/*
 669 * These helpers are used to track how many pages are reserved for
 670 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
 671 * is guaranteed to have their future faults succeed.
 672 *
 673 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
 674 * the reserve counters are updated with the hugetlb_lock held. It is safe
 675 * to reset the VMA at fork() time as it is not in use yet and there is no
 676 * chance of the global counters getting corrupted as a result of the values.
 677 *
 678 * The private mapping reservation is represented in a subtly different
 679 * manner to a shared mapping.  A shared mapping has a region map associated
 680 * with the underlying file, this region map represents the backing file
 681 * pages which have ever had a reservation assigned which this persists even
 682 * after the page is instantiated.  A private mapping has a region map
 683 * associated with the original mmap which is attached to all VMAs which
 684 * reference it, this region map represents those offsets which have consumed
 685 * reservation ie. where pages have been instantiated.
 686 */
 687static unsigned long get_vma_private_data(struct vm_area_struct *vma)
 688{
 689	return (unsigned long)vma->vm_private_data;
 690}
 691
 692static void set_vma_private_data(struct vm_area_struct *vma,
 693							unsigned long value)
 694{
 695	vma->vm_private_data = (void *)value;
 696}
 697
 698struct resv_map *resv_map_alloc(void)
 
 
 
 
 
 699{
 700	struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
 701	struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
 702
 703	if (!resv_map || !rg) {
 704		kfree(resv_map);
 705		kfree(rg);
 706		return NULL;
 707	}
 708
 709	kref_init(&resv_map->refs);
 710	spin_lock_init(&resv_map->lock);
 711	INIT_LIST_HEAD(&resv_map->regions);
 712
 713	resv_map->adds_in_progress = 0;
 714
 715	INIT_LIST_HEAD(&resv_map->region_cache);
 716	list_add(&rg->link, &resv_map->region_cache);
 717	resv_map->region_cache_count = 1;
 718
 719	return resv_map;
 720}
 721
 722void resv_map_release(struct kref *ref)
 723{
 724	struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 725	struct list_head *head = &resv_map->region_cache;
 726	struct file_region *rg, *trg;
 727
 728	/* Clear out any active regions before we release the map. */
 729	region_del(resv_map, 0, LONG_MAX);
 730
 731	/* ... and any entries left in the cache */
 732	list_for_each_entry_safe(rg, trg, head, link) {
 733		list_del(&rg->link);
 734		kfree(rg);
 735	}
 736
 737	VM_BUG_ON(resv_map->adds_in_progress);
 738
 739	kfree(resv_map);
 740}
 741
 742static inline struct resv_map *inode_resv_map(struct inode *inode)
 743{
 744	/*
 745	 * At inode evict time, i_mapping may not point to the original
 746	 * address space within the inode.  This original address space
 747	 * contains the pointer to the resv_map.  So, always use the
 748	 * address space embedded within the inode.
 749	 * The VERY common case is inode->mapping == &inode->i_data but,
 750	 * this may not be true for device special inodes.
 751	 */
 752	return (struct resv_map *)(&inode->i_data)->private_data;
 753}
 754
 755static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 756{
 757	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
 758	if (vma->vm_flags & VM_MAYSHARE) {
 759		struct address_space *mapping = vma->vm_file->f_mapping;
 760		struct inode *inode = mapping->host;
 761
 762		return inode_resv_map(inode);
 763
 764	} else {
 765		return (struct resv_map *)(get_vma_private_data(vma) &
 766							~HPAGE_RESV_MASK);
 767	}
 768}
 769
 770static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 771{
 772	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
 773	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
 774
 775	set_vma_private_data(vma, (get_vma_private_data(vma) &
 776				HPAGE_RESV_MASK) | (unsigned long)map);
 777}
 778
 779static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 780{
 781	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
 782	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
 783
 784	set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 785}
 786
 787static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
 788{
 789	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
 790
 791	return (get_vma_private_data(vma) & flag) != 0;
 792}
 793
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 794/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
 795void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 796{
 797	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
 798	if (!(vma->vm_flags & VM_MAYSHARE))
 799		vma->vm_private_data = (void *)0;
 800}
 801
 802/* Returns true if the VMA has associated reserve pages */
 803static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
 804{
 805	if (vma->vm_flags & VM_NORESERVE) {
 806		/*
 807		 * This address is already reserved by other process(chg == 0),
 808		 * so, we should decrement reserved count. Without decrementing,
 809		 * reserve count remains after releasing inode, because this
 810		 * allocated page will go into page cache and is regarded as
 811		 * coming from reserved pool in releasing step.  Currently, we
 812		 * don't have any other solution to deal with this situation
 813		 * properly, so add work-around here.
 814		 */
 815		if (vma->vm_flags & VM_MAYSHARE && chg == 0)
 816			return true;
 817		else
 818			return false;
 819	}
 820
 821	/* Shared mappings always use reserves */
 822	if (vma->vm_flags & VM_MAYSHARE) {
 823		/*
 824		 * We know VM_NORESERVE is not set.  Therefore, there SHOULD
 825		 * be a region map for all pages.  The only situation where
 826		 * there is no region map is if a hole was punched via
 827		 * fallocate.  In this case, there really are no reverves to
 828		 * use.  This situation is indicated if chg != 0.
 829		 */
 830		if (chg)
 831			return false;
 832		else
 833			return true;
 
 834	}
 
 835
 836	/*
 837	 * Only the process that called mmap() has reserves for
 838	 * private mappings.
 839	 */
 840	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 841		/*
 842		 * Like the shared case above, a hole punch or truncate
 843		 * could have been performed on the private mapping.
 844		 * Examine the value of chg to determine if reserves
 845		 * actually exist or were previously consumed.
 846		 * Very Subtle - The value of chg comes from a previous
 847		 * call to vma_needs_reserves().  The reserve map for
 848		 * private mappings has different (opposite) semantics
 849		 * than that of shared mappings.  vma_needs_reserves()
 850		 * has already taken this difference in semantics into
 851		 * account.  Therefore, the meaning of chg is the same
 852		 * as in the shared case above.  Code could easily be
 853		 * combined, but keeping it separate draws attention to
 854		 * subtle differences.
 855		 */
 856		if (chg)
 857			return false;
 858		else
 859			return true;
 860	}
 861
 862	return false;
 
 
 
 
 863}
 864
 865static void enqueue_huge_page(struct hstate *h, struct page *page)
 866{
 867	int nid = page_to_nid(page);
 868	list_move(&page->lru, &h->hugepage_freelists[nid]);
 869	h->free_huge_pages++;
 870	h->free_huge_pages_node[nid]++;
 871}
 872
 873static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
 874{
 875	struct page *page;
 876
 877	list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
 878		if (!PageHWPoison(page))
 879			break;
 880	/*
 881	 * if 'non-isolated free hugepage' not found on the list,
 882	 * the allocation fails.
 883	 */
 884	if (&h->hugepage_freelists[nid] == &page->lru)
 885		return NULL;
 886	list_move(&page->lru, &h->hugepage_activelist);
 
 887	set_page_refcounted(page);
 888	h->free_huge_pages--;
 889	h->free_huge_pages_node[nid]--;
 890	return page;
 891}
 892
 893static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
 894		nodemask_t *nmask)
 895{
 896	unsigned int cpuset_mems_cookie;
 897	struct zonelist *zonelist;
 898	struct zone *zone;
 899	struct zoneref *z;
 900	int node = NUMA_NO_NODE;
 901
 902	zonelist = node_zonelist(nid, gfp_mask);
 903
 904retry_cpuset:
 905	cpuset_mems_cookie = read_mems_allowed_begin();
 906	for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
 907		struct page *page;
 908
 909		if (!cpuset_zone_allowed(zone, gfp_mask))
 910			continue;
 911		/*
 912		 * no need to ask again on the same node. Pool is node rather than
 913		 * zone aware
 914		 */
 915		if (zone_to_nid(zone) == node)
 916			continue;
 917		node = zone_to_nid(zone);
 918
 919		page = dequeue_huge_page_node_exact(h, node);
 920		if (page)
 921			return page;
 922	}
 923	if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
 924		goto retry_cpuset;
 925
 926	return NULL;
 927}
 928
 929/* Movability of hugepages depends on migration support. */
 930static inline gfp_t htlb_alloc_mask(struct hstate *h)
 931{
 932	if (hugepage_movable_supported(h))
 933		return GFP_HIGHUSER_MOVABLE;
 934	else
 935		return GFP_HIGHUSER;
 936}
 937
 938static struct page *dequeue_huge_page_vma(struct hstate *h,
 939				struct vm_area_struct *vma,
 940				unsigned long address, int avoid_reserve,
 941				long chg)
 942{
 943	struct page *page;
 944	struct mempolicy *mpol;
 945	gfp_t gfp_mask;
 946	nodemask_t *nodemask;
 947	int nid;
 
 
 
 948
 
 
 
 
 949	/*
 950	 * A child process with MAP_PRIVATE mappings created by their parent
 951	 * have no page reserves. This check ensures that reservations are
 952	 * not "stolen". The child may still get SIGKILLed
 953	 */
 954	if (!vma_has_reserves(vma, chg) &&
 955			h->free_huge_pages - h->resv_huge_pages == 0)
 956		goto err;
 957
 958	/* If reserves cannot be used, ensure enough pages are in the pool */
 959	if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
 960		goto err;
 961
 962	gfp_mask = htlb_alloc_mask(h);
 963	nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
 964	page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
 965	if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
 966		SetPagePrivate(page);
 967		h->resv_huge_pages--;
 
 
 
 
 968	}
 969
 970	mpol_cond_put(mpol);
 
 
 971	return page;
 972
 973err:
 
 974	return NULL;
 975}
 976
 977/*
 978 * common helper functions for hstate_next_node_to_{alloc|free}.
 979 * We may have allocated or freed a huge page based on a different
 980 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
 981 * be outside of *nodes_allowed.  Ensure that we use an allowed
 982 * node for alloc or free.
 983 */
 984static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
 985{
 986	nid = next_node_in(nid, *nodes_allowed);
 987	VM_BUG_ON(nid >= MAX_NUMNODES);
 988
 989	return nid;
 990}
 991
 992static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
 993{
 994	if (!node_isset(nid, *nodes_allowed))
 995		nid = next_node_allowed(nid, nodes_allowed);
 996	return nid;
 997}
 998
 999/*
1000 * returns the previously saved node ["this node"] from which to
1001 * allocate a persistent huge page for the pool and advance the
1002 * next node from which to allocate, handling wrap at end of node
1003 * mask.
1004 */
1005static int hstate_next_node_to_alloc(struct hstate *h,
1006					nodemask_t *nodes_allowed)
1007{
1008	int nid;
1009
1010	VM_BUG_ON(!nodes_allowed);
1011
1012	nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1013	h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1014
1015	return nid;
1016}
1017
1018/*
1019 * helper for free_pool_huge_page() - return the previously saved
1020 * node ["this node"] from which to free a huge page.  Advance the
1021 * next node id whether or not we find a free huge page to free so
1022 * that the next attempt to free addresses the next node.
1023 */
1024static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1025{
1026	int nid;
1027
1028	VM_BUG_ON(!nodes_allowed);
1029
1030	nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1031	h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1032
1033	return nid;
1034}
1035
1036#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)		\
1037	for (nr_nodes = nodes_weight(*mask);				\
1038		nr_nodes > 0 &&						\
1039		((node = hstate_next_node_to_alloc(hs, mask)) || 1);	\
1040		nr_nodes--)
1041
1042#define for_each_node_mask_to_free(hs, nr_nodes, node, mask)		\
1043	for (nr_nodes = nodes_weight(*mask);				\
1044		nr_nodes > 0 &&						\
1045		((node = hstate_next_node_to_free(hs, mask)) || 1);	\
1046		nr_nodes--)
1047
1048#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1049static void destroy_compound_gigantic_page(struct page *page,
1050					unsigned int order)
1051{
1052	int i;
1053	int nr_pages = 1 << order;
1054	struct page *p = page + 1;
1055
1056	atomic_set(compound_mapcount_ptr(page), 0);
1057	for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1058		clear_compound_head(p);
1059		set_page_refcounted(p);
1060	}
1061
1062	set_compound_order(page, 0);
1063	__ClearPageHead(page);
1064}
1065
1066static void free_gigantic_page(struct page *page, unsigned int order)
1067{
1068	free_contig_range(page_to_pfn(page), 1 << order);
1069}
1070
1071#ifdef CONFIG_CONTIG_ALLOC
1072static int __alloc_gigantic_page(unsigned long start_pfn,
1073				unsigned long nr_pages, gfp_t gfp_mask)
1074{
1075	unsigned long end_pfn = start_pfn + nr_pages;
1076	return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1077				  gfp_mask);
1078}
1079
1080static bool pfn_range_valid_gigantic(struct zone *z,
1081			unsigned long start_pfn, unsigned long nr_pages)
1082{
1083	unsigned long i, end_pfn = start_pfn + nr_pages;
1084	struct page *page;
1085
1086	for (i = start_pfn; i < end_pfn; i++) {
1087		page = pfn_to_online_page(i);
1088		if (!page)
1089			return false;
1090
1091		if (page_zone(page) != z)
1092			return false;
1093
1094		if (PageReserved(page))
1095			return false;
1096
1097		if (page_count(page) > 0)
1098			return false;
1099
1100		if (PageHuge(page))
1101			return false;
1102	}
1103
1104	return true;
1105}
1106
1107static bool zone_spans_last_pfn(const struct zone *zone,
1108			unsigned long start_pfn, unsigned long nr_pages)
1109{
1110	unsigned long last_pfn = start_pfn + nr_pages - 1;
1111	return zone_spans_pfn(zone, last_pfn);
1112}
1113
1114static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1115		int nid, nodemask_t *nodemask)
1116{
1117	unsigned int order = huge_page_order(h);
1118	unsigned long nr_pages = 1 << order;
1119	unsigned long ret, pfn, flags;
1120	struct zonelist *zonelist;
1121	struct zone *zone;
1122	struct zoneref *z;
1123
1124	zonelist = node_zonelist(nid, gfp_mask);
1125	for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1126		spin_lock_irqsave(&zone->lock, flags);
1127
1128		pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1129		while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1130			if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1131				/*
1132				 * We release the zone lock here because
1133				 * alloc_contig_range() will also lock the zone
1134				 * at some point. If there's an allocation
1135				 * spinning on this lock, it may win the race
1136				 * and cause alloc_contig_range() to fail...
1137				 */
1138				spin_unlock_irqrestore(&zone->lock, flags);
1139				ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1140				if (!ret)
1141					return pfn_to_page(pfn);
1142				spin_lock_irqsave(&zone->lock, flags);
1143			}
1144			pfn += nr_pages;
1145		}
1146
1147		spin_unlock_irqrestore(&zone->lock, flags);
1148	}
1149
1150	return NULL;
1151}
1152
1153static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1154static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1155#else /* !CONFIG_CONTIG_ALLOC */
1156static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1157					int nid, nodemask_t *nodemask)
1158{
1159	return NULL;
1160}
1161#endif /* CONFIG_CONTIG_ALLOC */
1162
1163#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1164static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1165					int nid, nodemask_t *nodemask)
1166{
1167	return NULL;
1168}
1169static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1170static inline void destroy_compound_gigantic_page(struct page *page,
1171						unsigned int order) { }
1172#endif
1173
1174static void update_and_free_page(struct hstate *h, struct page *page)
1175{
1176	int i;
1177
1178	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1179		return;
1180
1181	h->nr_huge_pages--;
1182	h->nr_huge_pages_node[page_to_nid(page)]--;
1183	for (i = 0; i < pages_per_huge_page(h); i++) {
1184		page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1185				1 << PG_referenced | 1 << PG_dirty |
1186				1 << PG_active | 1 << PG_private |
1187				1 << PG_writeback);
1188	}
1189	VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1190	set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1191	set_page_refcounted(page);
1192	if (hstate_is_gigantic(h)) {
1193		destroy_compound_gigantic_page(page, huge_page_order(h));
1194		free_gigantic_page(page, huge_page_order(h));
1195	} else {
1196		__free_pages(page, huge_page_order(h));
1197	}
1198}
1199
1200struct hstate *size_to_hstate(unsigned long size)
1201{
1202	struct hstate *h;
1203
1204	for_each_hstate(h) {
1205		if (huge_page_size(h) == size)
1206			return h;
1207	}
1208	return NULL;
1209}
1210
1211/*
1212 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1213 * to hstate->hugepage_activelist.)
1214 *
1215 * This function can be called for tail pages, but never returns true for them.
1216 */
1217bool page_huge_active(struct page *page)
1218{
1219	VM_BUG_ON_PAGE(!PageHuge(page), page);
1220	return PageHead(page) && PagePrivate(&page[1]);
1221}
1222
1223/* never called for tail page */
1224static void set_page_huge_active(struct page *page)
1225{
1226	VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1227	SetPagePrivate(&page[1]);
1228}
1229
1230static void clear_page_huge_active(struct page *page)
1231{
1232	VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1233	ClearPagePrivate(&page[1]);
1234}
1235
1236/*
1237 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1238 * code
1239 */
1240static inline bool PageHugeTemporary(struct page *page)
1241{
1242	if (!PageHuge(page))
1243		return false;
1244
1245	return (unsigned long)page[2].mapping == -1U;
1246}
1247
1248static inline void SetPageHugeTemporary(struct page *page)
1249{
1250	page[2].mapping = (void *)-1U;
1251}
1252
1253static inline void ClearPageHugeTemporary(struct page *page)
1254{
1255	page[2].mapping = NULL;
1256}
1257
1258void free_huge_page(struct page *page)
1259{
1260	/*
1261	 * Can't pass hstate in here because it is called from the
1262	 * compound page destructor.
1263	 */
1264	struct hstate *h = page_hstate(page);
1265	int nid = page_to_nid(page);
1266	struct hugepage_subpool *spool =
1267		(struct hugepage_subpool *)page_private(page);
1268	bool restore_reserve;
1269
1270	VM_BUG_ON_PAGE(page_count(page), page);
1271	VM_BUG_ON_PAGE(page_mapcount(page), page);
1272
1273	set_page_private(page, 0);
1274	page->mapping = NULL;
1275	restore_reserve = PagePrivate(page);
1276	ClearPagePrivate(page);
1277
1278	/*
1279	 * If PagePrivate() was set on page, page allocation consumed a
1280	 * reservation.  If the page was associated with a subpool, there
1281	 * would have been a page reserved in the subpool before allocation
1282	 * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1283	 * reservtion, do not call hugepage_subpool_put_pages() as this will
1284	 * remove the reserved page from the subpool.
1285	 */
1286	if (!restore_reserve) {
1287		/*
1288		 * A return code of zero implies that the subpool will be
1289		 * under its minimum size if the reservation is not restored
1290		 * after page is free.  Therefore, force restore_reserve
1291		 * operation.
1292		 */
1293		if (hugepage_subpool_put_pages(spool, 1) == 0)
1294			restore_reserve = true;
1295	}
1296
1297	spin_lock(&hugetlb_lock);
1298	clear_page_huge_active(page);
1299	hugetlb_cgroup_uncharge_page(hstate_index(h),
1300				     pages_per_huge_page(h), page);
1301	if (restore_reserve)
1302		h->resv_huge_pages++;
1303
1304	if (PageHugeTemporary(page)) {
1305		list_del(&page->lru);
1306		ClearPageHugeTemporary(page);
1307		update_and_free_page(h, page);
1308	} else if (h->surplus_huge_pages_node[nid]) {
1309		/* remove the page from active list */
1310		list_del(&page->lru);
1311		update_and_free_page(h, page);
1312		h->surplus_huge_pages--;
1313		h->surplus_huge_pages_node[nid]--;
1314	} else {
1315		arch_clear_hugepage_flags(page);
1316		enqueue_huge_page(h, page);
1317	}
1318	spin_unlock(&hugetlb_lock);
 
1319}
1320
1321static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1322{
1323	INIT_LIST_HEAD(&page->lru);
1324	set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1325	spin_lock(&hugetlb_lock);
1326	set_hugetlb_cgroup(page, NULL);
1327	h->nr_huge_pages++;
1328	h->nr_huge_pages_node[nid]++;
1329	spin_unlock(&hugetlb_lock);
 
1330}
1331
1332static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1333{
1334	int i;
1335	int nr_pages = 1 << order;
1336	struct page *p = page + 1;
1337
1338	/* we rely on prep_new_huge_page to set the destructor */
1339	set_compound_order(page, order);
1340	__ClearPageReserved(page);
1341	__SetPageHead(page);
1342	for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1343		/*
1344		 * For gigantic hugepages allocated through bootmem at
1345		 * boot, it's safer to be consistent with the not-gigantic
1346		 * hugepages and clear the PG_reserved bit from all tail pages
1347		 * too.  Otherwse drivers using get_user_pages() to access tail
1348		 * pages may get the reference counting wrong if they see
1349		 * PG_reserved set on a tail page (despite the head page not
1350		 * having PG_reserved set).  Enforcing this consistency between
1351		 * head and tail pages allows drivers to optimize away a check
1352		 * on the head page when they need know if put_page() is needed
1353		 * after get_user_pages().
1354		 */
1355		__ClearPageReserved(p);
1356		set_page_count(p, 0);
1357		set_compound_head(p, page);
1358	}
1359	atomic_set(compound_mapcount_ptr(page), -1);
1360}
1361
1362/*
1363 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1364 * transparent huge pages.  See the PageTransHuge() documentation for more
1365 * details.
1366 */
1367int PageHuge(struct page *page)
1368{
 
 
1369	if (!PageCompound(page))
1370		return 0;
1371
1372	page = compound_head(page);
1373	return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
 
 
1374}
1375EXPORT_SYMBOL_GPL(PageHuge);
1376
1377/*
1378 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1379 * normal or transparent huge pages.
1380 */
1381int PageHeadHuge(struct page *page_head)
1382{
1383	if (!PageHead(page_head))
1384		return 0;
1385
1386	return get_compound_page_dtor(page_head) == free_huge_page;
1387}
1388
1389pgoff_t __basepage_index(struct page *page)
1390{
1391	struct page *page_head = compound_head(page);
1392	pgoff_t index = page_index(page_head);
1393	unsigned long compound_idx;
 
 
 
 
 
 
1394
1395	if (!PageHuge(page_head))
1396		return page_index(page);
1397
1398	if (compound_order(page_head) >= MAX_ORDER)
1399		compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1400	else
1401		compound_idx = page - page_head;
 
 
 
 
 
 
 
 
 
1402
1403	return (index << compound_order(page_head)) + compound_idx;
1404}
1405
1406static struct page *alloc_buddy_huge_page(struct hstate *h,
1407		gfp_t gfp_mask, int nid, nodemask_t *nmask,
1408		nodemask_t *node_alloc_noretry)
1409{
1410	int order = huge_page_order(h);
1411	struct page *page;
1412	bool alloc_try_hard = true;
 
1413
1414	/*
1415	 * By default we always try hard to allocate the page with
1416	 * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
1417	 * a loop (to adjust global huge page counts) and previous allocation
1418	 * failed, do not continue to try hard on the same node.  Use the
1419	 * node_alloc_noretry bitmap to manage this state information.
1420	 */
1421	if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1422		alloc_try_hard = false;
1423	gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1424	if (alloc_try_hard)
1425		gfp_mask |= __GFP_RETRY_MAYFAIL;
1426	if (nid == NUMA_NO_NODE)
1427		nid = numa_mem_id();
1428	page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1429	if (page)
1430		__count_vm_event(HTLB_BUDDY_PGALLOC);
1431	else
1432		__count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1433
1434	/*
1435	 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1436	 * indicates an overall state change.  Clear bit so that we resume
1437	 * normal 'try hard' allocations.
1438	 */
1439	if (node_alloc_noretry && page && !alloc_try_hard)
1440		node_clear(nid, *node_alloc_noretry);
1441
1442	/*
1443	 * If we tried hard to get a page but failed, set bit so that
1444	 * subsequent attempts will not try as hard until there is an
1445	 * overall state change.
1446	 */
1447	if (node_alloc_noretry && !page && alloc_try_hard)
1448		node_set(nid, *node_alloc_noretry);
1449
1450	return page;
1451}
1452
1453/*
1454 * Common helper to allocate a fresh hugetlb page. All specific allocators
1455 * should use this function to get new hugetlb pages
1456 */
1457static struct page *alloc_fresh_huge_page(struct hstate *h,
1458		gfp_t gfp_mask, int nid, nodemask_t *nmask,
1459		nodemask_t *node_alloc_noretry)
1460{
1461	struct page *page;
 
 
 
1462
1463	if (hstate_is_gigantic(h))
1464		page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1465	else
1466		page = alloc_buddy_huge_page(h, gfp_mask,
1467				nid, nmask, node_alloc_noretry);
1468	if (!page)
1469		return NULL;
 
 
 
 
1470
1471	if (hstate_is_gigantic(h))
1472		prep_compound_gigantic_page(page, huge_page_order(h));
1473	prep_new_huge_page(h, page, page_to_nid(page));
 
1474
1475	return page;
1476}
1477
1478/*
1479 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1480 * manner.
 
 
1481 */
1482static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1483				nodemask_t *node_alloc_noretry)
1484{
1485	struct page *page;
1486	int nr_nodes, node;
1487	gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1488
1489	for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1490		page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1491						node_alloc_noretry);
1492		if (page)
1493			break;
1494	}
1495
1496	if (!page)
1497		return 0;
1498
1499	put_page(page); /* free it into the hugepage allocator */
 
1500
1501	return 1;
1502}
1503
1504/*
1505 * Free huge page from pool from next node to free.
1506 * Attempt to keep persistent huge pages more or less
1507 * balanced over allowed nodes.
1508 * Called with hugetlb_lock locked.
1509 */
1510static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1511							 bool acct_surplus)
1512{
1513	int nr_nodes, node;
 
1514	int ret = 0;
1515
1516	for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
 
 
 
1517		/*
1518		 * If we're returning unused surplus pages, only examine
1519		 * nodes with surplus pages.
1520		 */
1521		if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1522		    !list_empty(&h->hugepage_freelists[node])) {
1523			struct page *page =
1524				list_entry(h->hugepage_freelists[node].next,
1525					  struct page, lru);
1526			list_del(&page->lru);
1527			h->free_huge_pages--;
1528			h->free_huge_pages_node[node]--;
1529			if (acct_surplus) {
1530				h->surplus_huge_pages--;
1531				h->surplus_huge_pages_node[node]--;
1532			}
1533			update_and_free_page(h, page);
1534			ret = 1;
1535			break;
1536		}
1537	}
 
1538
1539	return ret;
1540}
1541
1542/*
1543 * Dissolve a given free hugepage into free buddy pages. This function does
1544 * nothing for in-use hugepages and non-hugepages.
1545 * This function returns values like below:
1546 *
1547 *  -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1548 *          (allocated or reserved.)
1549 *       0: successfully dissolved free hugepages or the page is not a
1550 *          hugepage (considered as already dissolved)
1551 */
1552int dissolve_free_huge_page(struct page *page)
1553{
1554	int rc = -EBUSY;
1555
1556	/* Not to disrupt normal path by vainly holding hugetlb_lock */
1557	if (!PageHuge(page))
1558		return 0;
1559
1560	spin_lock(&hugetlb_lock);
1561	if (!PageHuge(page)) {
1562		rc = 0;
1563		goto out;
1564	}
1565
1566	if (!page_count(page)) {
1567		struct page *head = compound_head(page);
1568		struct hstate *h = page_hstate(head);
1569		int nid = page_to_nid(head);
1570		if (h->free_huge_pages - h->resv_huge_pages == 0)
1571			goto out;
1572		/*
1573		 * Move PageHWPoison flag from head page to the raw error page,
1574		 * which makes any subpages rather than the error page reusable.
1575		 */
1576		if (PageHWPoison(head) && page != head) {
1577			SetPageHWPoison(page);
1578			ClearPageHWPoison(head);
1579		}
1580		list_del(&head->lru);
1581		h->free_huge_pages--;
1582		h->free_huge_pages_node[nid]--;
1583		h->max_huge_pages--;
1584		update_and_free_page(h, head);
1585		rc = 0;
1586	}
1587out:
1588	spin_unlock(&hugetlb_lock);
1589	return rc;
1590}
1591
1592/*
1593 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1594 * make specified memory blocks removable from the system.
1595 * Note that this will dissolve a free gigantic hugepage completely, if any
1596 * part of it lies within the given range.
1597 * Also note that if dissolve_free_huge_page() returns with an error, all
1598 * free hugepages that were dissolved before that error are lost.
1599 */
1600int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1601{
1602	unsigned long pfn;
1603	struct page *page;
1604	int rc = 0;
1605
1606	if (!hugepages_supported())
1607		return rc;
1608
1609	for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1610		page = pfn_to_page(pfn);
1611		rc = dissolve_free_huge_page(page);
1612		if (rc)
1613			break;
1614	}
1615
1616	return rc;
1617}
1618
1619/*
1620 * Allocates a fresh surplus page from the page allocator.
1621 */
1622static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1623		int nid, nodemask_t *nmask)
1624{
1625	struct page *page = NULL;
1626
1627	if (hstate_is_gigantic(h))
1628		return NULL;
1629
1630	spin_lock(&hugetlb_lock);
1631	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1632		goto out_unlock;
1633	spin_unlock(&hugetlb_lock);
1634
1635	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1636	if (!page)
1637		return NULL;
1638
1639	spin_lock(&hugetlb_lock);
1640	/*
1641	 * We could have raced with the pool size change.
1642	 * Double check that and simply deallocate the new page
1643	 * if we would end up overcommiting the surpluses. Abuse
1644	 * temporary page to workaround the nasty free_huge_page
1645	 * codeflow
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1646	 */
 
1647	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1648		SetPageHugeTemporary(page);
1649		spin_unlock(&hugetlb_lock);
1650		put_page(page);
1651		return NULL;
1652	} else {
 
1653		h->surplus_huge_pages++;
1654		h->surplus_huge_pages_node[page_to_nid(page)]++;
1655	}
1656
1657out_unlock:
1658	spin_unlock(&hugetlb_lock);
1659
1660	return page;
1661}
1662
1663struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1664				     int nid, nodemask_t *nmask)
1665{
1666	struct page *page;
1667
1668	if (hstate_is_gigantic(h))
1669		return NULL;
1670
1671	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1672	if (!page)
1673		return NULL;
 
1674
1675	/*
1676	 * We do not account these pages as surplus because they are only
1677	 * temporary and will be released properly on the last reference
1678	 */
1679	SetPageHugeTemporary(page);
 
 
 
 
 
 
 
 
 
 
 
1680
1681	return page;
1682}
1683
1684/*
1685 * Use the VMA's mpolicy to allocate a huge page from the buddy.
 
 
1686 */
1687static
1688struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1689		struct vm_area_struct *vma, unsigned long addr)
1690{
1691	struct page *page;
1692	struct mempolicy *mpol;
1693	gfp_t gfp_mask = htlb_alloc_mask(h);
1694	int nid;
1695	nodemask_t *nodemask;
1696
1697	nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1698	page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1699	mpol_cond_put(mpol);
1700
1701	return page;
1702}
1703
1704/* page migration callback function */
1705struct page *alloc_huge_page_node(struct hstate *h, int nid)
1706{
1707	gfp_t gfp_mask = htlb_alloc_mask(h);
1708	struct page *page = NULL;
1709
1710	if (nid != NUMA_NO_NODE)
1711		gfp_mask |= __GFP_THISNODE;
1712
1713	spin_lock(&hugetlb_lock);
1714	if (h->free_huge_pages - h->resv_huge_pages > 0)
1715		page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1716	spin_unlock(&hugetlb_lock);
1717
1718	if (!page)
1719		page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1720
1721	return page;
1722}
1723
1724/* page migration callback function */
1725struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1726		nodemask_t *nmask)
1727{
1728	gfp_t gfp_mask = htlb_alloc_mask(h);
1729
1730	spin_lock(&hugetlb_lock);
1731	if (h->free_huge_pages - h->resv_huge_pages > 0) {
1732		struct page *page;
1733
1734		page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1735		if (page) {
1736			spin_unlock(&hugetlb_lock);
1737			return page;
1738		}
1739	}
1740	spin_unlock(&hugetlb_lock);
1741
1742	return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1743}
1744
1745/* mempolicy aware migration callback */
1746struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1747		unsigned long address)
1748{
1749	struct mempolicy *mpol;
1750	nodemask_t *nodemask;
1751	struct page *page;
1752	gfp_t gfp_mask;
1753	int node;
1754
1755	gfp_mask = htlb_alloc_mask(h);
1756	node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1757	page = alloc_huge_page_nodemask(h, node, nodemask);
1758	mpol_cond_put(mpol);
1759
1760	return page;
1761}
1762
1763/*
1764 * Increase the hugetlb pool such that it can accommodate a reservation
1765 * of size 'delta'.
1766 */
1767static int gather_surplus_pages(struct hstate *h, int delta)
1768{
1769	struct list_head surplus_list;
1770	struct page *page, *tmp;
1771	int ret, i;
1772	int needed, allocated;
1773	bool alloc_ok = true;
1774
1775	needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1776	if (needed <= 0) {
1777		h->resv_huge_pages += delta;
1778		return 0;
1779	}
1780
1781	allocated = 0;
1782	INIT_LIST_HEAD(&surplus_list);
1783
1784	ret = -ENOMEM;
1785retry:
1786	spin_unlock(&hugetlb_lock);
1787	for (i = 0; i < needed; i++) {
1788		page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1789				NUMA_NO_NODE, NULL);
1790		if (!page) {
1791			alloc_ok = false;
1792			break;
1793		}
1794		list_add(&page->lru, &surplus_list);
1795		cond_resched();
1796	}
1797	allocated += i;
1798
1799	/*
1800	 * After retaking hugetlb_lock, we need to recalculate 'needed'
1801	 * because either resv_huge_pages or free_huge_pages may have changed.
1802	 */
1803	spin_lock(&hugetlb_lock);
1804	needed = (h->resv_huge_pages + delta) -
1805			(h->free_huge_pages + allocated);
1806	if (needed > 0) {
1807		if (alloc_ok)
1808			goto retry;
1809		/*
1810		 * We were not able to allocate enough pages to
1811		 * satisfy the entire reservation so we free what
1812		 * we've allocated so far.
1813		 */
1814		goto free;
1815	}
1816	/*
1817	 * The surplus_list now contains _at_least_ the number of extra pages
1818	 * needed to accommodate the reservation.  Add the appropriate number
1819	 * of pages to the hugetlb pool and free the extras back to the buddy
1820	 * allocator.  Commit the entire reservation here to prevent another
1821	 * process from stealing the pages as they are added to the pool but
1822	 * before they are reserved.
1823	 */
1824	needed += allocated;
1825	h->resv_huge_pages += delta;
1826	ret = 0;
1827
1828	/* Free the needed pages to the hugetlb pool */
1829	list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1830		if ((--needed) < 0)
1831			break;
 
1832		/*
1833		 * This page is now managed by the hugetlb allocator and has
1834		 * no users -- drop the buddy allocator's reference.
1835		 */
1836		put_page_testzero(page);
1837		VM_BUG_ON_PAGE(page_count(page), page);
1838		enqueue_huge_page(h, page);
1839	}
1840free:
1841	spin_unlock(&hugetlb_lock);
1842
1843	/* Free unnecessary surplus pages to the buddy allocator */
1844	list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1845		put_page(page);
 
 
 
 
1846	spin_lock(&hugetlb_lock);
1847
1848	return ret;
1849}
1850
1851/*
1852 * This routine has two main purposes:
1853 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1854 *    in unused_resv_pages.  This corresponds to the prior adjustments made
1855 *    to the associated reservation map.
1856 * 2) Free any unused surplus pages that may have been allocated to satisfy
1857 *    the reservation.  As many as unused_resv_pages may be freed.
1858 *
1859 * Called with hugetlb_lock held.  However, the lock could be dropped (and
1860 * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1861 * we must make sure nobody else can claim pages we are in the process of
1862 * freeing.  Do this by ensuring resv_huge_page always is greater than the
1863 * number of huge pages we plan to free when dropping the lock.
1864 */
1865static void return_unused_surplus_pages(struct hstate *h,
1866					unsigned long unused_resv_pages)
1867{
1868	unsigned long nr_pages;
1869
 
 
 
1870	/* Cannot return gigantic pages currently */
1871	if (hstate_is_gigantic(h))
1872		goto out;
1873
1874	/*
1875	 * Part (or even all) of the reservation could have been backed
1876	 * by pre-allocated pages. Only free surplus pages.
1877	 */
1878	nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1879
1880	/*
1881	 * We want to release as many surplus pages as possible, spread
1882	 * evenly across all nodes with memory. Iterate across these nodes
1883	 * until we can no longer free unreserved surplus pages. This occurs
1884	 * when the nodes with surplus pages have no free pages.
1885	 * free_pool_huge_page() will balance the the freed pages across the
1886	 * on-line nodes with memory and will handle the hstate accounting.
1887	 *
1888	 * Note that we decrement resv_huge_pages as we free the pages.  If
1889	 * we drop the lock, resv_huge_pages will still be sufficiently large
1890	 * to cover subsequent pages we may free.
1891	 */
1892	while (nr_pages--) {
1893		h->resv_huge_pages--;
1894		unused_resv_pages--;
1895		if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1896			goto out;
1897		cond_resched_lock(&hugetlb_lock);
1898	}
1899
1900out:
1901	/* Fully uncommit the reservation */
1902	h->resv_huge_pages -= unused_resv_pages;
1903}
1904
1905
1906/*
1907 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1908 * are used by the huge page allocation routines to manage reservations.
1909 *
1910 * vma_needs_reservation is called to determine if the huge page at addr
1911 * within the vma has an associated reservation.  If a reservation is
1912 * needed, the value 1 is returned.  The caller is then responsible for
1913 * managing the global reservation and subpool usage counts.  After
1914 * the huge page has been allocated, vma_commit_reservation is called
1915 * to add the page to the reservation map.  If the page allocation fails,
1916 * the reservation must be ended instead of committed.  vma_end_reservation
1917 * is called in such cases.
1918 *
1919 * In the normal case, vma_commit_reservation returns the same value
1920 * as the preceding vma_needs_reservation call.  The only time this
1921 * is not the case is if a reserve map was changed between calls.  It
1922 * is the responsibility of the caller to notice the difference and
1923 * take appropriate action.
1924 *
1925 * vma_add_reservation is used in error paths where a reservation must
1926 * be restored when a newly allocated huge page must be freed.  It is
1927 * to be called after calling vma_needs_reservation to determine if a
1928 * reservation exists.
1929 */
1930enum vma_resv_mode {
1931	VMA_NEEDS_RESV,
1932	VMA_COMMIT_RESV,
1933	VMA_END_RESV,
1934	VMA_ADD_RESV,
1935};
1936static long __vma_reservation_common(struct hstate *h,
1937				struct vm_area_struct *vma, unsigned long addr,
1938				enum vma_resv_mode mode)
1939{
1940	struct resv_map *resv;
1941	pgoff_t idx;
1942	long ret;
1943
1944	resv = vma_resv_map(vma);
1945	if (!resv)
1946		return 1;
 
1947
1948	idx = vma_hugecache_offset(h, vma, addr);
1949	switch (mode) {
1950	case VMA_NEEDS_RESV:
1951		ret = region_chg(resv, idx, idx + 1);
1952		break;
1953	case VMA_COMMIT_RESV:
1954		ret = region_add(resv, idx, idx + 1);
1955		break;
1956	case VMA_END_RESV:
1957		region_abort(resv, idx, idx + 1);
1958		ret = 0;
1959		break;
1960	case VMA_ADD_RESV:
1961		if (vma->vm_flags & VM_MAYSHARE)
1962			ret = region_add(resv, idx, idx + 1);
1963		else {
1964			region_abort(resv, idx, idx + 1);
1965			ret = region_del(resv, idx, idx + 1);
1966		}
1967		break;
1968	default:
1969		BUG();
1970	}
1971
1972	if (vma->vm_flags & VM_MAYSHARE)
1973		return ret;
1974	else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1975		/*
1976		 * In most cases, reserves always exist for private mappings.
1977		 * However, a file associated with mapping could have been
1978		 * hole punched or truncated after reserves were consumed.
1979		 * As subsequent fault on such a range will not use reserves.
1980		 * Subtle - The reserve map for private mappings has the
1981		 * opposite meaning than that of shared mappings.  If NO
1982		 * entry is in the reserve map, it means a reservation exists.
1983		 * If an entry exists in the reserve map, it means the
1984		 * reservation has already been consumed.  As a result, the
1985		 * return value of this routine is the opposite of the
1986		 * value returned from reserve map manipulation routines above.
1987		 */
1988		if (ret)
1989			return 0;
1990		else
1991			return 1;
1992	}
1993	else
1994		return ret < 0 ? ret : 0;
1995}
1996
1997static long vma_needs_reservation(struct hstate *h,
1998			struct vm_area_struct *vma, unsigned long addr)
1999{
2000	return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2001}
2002
2003static long vma_commit_reservation(struct hstate *h,
2004			struct vm_area_struct *vma, unsigned long addr)
2005{
2006	return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2007}
2008
2009static void vma_end_reservation(struct hstate *h,
2010			struct vm_area_struct *vma, unsigned long addr)
2011{
2012	(void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2013}
2014
2015static long vma_add_reservation(struct hstate *h,
2016			struct vm_area_struct *vma, unsigned long addr)
2017{
2018	return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2019}
2020
2021/*
2022 * This routine is called to restore a reservation on error paths.  In the
2023 * specific error paths, a huge page was allocated (via alloc_huge_page)
2024 * and is about to be freed.  If a reservation for the page existed,
2025 * alloc_huge_page would have consumed the reservation and set PagePrivate
2026 * in the newly allocated page.  When the page is freed via free_huge_page,
2027 * the global reservation count will be incremented if PagePrivate is set.
2028 * However, free_huge_page can not adjust the reserve map.  Adjust the
2029 * reserve map here to be consistent with global reserve count adjustments
2030 * to be made by free_huge_page.
2031 */
2032static void restore_reserve_on_error(struct hstate *h,
2033			struct vm_area_struct *vma, unsigned long address,
2034			struct page *page)
2035{
2036	if (unlikely(PagePrivate(page))) {
2037		long rc = vma_needs_reservation(h, vma, address);
2038
2039		if (unlikely(rc < 0)) {
2040			/*
2041			 * Rare out of memory condition in reserve map
2042			 * manipulation.  Clear PagePrivate so that
2043			 * global reserve count will not be incremented
2044			 * by free_huge_page.  This will make it appear
2045			 * as though the reservation for this page was
2046			 * consumed.  This may prevent the task from
2047			 * faulting in the page at a later time.  This
2048			 * is better than inconsistent global huge page
2049			 * accounting of reserve counts.
2050			 */
2051			ClearPagePrivate(page);
2052		} else if (rc) {
2053			rc = vma_add_reservation(h, vma, address);
2054			if (unlikely(rc < 0))
2055				/*
2056				 * See above comment about rare out of
2057				 * memory condition.
2058				 */
2059				ClearPagePrivate(page);
2060		} else
2061			vma_end_reservation(h, vma, address);
2062	}
2063}
2064
2065struct page *alloc_huge_page(struct vm_area_struct *vma,
2066				    unsigned long addr, int avoid_reserve)
2067{
2068	struct hugepage_subpool *spool = subpool_vma(vma);
2069	struct hstate *h = hstate_vma(vma);
2070	struct page *page;
2071	long map_chg, map_commit;
2072	long gbl_chg;
2073	int ret, idx;
2074	struct hugetlb_cgroup *h_cg;
2075
2076	idx = hstate_index(h);
2077	/*
2078	 * Examine the region/reserve map to determine if the process
2079	 * has a reservation for the page to be allocated.  A return
2080	 * code of zero indicates a reservation exists (no change).
2081	 */
2082	map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2083	if (map_chg < 0)
2084		return ERR_PTR(-ENOMEM);
2085
2086	/*
2087	 * Processes that did not create the mapping will have no
2088	 * reserves as indicated by the region/reserve map. Check
2089	 * that the allocation will not exceed the subpool limit.
2090	 * Allocations for MAP_NORESERVE mappings also need to be
2091	 * checked against any subpool limit.
2092	 */
2093	if (map_chg || avoid_reserve) {
2094		gbl_chg = hugepage_subpool_get_pages(spool, 1);
2095		if (gbl_chg < 0) {
2096			vma_end_reservation(h, vma, addr);
2097			return ERR_PTR(-ENOSPC);
2098		}
2099
2100		/*
2101		 * Even though there was no reservation in the region/reserve
2102		 * map, there could be reservations associated with the
2103		 * subpool that can be used.  This would be indicated if the
2104		 * return value of hugepage_subpool_get_pages() is zero.
2105		 * However, if avoid_reserve is specified we still avoid even
2106		 * the subpool reservations.
2107		 */
2108		if (avoid_reserve)
2109			gbl_chg = 1;
2110	}
2111
2112	ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2113	if (ret)
2114		goto out_subpool_put;
2115
2116	spin_lock(&hugetlb_lock);
2117	/*
2118	 * glb_chg is passed to indicate whether or not a page must be taken
2119	 * from the global free pool (global change).  gbl_chg == 0 indicates
2120	 * a reservation exists for the allocation.
2121	 */
2122	page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2123	if (!page) {
2124		spin_unlock(&hugetlb_lock);
2125		page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2126		if (!page)
2127			goto out_uncharge_cgroup;
2128		if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2129			SetPagePrivate(page);
2130			h->resv_huge_pages--;
2131		}
2132		spin_lock(&hugetlb_lock);
2133		list_move(&page->lru, &h->hugepage_activelist);
2134		/* Fall through */
2135	}
2136	hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2137	spin_unlock(&hugetlb_lock);
2138
2139	set_page_private(page, (unsigned long)spool);
2140
2141	map_commit = vma_commit_reservation(h, vma, addr);
2142	if (unlikely(map_chg > map_commit)) {
2143		/*
2144		 * The page was added to the reservation map between
2145		 * vma_needs_reservation and vma_commit_reservation.
2146		 * This indicates a race with hugetlb_reserve_pages.
2147		 * Adjust for the subpool count incremented above AND
2148		 * in hugetlb_reserve_pages for the same page.  Also,
2149		 * the reservation count added in hugetlb_reserve_pages
2150		 * no longer applies.
2151		 */
2152		long rsv_adjust;
2153
2154		rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2155		hugetlb_acct_memory(h, -rsv_adjust);
2156	}
2157	return page;
2158
2159out_uncharge_cgroup:
2160	hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2161out_subpool_put:
2162	if (map_chg || avoid_reserve)
2163		hugepage_subpool_put_pages(spool, 1);
2164	vma_end_reservation(h, vma, addr);
2165	return ERR_PTR(-ENOSPC);
2166}
2167
2168int alloc_bootmem_huge_page(struct hstate *h)
2169	__attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2170int __alloc_bootmem_huge_page(struct hstate *h)
2171{
2172	struct huge_bootmem_page *m;
2173	int nr_nodes, node;
2174
2175	for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2176		void *addr;
2177
2178		addr = memblock_alloc_try_nid_raw(
2179				huge_page_size(h), huge_page_size(h),
2180				0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
 
 
2181		if (addr) {
2182			/*
2183			 * Use the beginning of the huge page to store the
2184			 * huge_bootmem_page struct (until gather_bootmem
2185			 * puts them into the mem_map).
2186			 */
2187			m = addr;
2188			goto found;
2189		}
 
2190	}
2191	return 0;
2192
2193found:
2194	BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2195	/* Put them into a private list first because mem_map is not up yet */
2196	INIT_LIST_HEAD(&m->list);
2197	list_add(&m->list, &huge_boot_pages);
2198	m->hstate = h;
2199	return 1;
2200}
2201
2202static void __init prep_compound_huge_page(struct page *page,
2203		unsigned int order)
2204{
2205	if (unlikely(order > (MAX_ORDER - 1)))
2206		prep_compound_gigantic_page(page, order);
2207	else
2208		prep_compound_page(page, order);
2209}
2210
2211/* Put bootmem huge pages into the standard lists after mem_map is up */
2212static void __init gather_bootmem_prealloc(void)
2213{
2214	struct huge_bootmem_page *m;
2215
2216	list_for_each_entry(m, &huge_boot_pages, list) {
2217		struct page *page = virt_to_page(m);
2218		struct hstate *h = m->hstate;
 
2219
 
 
 
 
 
 
 
 
2220		WARN_ON(page_count(page) != 1);
2221		prep_compound_huge_page(page, h->order);
2222		WARN_ON(PageReserved(page));
2223		prep_new_huge_page(h, page, page_to_nid(page));
2224		put_page(page); /* free it into the hugepage allocator */
2225
2226		/*
2227		 * If we had gigantic hugepages allocated at boot time, we need
2228		 * to restore the 'stolen' pages to totalram_pages in order to
2229		 * fix confusing memory reports from free(1) and another
2230		 * side-effects, like CommitLimit going negative.
2231		 */
2232		if (hstate_is_gigantic(h))
2233			adjust_managed_page_count(page, 1 << h->order);
2234		cond_resched();
2235	}
2236}
2237
2238static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2239{
2240	unsigned long i;
2241	nodemask_t *node_alloc_noretry;
2242
2243	if (!hstate_is_gigantic(h)) {
2244		/*
2245		 * Bit mask controlling how hard we retry per-node allocations.
2246		 * Ignore errors as lower level routines can deal with
2247		 * node_alloc_noretry == NULL.  If this kmalloc fails at boot
2248		 * time, we are likely in bigger trouble.
2249		 */
2250		node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2251						GFP_KERNEL);
2252	} else {
2253		/* allocations done at boot time */
2254		node_alloc_noretry = NULL;
2255	}
2256
2257	/* bit mask controlling how hard we retry per-node allocations */
2258	if (node_alloc_noretry)
2259		nodes_clear(*node_alloc_noretry);
2260
2261	for (i = 0; i < h->max_huge_pages; ++i) {
2262		if (hstate_is_gigantic(h)) {
2263			if (!alloc_bootmem_huge_page(h))
2264				break;
2265		} else if (!alloc_pool_huge_page(h,
2266					 &node_states[N_MEMORY],
2267					 node_alloc_noretry))
2268			break;
2269		cond_resched();
2270	}
2271	if (i < h->max_huge_pages) {
2272		char buf[32];
2273
2274		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2275		pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2276			h->max_huge_pages, buf, i);
2277		h->max_huge_pages = i;
2278	}
2279
2280	kfree(node_alloc_noretry);
2281}
2282
2283static void __init hugetlb_init_hstates(void)
2284{
2285	struct hstate *h;
2286
2287	for_each_hstate(h) {
2288		if (minimum_order > huge_page_order(h))
2289			minimum_order = huge_page_order(h);
2290
2291		/* oversize hugepages were init'ed in early boot */
2292		if (!hstate_is_gigantic(h))
2293			hugetlb_hstate_alloc_pages(h);
2294	}
2295	VM_BUG_ON(minimum_order == UINT_MAX);
 
 
 
 
 
 
 
 
 
 
2296}
2297
2298static void __init report_hugepages(void)
2299{
2300	struct hstate *h;
2301
2302	for_each_hstate(h) {
2303		char buf[32];
2304
2305		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2306		pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2307			buf, h->free_huge_pages);
2308	}
2309}
2310
2311#ifdef CONFIG_HIGHMEM
2312static void try_to_free_low(struct hstate *h, unsigned long count,
2313						nodemask_t *nodes_allowed)
2314{
2315	int i;
2316
2317	if (hstate_is_gigantic(h))
2318		return;
2319
2320	for_each_node_mask(i, *nodes_allowed) {
2321		struct page *page, *next;
2322		struct list_head *freel = &h->hugepage_freelists[i];
2323		list_for_each_entry_safe(page, next, freel, lru) {
2324			if (count >= h->nr_huge_pages)
2325				return;
2326			if (PageHighMem(page))
2327				continue;
2328			list_del(&page->lru);
2329			update_and_free_page(h, page);
2330			h->free_huge_pages--;
2331			h->free_huge_pages_node[page_to_nid(page)]--;
2332		}
2333	}
2334}
2335#else
2336static inline void try_to_free_low(struct hstate *h, unsigned long count,
2337						nodemask_t *nodes_allowed)
2338{
2339}
2340#endif
2341
2342/*
2343 * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2344 * balanced by operating on them in a round-robin fashion.
2345 * Returns 1 if an adjustment was made.
2346 */
2347static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2348				int delta)
2349{
2350	int nr_nodes, node;
 
2351
2352	VM_BUG_ON(delta != -1 && delta != 1);
2353
2354	if (delta < 0) {
2355		for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2356			if (h->surplus_huge_pages_node[node])
2357				goto found;
 
 
 
 
 
 
 
 
 
 
 
 
 
2358		}
2359	} else {
2360		for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2361			if (h->surplus_huge_pages_node[node] <
2362					h->nr_huge_pages_node[node])
2363				goto found;
 
 
 
 
 
2364		}
2365	}
2366	return 0;
2367
2368found:
2369	h->surplus_huge_pages += delta;
2370	h->surplus_huge_pages_node[node] += delta;
2371	return 1;
 
 
 
2372}
2373
2374#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2375static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2376			      nodemask_t *nodes_allowed)
2377{
2378	unsigned long min_count, ret;
2379	NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2380
2381	/*
2382	 * Bit mask controlling how hard we retry per-node allocations.
2383	 * If we can not allocate the bit mask, do not attempt to allocate
2384	 * the requested huge pages.
2385	 */
2386	if (node_alloc_noretry)
2387		nodes_clear(*node_alloc_noretry);
2388	else
2389		return -ENOMEM;
2390
2391	spin_lock(&hugetlb_lock);
2392
2393	/*
2394	 * Check for a node specific request.
2395	 * Changing node specific huge page count may require a corresponding
2396	 * change to the global count.  In any case, the passed node mask
2397	 * (nodes_allowed) will restrict alloc/free to the specified node.
2398	 */
2399	if (nid != NUMA_NO_NODE) {
2400		unsigned long old_count = count;
2401
2402		count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2403		/*
2404		 * User may have specified a large count value which caused the
2405		 * above calculation to overflow.  In this case, they wanted
2406		 * to allocate as many huge pages as possible.  Set count to
2407		 * largest possible value to align with their intention.
2408		 */
2409		if (count < old_count)
2410			count = ULONG_MAX;
2411	}
2412
2413	/*
2414	 * Gigantic pages runtime allocation depend on the capability for large
2415	 * page range allocation.
2416	 * If the system does not provide this feature, return an error when
2417	 * the user tries to allocate gigantic pages but let the user free the
2418	 * boottime allocated gigantic pages.
2419	 */
2420	if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2421		if (count > persistent_huge_pages(h)) {
2422			spin_unlock(&hugetlb_lock);
2423			NODEMASK_FREE(node_alloc_noretry);
2424			return -EINVAL;
2425		}
2426		/* Fall through to decrease pool */
2427	}
2428
2429	/*
2430	 * Increase the pool size
2431	 * First take pages out of surplus state.  Then make up the
2432	 * remaining difference by allocating fresh huge pages.
2433	 *
2434	 * We might race with alloc_surplus_huge_page() here and be unable
2435	 * to convert a surplus huge page to a normal huge page. That is
2436	 * not critical, though, it just means the overall size of the
2437	 * pool might be one hugepage larger than it needs to be, but
2438	 * within all the constraints specified by the sysctls.
2439	 */
 
2440	while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2441		if (!adjust_pool_surplus(h, nodes_allowed, -1))
2442			break;
2443	}
2444
2445	while (count > persistent_huge_pages(h)) {
2446		/*
2447		 * If this allocation races such that we no longer need the
2448		 * page, free_huge_page will handle it by freeing the page
2449		 * and reducing the surplus.
2450		 */
2451		spin_unlock(&hugetlb_lock);
2452
2453		/* yield cpu to avoid soft lockup */
2454		cond_resched();
2455
2456		ret = alloc_pool_huge_page(h, nodes_allowed,
2457						node_alloc_noretry);
2458		spin_lock(&hugetlb_lock);
2459		if (!ret)
2460			goto out;
2461
2462		/* Bail for signals. Probably ctrl-c from user */
2463		if (signal_pending(current))
2464			goto out;
2465	}
2466
2467	/*
2468	 * Decrease the pool size
2469	 * First return free pages to the buddy allocator (being careful
2470	 * to keep enough around to satisfy reservations).  Then place
2471	 * pages into surplus state as needed so the pool will shrink
2472	 * to the desired size as pages become free.
2473	 *
2474	 * By placing pages into the surplus state independent of the
2475	 * overcommit value, we are allowing the surplus pool size to
2476	 * exceed overcommit. There are few sane options here. Since
2477	 * alloc_surplus_huge_page() is checking the global counter,
2478	 * though, we'll note that we're not allowed to exceed surplus
2479	 * and won't grow the pool anywhere else. Not until one of the
2480	 * sysctls are changed, or the surplus pages go out of use.
2481	 */
2482	min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2483	min_count = max(count, min_count);
2484	try_to_free_low(h, min_count, nodes_allowed);
2485	while (min_count < persistent_huge_pages(h)) {
2486		if (!free_pool_huge_page(h, nodes_allowed, 0))
2487			break;
2488		cond_resched_lock(&hugetlb_lock);
2489	}
2490	while (count < persistent_huge_pages(h)) {
2491		if (!adjust_pool_surplus(h, nodes_allowed, 1))
2492			break;
2493	}
2494out:
2495	h->max_huge_pages = persistent_huge_pages(h);
2496	spin_unlock(&hugetlb_lock);
2497
2498	NODEMASK_FREE(node_alloc_noretry);
2499
2500	return 0;
2501}
2502
2503#define HSTATE_ATTR_RO(_name) \
2504	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2505
2506#define HSTATE_ATTR(_name) \
2507	static struct kobj_attribute _name##_attr = \
2508		__ATTR(_name, 0644, _name##_show, _name##_store)
2509
2510static struct kobject *hugepages_kobj;
2511static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2512
2513static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2514
2515static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2516{
2517	int i;
2518
2519	for (i = 0; i < HUGE_MAX_HSTATE; i++)
2520		if (hstate_kobjs[i] == kobj) {
2521			if (nidp)
2522				*nidp = NUMA_NO_NODE;
2523			return &hstates[i];
2524		}
2525
2526	return kobj_to_node_hstate(kobj, nidp);
2527}
2528
2529static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2530					struct kobj_attribute *attr, char *buf)
2531{
2532	struct hstate *h;
2533	unsigned long nr_huge_pages;
2534	int nid;
2535
2536	h = kobj_to_hstate(kobj, &nid);
2537	if (nid == NUMA_NO_NODE)
2538		nr_huge_pages = h->nr_huge_pages;
2539	else
2540		nr_huge_pages = h->nr_huge_pages_node[nid];
2541
2542	return sprintf(buf, "%lu\n", nr_huge_pages);
2543}
2544
2545static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2546					   struct hstate *h, int nid,
2547					   unsigned long count, size_t len)
2548{
2549	int err;
2550	nodemask_t nodes_allowed, *n_mask;
 
 
 
2551
2552	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2553		return -EINVAL;
 
 
 
 
 
 
 
2554
2555	if (nid == NUMA_NO_NODE) {
2556		/*
2557		 * global hstate attribute
2558		 */
2559		if (!(obey_mempolicy &&
2560				init_nodemask_of_mempolicy(&nodes_allowed)))
2561			n_mask = &node_states[N_MEMORY];
2562		else
2563			n_mask = &nodes_allowed;
2564	} else {
2565		/*
2566		 * Node specific request.  count adjustment happens in
2567		 * set_max_huge_pages() after acquiring hugetlb_lock.
2568		 */
2569		init_nodemask_of_node(&nodes_allowed, nid);
2570		n_mask = &nodes_allowed;
2571	}
2572
2573	err = set_max_huge_pages(h, count, nid, n_mask);
2574
2575	return err ? err : len;
2576}
2577
2578static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2579					 struct kobject *kobj, const char *buf,
2580					 size_t len)
2581{
2582	struct hstate *h;
2583	unsigned long count;
2584	int nid;
2585	int err;
2586
2587	err = kstrtoul(buf, 10, &count);
2588	if (err)
2589		return err;
2590
2591	h = kobj_to_hstate(kobj, &nid);
2592	return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
 
 
2593}
2594
2595static ssize_t nr_hugepages_show(struct kobject *kobj,
2596				       struct kobj_attribute *attr, char *buf)
2597{
2598	return nr_hugepages_show_common(kobj, attr, buf);
2599}
2600
2601static ssize_t nr_hugepages_store(struct kobject *kobj,
2602	       struct kobj_attribute *attr, const char *buf, size_t len)
2603{
2604	return nr_hugepages_store_common(false, kobj, buf, len);
2605}
2606HSTATE_ATTR(nr_hugepages);
2607
2608#ifdef CONFIG_NUMA
2609
2610/*
2611 * hstate attribute for optionally mempolicy-based constraint on persistent
2612 * huge page alloc/free.
2613 */
2614static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2615				       struct kobj_attribute *attr, char *buf)
2616{
2617	return nr_hugepages_show_common(kobj, attr, buf);
2618}
2619
2620static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2621	       struct kobj_attribute *attr, const char *buf, size_t len)
2622{
2623	return nr_hugepages_store_common(true, kobj, buf, len);
2624}
2625HSTATE_ATTR(nr_hugepages_mempolicy);
2626#endif
2627
2628
2629static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2630					struct kobj_attribute *attr, char *buf)
2631{
2632	struct hstate *h = kobj_to_hstate(kobj, NULL);
2633	return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2634}
2635
2636static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2637		struct kobj_attribute *attr, const char *buf, size_t count)
2638{
2639	int err;
2640	unsigned long input;
2641	struct hstate *h = kobj_to_hstate(kobj, NULL);
2642
2643	if (hstate_is_gigantic(h))
2644		return -EINVAL;
2645
2646	err = kstrtoul(buf, 10, &input);
2647	if (err)
2648		return err;
2649
2650	spin_lock(&hugetlb_lock);
2651	h->nr_overcommit_huge_pages = input;
2652	spin_unlock(&hugetlb_lock);
2653
2654	return count;
2655}
2656HSTATE_ATTR(nr_overcommit_hugepages);
2657
2658static ssize_t free_hugepages_show(struct kobject *kobj,
2659					struct kobj_attribute *attr, char *buf)
2660{
2661	struct hstate *h;
2662	unsigned long free_huge_pages;
2663	int nid;
2664
2665	h = kobj_to_hstate(kobj, &nid);
2666	if (nid == NUMA_NO_NODE)
2667		free_huge_pages = h->free_huge_pages;
2668	else
2669		free_huge_pages = h->free_huge_pages_node[nid];
2670
2671	return sprintf(buf, "%lu\n", free_huge_pages);
2672}
2673HSTATE_ATTR_RO(free_hugepages);
2674
2675static ssize_t resv_hugepages_show(struct kobject *kobj,
2676					struct kobj_attribute *attr, char *buf)
2677{
2678	struct hstate *h = kobj_to_hstate(kobj, NULL);
2679	return sprintf(buf, "%lu\n", h->resv_huge_pages);
2680}
2681HSTATE_ATTR_RO(resv_hugepages);
2682
2683static ssize_t surplus_hugepages_show(struct kobject *kobj,
2684					struct kobj_attribute *attr, char *buf)
2685{
2686	struct hstate *h;
2687	unsigned long surplus_huge_pages;
2688	int nid;
2689
2690	h = kobj_to_hstate(kobj, &nid);
2691	if (nid == NUMA_NO_NODE)
2692		surplus_huge_pages = h->surplus_huge_pages;
2693	else
2694		surplus_huge_pages = h->surplus_huge_pages_node[nid];
2695
2696	return sprintf(buf, "%lu\n", surplus_huge_pages);
2697}
2698HSTATE_ATTR_RO(surplus_hugepages);
2699
2700static struct attribute *hstate_attrs[] = {
2701	&nr_hugepages_attr.attr,
2702	&nr_overcommit_hugepages_attr.attr,
2703	&free_hugepages_attr.attr,
2704	&resv_hugepages_attr.attr,
2705	&surplus_hugepages_attr.attr,
2706#ifdef CONFIG_NUMA
2707	&nr_hugepages_mempolicy_attr.attr,
2708#endif
2709	NULL,
2710};
2711
2712static const struct attribute_group hstate_attr_group = {
2713	.attrs = hstate_attrs,
2714};
2715
2716static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2717				    struct kobject **hstate_kobjs,
2718				    const struct attribute_group *hstate_attr_group)
2719{
2720	int retval;
2721	int hi = hstate_index(h);
2722
2723	hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2724	if (!hstate_kobjs[hi])
2725		return -ENOMEM;
2726
2727	retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2728	if (retval)
2729		kobject_put(hstate_kobjs[hi]);
2730
2731	return retval;
2732}
2733
2734static void __init hugetlb_sysfs_init(void)
2735{
2736	struct hstate *h;
2737	int err;
2738
2739	hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2740	if (!hugepages_kobj)
2741		return;
2742
2743	for_each_hstate(h) {
2744		err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2745					 hstate_kobjs, &hstate_attr_group);
2746		if (err)
2747			pr_err("Hugetlb: Unable to add hstate %s", h->name);
 
2748	}
2749}
2750
2751#ifdef CONFIG_NUMA
2752
2753/*
2754 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2755 * with node devices in node_devices[] using a parallel array.  The array
2756 * index of a node device or _hstate == node id.
2757 * This is here to avoid any static dependency of the node device driver, in
2758 * the base kernel, on the hugetlb module.
2759 */
2760struct node_hstate {
2761	struct kobject		*hugepages_kobj;
2762	struct kobject		*hstate_kobjs[HUGE_MAX_HSTATE];
2763};
2764static struct node_hstate node_hstates[MAX_NUMNODES];
2765
2766/*
2767 * A subset of global hstate attributes for node devices
2768 */
2769static struct attribute *per_node_hstate_attrs[] = {
2770	&nr_hugepages_attr.attr,
2771	&free_hugepages_attr.attr,
2772	&surplus_hugepages_attr.attr,
2773	NULL,
2774};
2775
2776static const struct attribute_group per_node_hstate_attr_group = {
2777	.attrs = per_node_hstate_attrs,
2778};
2779
2780/*
2781 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2782 * Returns node id via non-NULL nidp.
2783 */
2784static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2785{
2786	int nid;
2787
2788	for (nid = 0; nid < nr_node_ids; nid++) {
2789		struct node_hstate *nhs = &node_hstates[nid];
2790		int i;
2791		for (i = 0; i < HUGE_MAX_HSTATE; i++)
2792			if (nhs->hstate_kobjs[i] == kobj) {
2793				if (nidp)
2794					*nidp = nid;
2795				return &hstates[i];
2796			}
2797	}
2798
2799	BUG();
2800	return NULL;
2801}
2802
2803/*
2804 * Unregister hstate attributes from a single node device.
2805 * No-op if no hstate attributes attached.
2806 */
2807static void hugetlb_unregister_node(struct node *node)
2808{
2809	struct hstate *h;
2810	struct node_hstate *nhs = &node_hstates[node->dev.id];
2811
2812	if (!nhs->hugepages_kobj)
2813		return;		/* no hstate attributes */
2814
2815	for_each_hstate(h) {
2816		int idx = hstate_index(h);
2817		if (nhs->hstate_kobjs[idx]) {
2818			kobject_put(nhs->hstate_kobjs[idx]);
2819			nhs->hstate_kobjs[idx] = NULL;
2820		}
2821	}
2822
2823	kobject_put(nhs->hugepages_kobj);
2824	nhs->hugepages_kobj = NULL;
2825}
2826
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2827
2828/*
2829 * Register hstate attributes for a single node device.
2830 * No-op if attributes already registered.
2831 */
2832static void hugetlb_register_node(struct node *node)
2833{
2834	struct hstate *h;
2835	struct node_hstate *nhs = &node_hstates[node->dev.id];
2836	int err;
2837
2838	if (nhs->hugepages_kobj)
2839		return;		/* already allocated */
2840
2841	nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2842							&node->dev.kobj);
2843	if (!nhs->hugepages_kobj)
2844		return;
2845
2846	for_each_hstate(h) {
2847		err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2848						nhs->hstate_kobjs,
2849						&per_node_hstate_attr_group);
2850		if (err) {
2851			pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2852				h->name, node->dev.id);
 
2853			hugetlb_unregister_node(node);
2854			break;
2855		}
2856	}
2857}
2858
2859/*
2860 * hugetlb init time:  register hstate attributes for all registered node
2861 * devices of nodes that have memory.  All on-line nodes should have
2862 * registered their associated device by this time.
2863 */
2864static void __init hugetlb_register_all_nodes(void)
2865{
2866	int nid;
2867
2868	for_each_node_state(nid, N_MEMORY) {
2869		struct node *node = node_devices[nid];
2870		if (node->dev.id == nid)
2871			hugetlb_register_node(node);
2872	}
2873
2874	/*
2875	 * Let the node device driver know we're here so it can
2876	 * [un]register hstate attributes on node hotplug.
2877	 */
2878	register_hugetlbfs_with_node(hugetlb_register_node,
2879				     hugetlb_unregister_node);
2880}
2881#else	/* !CONFIG_NUMA */
2882
2883static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2884{
2885	BUG();
2886	if (nidp)
2887		*nidp = -1;
2888	return NULL;
2889}
2890
 
 
2891static void hugetlb_register_all_nodes(void) { }
2892
2893#endif
2894
2895static int __init hugetlb_init(void)
2896{
2897	int i;
 
 
 
 
 
 
2898
2899	if (!hugepages_supported())
 
 
 
 
 
 
 
 
 
 
2900		return 0;
2901
2902	if (!size_to_hstate(default_hstate_size)) {
2903		if (default_hstate_size != 0) {
2904			pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2905			       default_hstate_size, HPAGE_SIZE);
2906		}
2907
2908		default_hstate_size = HPAGE_SIZE;
2909		if (!size_to_hstate(default_hstate_size))
2910			hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2911	}
2912	default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2913	if (default_hstate_max_huge_pages) {
2914		if (!default_hstate.max_huge_pages)
2915			default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2916	}
2917
2918	hugetlb_init_hstates();
 
2919	gather_bootmem_prealloc();
 
2920	report_hugepages();
2921
2922	hugetlb_sysfs_init();
2923	hugetlb_register_all_nodes();
2924	hugetlb_cgroup_file_init();
2925
2926#ifdef CONFIG_SMP
2927	num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2928#else
2929	num_fault_mutexes = 1;
2930#endif
2931	hugetlb_fault_mutex_table =
2932		kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2933			      GFP_KERNEL);
2934	BUG_ON(!hugetlb_fault_mutex_table);
2935
2936	for (i = 0; i < num_fault_mutexes; i++)
2937		mutex_init(&hugetlb_fault_mutex_table[i]);
2938	return 0;
2939}
2940subsys_initcall(hugetlb_init);
2941
2942/* Should be called on processing a hugepagesz=... option */
2943void __init hugetlb_bad_size(void)
2944{
2945	parsed_valid_hugepagesz = false;
2946}
2947
2948void __init hugetlb_add_hstate(unsigned int order)
2949{
2950	struct hstate *h;
2951	unsigned long i;
2952
2953	if (size_to_hstate(PAGE_SIZE << order)) {
2954		pr_warn("hugepagesz= specified twice, ignoring\n");
2955		return;
2956	}
2957	BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2958	BUG_ON(order == 0);
2959	h = &hstates[hugetlb_max_hstate++];
2960	h->order = order;
2961	h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2962	h->nr_huge_pages = 0;
2963	h->free_huge_pages = 0;
2964	for (i = 0; i < MAX_NUMNODES; ++i)
2965		INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2966	INIT_LIST_HEAD(&h->hugepage_activelist);
2967	h->next_nid_to_alloc = first_memory_node;
2968	h->next_nid_to_free = first_memory_node;
2969	snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2970					huge_page_size(h)/1024);
2971
2972	parsed_hstate = h;
2973}
2974
2975static int __init hugetlb_nrpages_setup(char *s)
2976{
2977	unsigned long *mhp;
2978	static unsigned long *last_mhp;
2979
2980	if (!parsed_valid_hugepagesz) {
2981		pr_warn("hugepages = %s preceded by "
2982			"an unsupported hugepagesz, ignoring\n", s);
2983		parsed_valid_hugepagesz = true;
2984		return 1;
2985	}
2986	/*
2987	 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2988	 * so this hugepages= parameter goes to the "default hstate".
2989	 */
2990	else if (!hugetlb_max_hstate)
2991		mhp = &default_hstate_max_huge_pages;
2992	else
2993		mhp = &parsed_hstate->max_huge_pages;
2994
2995	if (mhp == last_mhp) {
2996		pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
 
2997		return 1;
2998	}
2999
3000	if (sscanf(s, "%lu", mhp) <= 0)
3001		*mhp = 0;
3002
3003	/*
3004	 * Global state is always initialized later in hugetlb_init.
3005	 * But we need to allocate >= MAX_ORDER hstates here early to still
3006	 * use the bootmem allocator.
3007	 */
3008	if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3009		hugetlb_hstate_alloc_pages(parsed_hstate);
3010
3011	last_mhp = mhp;
3012
3013	return 1;
3014}
3015__setup("hugepages=", hugetlb_nrpages_setup);
3016
3017static int __init hugetlb_default_setup(char *s)
3018{
3019	default_hstate_size = memparse(s, &s);
3020	return 1;
3021}
3022__setup("default_hugepagesz=", hugetlb_default_setup);
3023
3024static unsigned int cpuset_mems_nr(unsigned int *array)
3025{
3026	int node;
3027	unsigned int nr = 0;
3028
3029	for_each_node_mask(node, cpuset_current_mems_allowed)
3030		nr += array[node];
3031
3032	return nr;
3033}
3034
3035#ifdef CONFIG_SYSCTL
3036static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3037			 struct ctl_table *table, int write,
3038			 void __user *buffer, size_t *length, loff_t *ppos)
3039{
3040	struct hstate *h = &default_hstate;
3041	unsigned long tmp = h->max_huge_pages;
3042	int ret;
3043
3044	if (!hugepages_supported())
3045		return -EOPNOTSUPP;
 
 
3046
3047	table->data = &tmp;
3048	table->maxlen = sizeof(unsigned long);
3049	ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3050	if (ret)
3051		goto out;
3052
3053	if (write)
3054		ret = __nr_hugepages_store_common(obey_mempolicy, h,
3055						  NUMA_NO_NODE, tmp, *length);
 
 
 
 
 
 
 
 
 
 
3056out:
3057	return ret;
3058}
3059
3060int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3061			  void __user *buffer, size_t *length, loff_t *ppos)
3062{
3063
3064	return hugetlb_sysctl_handler_common(false, table, write,
3065							buffer, length, ppos);
3066}
3067
3068#ifdef CONFIG_NUMA
3069int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3070			  void __user *buffer, size_t *length, loff_t *ppos)
3071{
3072	return hugetlb_sysctl_handler_common(true, table, write,
3073							buffer, length, ppos);
3074}
3075#endif /* CONFIG_NUMA */
3076
 
 
 
 
 
 
 
 
 
 
 
 
3077int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3078			void __user *buffer,
3079			size_t *length, loff_t *ppos)
3080{
3081	struct hstate *h = &default_hstate;
3082	unsigned long tmp;
3083	int ret;
3084
3085	if (!hugepages_supported())
3086		return -EOPNOTSUPP;
3087
3088	tmp = h->nr_overcommit_huge_pages;
3089
3090	if (write && hstate_is_gigantic(h))
3091		return -EINVAL;
3092
3093	table->data = &tmp;
3094	table->maxlen = sizeof(unsigned long);
3095	ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3096	if (ret)
3097		goto out;
3098
3099	if (write) {
3100		spin_lock(&hugetlb_lock);
3101		h->nr_overcommit_huge_pages = tmp;
3102		spin_unlock(&hugetlb_lock);
3103	}
3104out:
3105	return ret;
3106}
3107
3108#endif /* CONFIG_SYSCTL */
3109
3110void hugetlb_report_meminfo(struct seq_file *m)
3111{
3112	struct hstate *h;
3113	unsigned long total = 0;
3114
3115	if (!hugepages_supported())
3116		return;
3117
3118	for_each_hstate(h) {
3119		unsigned long count = h->nr_huge_pages;
3120
3121		total += (PAGE_SIZE << huge_page_order(h)) * count;
3122
3123		if (h == &default_hstate)
3124			seq_printf(m,
3125				   "HugePages_Total:   %5lu\n"
3126				   "HugePages_Free:    %5lu\n"
3127				   "HugePages_Rsvd:    %5lu\n"
3128				   "HugePages_Surp:    %5lu\n"
3129				   "Hugepagesize:   %8lu kB\n",
3130				   count,
3131				   h->free_huge_pages,
3132				   h->resv_huge_pages,
3133				   h->surplus_huge_pages,
3134				   (PAGE_SIZE << huge_page_order(h)) / 1024);
3135	}
3136
3137	seq_printf(m, "Hugetlb:        %8lu kB\n", total / 1024);
3138}
3139
3140int hugetlb_report_node_meminfo(int nid, char *buf)
3141{
3142	struct hstate *h = &default_hstate;
3143	if (!hugepages_supported())
3144		return 0;
3145	return sprintf(buf,
3146		"Node %d HugePages_Total: %5u\n"
3147		"Node %d HugePages_Free:  %5u\n"
3148		"Node %d HugePages_Surp:  %5u\n",
3149		nid, h->nr_huge_pages_node[nid],
3150		nid, h->free_huge_pages_node[nid],
3151		nid, h->surplus_huge_pages_node[nid]);
3152}
3153
3154void hugetlb_show_meminfo(void)
3155{
3156	struct hstate *h;
3157	int nid;
3158
3159	if (!hugepages_supported())
3160		return;
3161
3162	for_each_node_state(nid, N_MEMORY)
3163		for_each_hstate(h)
3164			pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3165				nid,
3166				h->nr_huge_pages_node[nid],
3167				h->free_huge_pages_node[nid],
3168				h->surplus_huge_pages_node[nid],
3169				1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3170}
3171
3172void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3173{
3174	seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3175		   atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3176}
3177
3178/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3179unsigned long hugetlb_total_pages(void)
3180{
3181	struct hstate *h;
3182	unsigned long nr_total_pages = 0;
3183
3184	for_each_hstate(h)
3185		nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3186	return nr_total_pages;
3187}
3188
3189static int hugetlb_acct_memory(struct hstate *h, long delta)
3190{
3191	int ret = -ENOMEM;
3192
3193	spin_lock(&hugetlb_lock);
3194	/*
3195	 * When cpuset is configured, it breaks the strict hugetlb page
3196	 * reservation as the accounting is done on a global variable. Such
3197	 * reservation is completely rubbish in the presence of cpuset because
3198	 * the reservation is not checked against page availability for the
3199	 * current cpuset. Application can still potentially OOM'ed by kernel
3200	 * with lack of free htlb page in cpuset that the task is in.
3201	 * Attempt to enforce strict accounting with cpuset is almost
3202	 * impossible (or too ugly) because cpuset is too fluid that
3203	 * task or memory node can be dynamically moved between cpusets.
3204	 *
3205	 * The change of semantics for shared hugetlb mapping with cpuset is
3206	 * undesirable. However, in order to preserve some of the semantics,
3207	 * we fall back to check against current free page availability as
3208	 * a best attempt and hopefully to minimize the impact of changing
3209	 * semantics that cpuset has.
3210	 */
3211	if (delta > 0) {
3212		if (gather_surplus_pages(h, delta) < 0)
3213			goto out;
3214
3215		if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3216			return_unused_surplus_pages(h, delta);
3217			goto out;
3218		}
3219	}
3220
3221	ret = 0;
3222	if (delta < 0)
3223		return_unused_surplus_pages(h, (unsigned long) -delta);
3224
3225out:
3226	spin_unlock(&hugetlb_lock);
3227	return ret;
3228}
3229
3230static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3231{
3232	struct resv_map *resv = vma_resv_map(vma);
3233
3234	/*
3235	 * This new VMA should share its siblings reservation map if present.
3236	 * The VMA will only ever have a valid reservation map pointer where
3237	 * it is being copied for another still existing VMA.  As that VMA
3238	 * has a reference to the reservation map it cannot disappear until
3239	 * after this open call completes.  It is therefore safe to take a
3240	 * new reference here without additional locking.
3241	 */
3242	if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3243		kref_get(&resv->refs);
3244}
3245
3246static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3247{
3248	struct hstate *h = hstate_vma(vma);
3249	struct resv_map *resv = vma_resv_map(vma);
3250	struct hugepage_subpool *spool = subpool_vma(vma);
3251	unsigned long reserve, start, end;
3252	long gbl_reserve;
3253
3254	if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3255		return;
3256
3257	start = vma_hugecache_offset(h, vma, vma->vm_start);
3258	end = vma_hugecache_offset(h, vma, vma->vm_end);
3259
3260	reserve = (end - start) - region_count(resv, start, end);
3261
3262	kref_put(&resv->refs, resv_map_release);
3263
3264	if (reserve) {
3265		/*
3266		 * Decrement reserve counts.  The global reserve count may be
3267		 * adjusted if the subpool has a minimum size.
3268		 */
3269		gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3270		hugetlb_acct_memory(h, -gbl_reserve);
3271	}
3272}
3273
3274static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3275{
3276	if (addr & ~(huge_page_mask(hstate_vma(vma))))
3277		return -EINVAL;
3278	return 0;
3279}
3280
3281static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3282{
3283	struct hstate *hstate = hstate_vma(vma);
3284
3285	return 1UL << huge_page_shift(hstate);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3286}
3287
3288/*
3289 * We cannot handle pagefaults against hugetlb pages at all.  They cause
3290 * handle_mm_fault() to try to instantiate regular-sized pages in the
3291 * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3292 * this far.
3293 */
3294static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3295{
3296	BUG();
3297	return 0;
3298}
3299
3300/*
3301 * When a new function is introduced to vm_operations_struct and added
3302 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3303 * This is because under System V memory model, mappings created via
3304 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3305 * their original vm_ops are overwritten with shm_vm_ops.
3306 */
3307const struct vm_operations_struct hugetlb_vm_ops = {
3308	.fault = hugetlb_vm_op_fault,
3309	.open = hugetlb_vm_op_open,
3310	.close = hugetlb_vm_op_close,
3311	.split = hugetlb_vm_op_split,
3312	.pagesize = hugetlb_vm_op_pagesize,
3313};
3314
3315static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3316				int writable)
3317{
3318	pte_t entry;
3319
3320	if (writable) {
3321		entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3322					 vma->vm_page_prot)));
3323	} else {
3324		entry = huge_pte_wrprotect(mk_huge_pte(page,
3325					   vma->vm_page_prot));
3326	}
3327	entry = pte_mkyoung(entry);
3328	entry = pte_mkhuge(entry);
3329	entry = arch_make_huge_pte(entry, vma, page, writable);
3330
3331	return entry;
3332}
3333
3334static void set_huge_ptep_writable(struct vm_area_struct *vma,
3335				   unsigned long address, pte_t *ptep)
3336{
3337	pte_t entry;
3338
3339	entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3340	if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3341		update_mmu_cache(vma, address, ptep);
3342}
3343
3344bool is_hugetlb_entry_migration(pte_t pte)
3345{
3346	swp_entry_t swp;
3347
3348	if (huge_pte_none(pte) || pte_present(pte))
3349		return false;
3350	swp = pte_to_swp_entry(pte);
3351	if (non_swap_entry(swp) && is_migration_entry(swp))
3352		return true;
3353	else
3354		return false;
3355}
3356
3357static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3358{
3359	swp_entry_t swp;
3360
3361	if (huge_pte_none(pte) || pte_present(pte))
3362		return 0;
3363	swp = pte_to_swp_entry(pte);
3364	if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3365		return 1;
3366	else
3367		return 0;
3368}
3369
3370int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3371			    struct vm_area_struct *vma)
3372{
3373	pte_t *src_pte, *dst_pte, entry, dst_entry;
3374	struct page *ptepage;
3375	unsigned long addr;
3376	int cow;
3377	struct hstate *h = hstate_vma(vma);
3378	unsigned long sz = huge_page_size(h);
3379	struct mmu_notifier_range range;
3380	int ret = 0;
3381
3382	cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3383
3384	if (cow) {
3385		mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3386					vma->vm_start,
3387					vma->vm_end);
3388		mmu_notifier_invalidate_range_start(&range);
3389	}
3390
3391	for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3392		spinlock_t *src_ptl, *dst_ptl;
3393		src_pte = huge_pte_offset(src, addr, sz);
3394		if (!src_pte)
3395			continue;
3396		dst_pte = huge_pte_alloc(dst, addr, sz);
3397		if (!dst_pte) {
3398			ret = -ENOMEM;
3399			break;
3400		}
3401
3402		/*
3403		 * If the pagetables are shared don't copy or take references.
3404		 * dst_pte == src_pte is the common case of src/dest sharing.
3405		 *
3406		 * However, src could have 'unshared' and dst shares with
3407		 * another vma.  If dst_pte !none, this implies sharing.
3408		 * Check here before taking page table lock, and once again
3409		 * after taking the lock below.
3410		 */
3411		dst_entry = huge_ptep_get(dst_pte);
3412		if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3413			continue;
3414
3415		dst_ptl = huge_pte_lock(h, dst, dst_pte);
3416		src_ptl = huge_pte_lockptr(h, src, src_pte);
3417		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3418		entry = huge_ptep_get(src_pte);
3419		dst_entry = huge_ptep_get(dst_pte);
3420		if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3421			/*
3422			 * Skip if src entry none.  Also, skip in the
3423			 * unlikely case dst entry !none as this implies
3424			 * sharing with another vma.
3425			 */
3426			;
3427		} else if (unlikely(is_hugetlb_entry_migration(entry) ||
3428				    is_hugetlb_entry_hwpoisoned(entry))) {
3429			swp_entry_t swp_entry = pte_to_swp_entry(entry);
3430
3431			if (is_write_migration_entry(swp_entry) && cow) {
3432				/*
3433				 * COW mappings require pages in both
3434				 * parent and child to be set to read.
3435				 */
3436				make_migration_entry_read(&swp_entry);
3437				entry = swp_entry_to_pte(swp_entry);
3438				set_huge_swap_pte_at(src, addr, src_pte,
3439						     entry, sz);
3440			}
3441			set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3442		} else {
3443			if (cow) {
3444				/*
3445				 * No need to notify as we are downgrading page
3446				 * table protection not changing it to point
3447				 * to a new page.
3448				 *
3449				 * See Documentation/vm/mmu_notifier.rst
3450				 */
3451				huge_ptep_set_wrprotect(src, addr, src_pte);
3452			}
3453			entry = huge_ptep_get(src_pte);
3454			ptepage = pte_page(entry);
3455			get_page(ptepage);
3456			page_dup_rmap(ptepage, true);
3457			set_huge_pte_at(dst, addr, dst_pte, entry);
3458			hugetlb_count_add(pages_per_huge_page(h), dst);
3459		}
3460		spin_unlock(src_ptl);
3461		spin_unlock(dst_ptl);
3462	}
 
3463
3464	if (cow)
3465		mmu_notifier_invalidate_range_end(&range);
 
3466
3467	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3468}
3469
3470void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3471			    unsigned long start, unsigned long end,
3472			    struct page *ref_page)
3473{
3474	struct mm_struct *mm = vma->vm_mm;
3475	unsigned long address;
3476	pte_t *ptep;
3477	pte_t pte;
3478	spinlock_t *ptl;
3479	struct page *page;
 
3480	struct hstate *h = hstate_vma(vma);
3481	unsigned long sz = huge_page_size(h);
3482	struct mmu_notifier_range range;
 
 
 
 
 
 
3483
3484	WARN_ON(!is_vm_hugetlb_page(vma));
3485	BUG_ON(start & ~huge_page_mask(h));
3486	BUG_ON(end & ~huge_page_mask(h));
3487
3488	/*
3489	 * This is a hugetlb vma, all the pte entries should point
3490	 * to huge page.
3491	 */
3492	tlb_change_page_size(tlb, sz);
3493	tlb_start_vma(tlb, vma);
3494
3495	/*
3496	 * If sharing possible, alert mmu notifiers of worst case.
3497	 */
3498	mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3499				end);
3500	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3501	mmu_notifier_invalidate_range_start(&range);
3502	address = start;
3503	for (; address < end; address += sz) {
3504		ptep = huge_pte_offset(mm, address, sz);
3505		if (!ptep)
3506			continue;
3507
3508		ptl = huge_pte_lock(h, mm, ptep);
3509		if (huge_pmd_unshare(mm, &address, ptep)) {
3510			spin_unlock(ptl);
3511			/*
3512			 * We just unmapped a page of PMDs by clearing a PUD.
3513			 * The caller's TLB flush range should cover this area.
3514			 */
3515			continue;
3516		}
3517
3518		pte = huge_ptep_get(ptep);
3519		if (huge_pte_none(pte)) {
3520			spin_unlock(ptl);
3521			continue;
3522		}
3523
3524		/*
3525		 * Migrating hugepage or HWPoisoned hugepage is already
3526		 * unmapped and its refcount is dropped, so just clear pte here.
3527		 */
3528		if (unlikely(!pte_present(pte))) {
3529			huge_pte_clear(mm, address, ptep, sz);
3530			spin_unlock(ptl);
3531			continue;
3532		}
3533
3534		page = pte_page(pte);
3535		/*
3536		 * If a reference page is supplied, it is because a specific
3537		 * page is being unmapped, not a range. Ensure the page we
3538		 * are about to unmap is the actual page of interest.
3539		 */
3540		if (ref_page) {
3541			if (page != ref_page) {
3542				spin_unlock(ptl);
3543				continue;
3544			}
3545			/*
3546			 * Mark the VMA as having unmapped its page so that
3547			 * future faults in this VMA will fail rather than
3548			 * looking like data was lost
3549			 */
3550			set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3551		}
3552
3553		pte = huge_ptep_get_and_clear(mm, address, ptep);
3554		tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3555		if (huge_pte_dirty(pte))
3556			set_page_dirty(page);
 
3557
3558		hugetlb_count_sub(pages_per_huge_page(h), mm);
3559		page_remove_rmap(page, true);
3560
3561		spin_unlock(ptl);
3562		tlb_remove_page_size(tlb, page, huge_page_size(h));
3563		/*
3564		 * Bail out after unmapping reference page if supplied
3565		 */
3566		if (ref_page)
3567			break;
3568	}
3569	mmu_notifier_invalidate_range_end(&range);
3570	tlb_end_vma(tlb, vma);
 
 
 
 
 
 
3571}
3572
3573void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3574			  struct vm_area_struct *vma, unsigned long start,
3575			  unsigned long end, struct page *ref_page)
3576{
3577	__unmap_hugepage_range(tlb, vma, start, end, ref_page);
3578
3579	/*
3580	 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3581	 * test will fail on a vma being torn down, and not grab a page table
3582	 * on its way out.  We're lucky that the flag has such an appropriate
3583	 * name, and can in fact be safely cleared here. We could clear it
3584	 * before the __unmap_hugepage_range above, but all that's necessary
3585	 * is to clear it before releasing the i_mmap_rwsem. This works
3586	 * because in the context this is called, the VMA is about to be
3587	 * destroyed and the i_mmap_rwsem is held.
 
 
 
 
 
3588	 */
3589	vma->vm_flags &= ~VM_MAYSHARE;
3590}
3591
3592void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3593			  unsigned long end, struct page *ref_page)
3594{
3595	struct mm_struct *mm;
3596	struct mmu_gather tlb;
3597	unsigned long tlb_start = start;
3598	unsigned long tlb_end = end;
3599
3600	/*
3601	 * If shared PMDs were possibly used within this vma range, adjust
3602	 * start/end for worst case tlb flushing.
3603	 * Note that we can not be sure if PMDs are shared until we try to
3604	 * unmap pages.  However, we want to make sure TLB flushing covers
3605	 * the largest possible range.
3606	 */
3607	adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3608
3609	mm = vma->vm_mm;
3610
3611	tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3612	__unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3613	tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3614}
3615
3616/*
3617 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3618 * mappping it owns the reserve page for. The intention is to unmap the page
3619 * from other VMAs and let the children be SIGKILLed if they are faulting the
3620 * same region.
3621 */
3622static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3623			      struct page *page, unsigned long address)
3624{
3625	struct hstate *h = hstate_vma(vma);
3626	struct vm_area_struct *iter_vma;
3627	struct address_space *mapping;
 
3628	pgoff_t pgoff;
3629
3630	/*
3631	 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3632	 * from page cache lookup which is in HPAGE_SIZE units.
3633	 */
3634	address = address & huge_page_mask(h);
3635	pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3636			vma->vm_pgoff;
3637	mapping = vma->vm_file->f_mapping;
3638
3639	/*
3640	 * Take the mapping lock for the duration of the table walk. As
3641	 * this mapping should be shared between all the VMAs,
3642	 * __unmap_hugepage_range() is called as the lock is already held
3643	 */
3644	i_mmap_lock_write(mapping);
3645	vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3646		/* Do not unmap the current VMA */
3647		if (iter_vma == vma)
3648			continue;
3649
3650		/*
3651		 * Shared VMAs have their own reserves and do not affect
3652		 * MAP_PRIVATE accounting but it is possible that a shared
3653		 * VMA is using the same page so check and skip such VMAs.
3654		 */
3655		if (iter_vma->vm_flags & VM_MAYSHARE)
3656			continue;
3657
3658		/*
3659		 * Unmap the page from other VMAs without their own reserves.
3660		 * They get marked to be SIGKILLed if they fault in these
3661		 * areas. This is because a future no-page fault on this VMA
3662		 * could insert a zeroed page instead of the data existing
3663		 * from the time of fork. This would look like data corruption
3664		 */
3665		if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3666			unmap_hugepage_range(iter_vma, address,
3667					     address + huge_page_size(h), page);
 
3668	}
3669	i_mmap_unlock_write(mapping);
 
 
3670}
3671
3672/*
3673 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3674 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3675 * cannot race with other handlers or page migration.
3676 * Keep the pte_same checks anyway to make transition from the mutex easier.
3677 */
3678static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3679		       unsigned long address, pte_t *ptep,
3680		       struct page *pagecache_page, spinlock_t *ptl)
3681{
3682	pte_t pte;
3683	struct hstate *h = hstate_vma(vma);
3684	struct page *old_page, *new_page;
 
3685	int outside_reserve = 0;
3686	vm_fault_t ret = 0;
3687	unsigned long haddr = address & huge_page_mask(h);
3688	struct mmu_notifier_range range;
3689
3690	pte = huge_ptep_get(ptep);
3691	old_page = pte_page(pte);
3692
3693retry_avoidcopy:
3694	/* If no-one else is actually using this page, avoid the copy
3695	 * and just make the page writable */
3696	if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3697		page_move_anon_rmap(old_page, vma);
3698		set_huge_ptep_writable(vma, haddr, ptep);
 
 
3699		return 0;
3700	}
3701
3702	/*
3703	 * If the process that created a MAP_PRIVATE mapping is about to
3704	 * perform a COW due to a shared page count, attempt to satisfy
3705	 * the allocation without using the existing reserves. The pagecache
3706	 * page is used to determine if the reserve at this address was
3707	 * consumed or not. If reserves were used, a partial faulted mapping
3708	 * at the time of fork() could consume its reserves on COW instead
3709	 * of the full address range.
3710	 */
3711	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
 
3712			old_page != pagecache_page)
3713		outside_reserve = 1;
3714
3715	get_page(old_page);
3716
3717	/*
3718	 * Drop page table lock as buddy allocator may be called. It will
3719	 * be acquired again before returning to the caller, as expected.
3720	 */
3721	spin_unlock(ptl);
3722	new_page = alloc_huge_page(vma, haddr, outside_reserve);
3723
3724	if (IS_ERR(new_page)) {
 
 
3725		/*
3726		 * If a process owning a MAP_PRIVATE mapping fails to COW,
3727		 * it is due to references held by a child and an insufficient
3728		 * huge page pool. To guarantee the original mappers
3729		 * reliability, unmap the page from child processes. The child
3730		 * may get SIGKILLed if it later faults.
3731		 */
3732		if (outside_reserve) {
3733			put_page(old_page);
3734			BUG_ON(huge_pte_none(pte));
3735			unmap_ref_private(mm, vma, old_page, haddr);
3736			BUG_ON(huge_pte_none(pte));
3737			spin_lock(ptl);
3738			ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3739			if (likely(ptep &&
3740				   pte_same(huge_ptep_get(ptep), pte)))
3741				goto retry_avoidcopy;
3742			/*
3743			 * race occurs while re-acquiring page table
3744			 * lock, and our job is done.
3745			 */
3746			return 0;
 
3747		}
3748
3749		ret = vmf_error(PTR_ERR(new_page));
3750		goto out_release_old;
 
3751	}
3752
3753	/*
3754	 * When the original hugepage is shared one, it does not have
3755	 * anon_vma prepared.
3756	 */
3757	if (unlikely(anon_vma_prepare(vma))) {
3758		ret = VM_FAULT_OOM;
3759		goto out_release_all;
 
 
 
3760	}
3761
3762	copy_user_huge_page(new_page, old_page, address, vma,
3763			    pages_per_huge_page(h));
3764	__SetPageUptodate(new_page);
3765
3766	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3767				haddr + huge_page_size(h));
3768	mmu_notifier_invalidate_range_start(&range);
3769
3770	/*
3771	 * Retake the page table lock to check for racing updates
3772	 * before the page tables are altered
3773	 */
3774	spin_lock(ptl);
3775	ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3776	if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3777		ClearPagePrivate(new_page);
3778
3779		/* Break COW */
3780		huge_ptep_clear_flush(vma, haddr, ptep);
3781		mmu_notifier_invalidate_range(mm, range.start, range.end);
3782		set_huge_pte_at(mm, haddr, ptep,
 
 
3783				make_huge_pte(vma, new_page, 1));
3784		page_remove_rmap(old_page, true);
3785		hugepage_add_new_anon_rmap(new_page, vma, haddr);
3786		set_page_huge_active(new_page);
3787		/* Make the old page be freed below */
3788		new_page = old_page;
 
 
 
3789	}
3790	spin_unlock(ptl);
3791	mmu_notifier_invalidate_range_end(&range);
3792out_release_all:
3793	restore_reserve_on_error(h, vma, haddr, new_page);
3794	put_page(new_page);
3795out_release_old:
3796	put_page(old_page);
3797
3798	spin_lock(ptl); /* Caller expects lock to be held */
3799	return ret;
3800}
3801
3802/* Return the pagecache page at a given address within a VMA */
3803static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3804			struct vm_area_struct *vma, unsigned long address)
3805{
3806	struct address_space *mapping;
3807	pgoff_t idx;
3808
3809	mapping = vma->vm_file->f_mapping;
3810	idx = vma_hugecache_offset(h, vma, address);
3811
3812	return find_lock_page(mapping, idx);
3813}
3814
3815/*
3816 * Return whether there is a pagecache page to back given address within VMA.
3817 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3818 */
3819static bool hugetlbfs_pagecache_present(struct hstate *h,
3820			struct vm_area_struct *vma, unsigned long address)
3821{
3822	struct address_space *mapping;
3823	pgoff_t idx;
3824	struct page *page;
3825
3826	mapping = vma->vm_file->f_mapping;
3827	idx = vma_hugecache_offset(h, vma, address);
3828
3829	page = find_get_page(mapping, idx);
3830	if (page)
3831		put_page(page);
3832	return page != NULL;
3833}
3834
3835int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3836			   pgoff_t idx)
3837{
3838	struct inode *inode = mapping->host;
3839	struct hstate *h = hstate_inode(inode);
3840	int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3841
3842	if (err)
3843		return err;
3844	ClearPagePrivate(page);
3845
3846	/*
3847	 * set page dirty so that it will not be removed from cache/file
3848	 * by non-hugetlbfs specific code paths.
3849	 */
3850	set_page_dirty(page);
3851
3852	spin_lock(&inode->i_lock);
3853	inode->i_blocks += blocks_per_huge_page(h);
3854	spin_unlock(&inode->i_lock);
3855	return 0;
3856}
3857
3858static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3859			struct vm_area_struct *vma,
3860			struct address_space *mapping, pgoff_t idx,
3861			unsigned long address, pte_t *ptep, unsigned int flags)
3862{
3863	struct hstate *h = hstate_vma(vma);
3864	vm_fault_t ret = VM_FAULT_SIGBUS;
3865	int anon_rmap = 0;
 
3866	unsigned long size;
3867	struct page *page;
 
3868	pte_t new_pte;
3869	spinlock_t *ptl;
3870	unsigned long haddr = address & huge_page_mask(h);
3871	bool new_page = false;
3872
3873	/*
3874	 * Currently, we are forced to kill the process in the event the
3875	 * original mapper has unmapped pages from the child due to a failed
3876	 * COW. Warn that such a situation has occurred as it may not be obvious
3877	 */
3878	if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3879		pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3880			   current->pid);
 
3881		return ret;
3882	}
3883
 
 
 
3884	/*
3885	 * Use page lock to guard against racing truncation
3886	 * before we get page_table_lock.
3887	 */
3888retry:
3889	page = find_lock_page(mapping, idx);
3890	if (!page) {
3891		size = i_size_read(mapping->host) >> huge_page_shift(h);
3892		if (idx >= size)
3893			goto out;
3894
3895		/*
3896		 * Check for page in userfault range
3897		 */
3898		if (userfaultfd_missing(vma)) {
3899			u32 hash;
3900			struct vm_fault vmf = {
3901				.vma = vma,
3902				.address = haddr,
3903				.flags = flags,
3904				/*
3905				 * Hard to debug if it ends up being
3906				 * used by a callee that assumes
3907				 * something about the other
3908				 * uninitialized fields... same as in
3909				 * memory.c
3910				 */
3911			};
3912
3913			/*
3914			 * hugetlb_fault_mutex must be dropped before
3915			 * handling userfault.  Reacquire after handling
3916			 * fault to make calling code simpler.
3917			 */
3918			hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
3919			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3920			ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3921			mutex_lock(&hugetlb_fault_mutex_table[hash]);
3922			goto out;
3923		}
3924
3925		page = alloc_huge_page(vma, haddr, 0);
3926		if (IS_ERR(page)) {
3927			/*
3928			 * Returning error will result in faulting task being
3929			 * sent SIGBUS.  The hugetlb fault mutex prevents two
3930			 * tasks from racing to fault in the same page which
3931			 * could result in false unable to allocate errors.
3932			 * Page migration does not take the fault mutex, but
3933			 * does a clear then write of pte's under page table
3934			 * lock.  Page fault code could race with migration,
3935			 * notice the clear pte and try to allocate a page
3936			 * here.  Before returning error, get ptl and make
3937			 * sure there really is no pte entry.
3938			 */
3939			ptl = huge_pte_lock(h, mm, ptep);
3940			if (!huge_pte_none(huge_ptep_get(ptep))) {
3941				ret = 0;
3942				spin_unlock(ptl);
3943				goto out;
3944			}
3945			spin_unlock(ptl);
3946			ret = vmf_error(PTR_ERR(page));
3947			goto out;
3948		}
3949		clear_huge_page(page, address, pages_per_huge_page(h));
3950		__SetPageUptodate(page);
3951		new_page = true;
3952
3953		if (vma->vm_flags & VM_MAYSHARE) {
3954			int err = huge_add_to_page_cache(page, mapping, idx);
 
 
 
3955			if (err) {
3956				put_page(page);
3957				if (err == -EEXIST)
3958					goto retry;
3959				goto out;
3960			}
 
 
 
 
3961		} else {
3962			lock_page(page);
3963			if (unlikely(anon_vma_prepare(vma))) {
3964				ret = VM_FAULT_OOM;
3965				goto backout_unlocked;
3966			}
3967			anon_rmap = 1;
3968		}
3969	} else {
3970		/*
3971		 * If memory error occurs between mmap() and fault, some process
3972		 * don't have hwpoisoned swap entry for errored virtual address.
3973		 * So we need to block hugepage fault by PG_hwpoison bit check.
3974		 */
3975		if (unlikely(PageHWPoison(page))) {
3976			ret = VM_FAULT_HWPOISON |
3977				VM_FAULT_SET_HINDEX(hstate_index(h));
3978			goto backout_unlocked;
3979		}
3980	}
3981
3982	/*
3983	 * If we are going to COW a private mapping later, we examine the
3984	 * pending reservations for this page now. This will ensure that
3985	 * any allocations necessary to record that reservation occur outside
3986	 * the spinlock.
3987	 */
3988	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3989		if (vma_needs_reservation(h, vma, haddr) < 0) {
3990			ret = VM_FAULT_OOM;
3991			goto backout_unlocked;
3992		}
3993		/* Just decrements count, does not deallocate */
3994		vma_end_reservation(h, vma, haddr);
3995	}
3996
3997	ptl = huge_pte_lock(h, mm, ptep);
3998	size = i_size_read(mapping->host) >> huge_page_shift(h);
3999	if (idx >= size)
4000		goto backout;
4001
4002	ret = 0;
4003	if (!huge_pte_none(huge_ptep_get(ptep)))
4004		goto backout;
4005
4006	if (anon_rmap) {
4007		ClearPagePrivate(page);
4008		hugepage_add_new_anon_rmap(page, vma, haddr);
4009	} else
4010		page_dup_rmap(page, true);
4011	new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4012				&& (vma->vm_flags & VM_SHARED)));
4013	set_huge_pte_at(mm, haddr, ptep, new_pte);
4014
4015	hugetlb_count_add(pages_per_huge_page(h), mm);
4016	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4017		/* Optimization, do the COW without a second fault */
4018		ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4019	}
4020
4021	spin_unlock(ptl);
4022
4023	/*
4024	 * Only make newly allocated pages active.  Existing pages found
4025	 * in the pagecache could be !page_huge_active() if they have been
4026	 * isolated for migration.
4027	 */
4028	if (new_page)
4029		set_page_huge_active(page);
4030
4031	unlock_page(page);
4032out:
4033	return ret;
4034
4035backout:
4036	spin_unlock(ptl);
4037backout_unlocked:
4038	unlock_page(page);
4039	restore_reserve_on_error(h, vma, haddr, page);
4040	put_page(page);
4041	goto out;
4042}
4043
4044#ifdef CONFIG_SMP
4045u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
4046			    pgoff_t idx, unsigned long address)
4047{
4048	unsigned long key[2];
4049	u32 hash;
4050
4051	key[0] = (unsigned long) mapping;
4052	key[1] = idx;
4053
4054	hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
4055
4056	return hash & (num_fault_mutexes - 1);
4057}
4058#else
4059/*
4060 * For uniprocesor systems we always use a single mutex, so just
4061 * return 0 and avoid the hashing overhead.
4062 */
4063u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
4064			    pgoff_t idx, unsigned long address)
4065{
4066	return 0;
4067}
4068#endif
4069
4070vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4071			unsigned long address, unsigned int flags)
4072{
4073	pte_t *ptep, entry;
4074	spinlock_t *ptl;
4075	vm_fault_t ret;
4076	u32 hash;
4077	pgoff_t idx;
4078	struct page *page = NULL;
4079	struct page *pagecache_page = NULL;
 
4080	struct hstate *h = hstate_vma(vma);
4081	struct address_space *mapping;
4082	int need_wait_lock = 0;
4083	unsigned long haddr = address & huge_page_mask(h);
4084
4085	ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
 
 
4086	if (ptep) {
4087		entry = huge_ptep_get(ptep);
4088		if (unlikely(is_hugetlb_entry_migration(entry))) {
4089			migration_entry_wait_huge(vma, mm, ptep);
4090			return 0;
4091		} else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4092			return VM_FAULT_HWPOISON_LARGE |
4093				VM_FAULT_SET_HINDEX(hstate_index(h));
4094	} else {
4095		ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4096		if (!ptep)
4097			return VM_FAULT_OOM;
4098	}
4099
4100	mapping = vma->vm_file->f_mapping;
4101	idx = vma_hugecache_offset(h, vma, haddr);
 
4102
4103	/*
4104	 * Serialize hugepage allocation and instantiation, so that we don't
4105	 * get spurious allocation failures if two CPUs race to instantiate
4106	 * the same page in the page cache.
4107	 */
4108	hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
4109	mutex_lock(&hugetlb_fault_mutex_table[hash]);
4110
4111	entry = huge_ptep_get(ptep);
4112	if (huge_pte_none(entry)) {
4113		ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4114		goto out_mutex;
4115	}
4116
4117	ret = 0;
4118
4119	/*
4120	 * entry could be a migration/hwpoison entry at this point, so this
4121	 * check prevents the kernel from going below assuming that we have
4122	 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4123	 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4124	 * handle it.
4125	 */
4126	if (!pte_present(entry))
4127		goto out_mutex;
4128
4129	/*
4130	 * If we are going to COW the mapping later, we examine the pending
4131	 * reservations for this page now. This will ensure that any
4132	 * allocations necessary to record that reservation occur outside the
4133	 * spinlock. For private mappings, we also lookup the pagecache
4134	 * page now as it is used to determine if a reservation has been
4135	 * consumed.
4136	 */
4137	if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4138		if (vma_needs_reservation(h, vma, haddr) < 0) {
4139			ret = VM_FAULT_OOM;
4140			goto out_mutex;
4141		}
4142		/* Just decrements count, does not deallocate */
4143		vma_end_reservation(h, vma, haddr);
4144
4145		if (!(vma->vm_flags & VM_MAYSHARE))
4146			pagecache_page = hugetlbfs_pagecache_page(h,
4147								vma, haddr);
4148	}
4149
4150	ptl = huge_pte_lock(h, mm, ptep);
4151
4152	/* Check for a racing update before calling hugetlb_cow */
4153	if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4154		goto out_ptl;
4155
4156	/*
4157	 * hugetlb_cow() requires page locks of pte_page(entry) and
4158	 * pagecache_page, so here we need take the former one
4159	 * when page != pagecache_page or !pagecache_page.
 
 
4160	 */
4161	page = pte_page(entry);
 
4162	if (page != pagecache_page)
4163		if (!trylock_page(page)) {
4164			need_wait_lock = 1;
4165			goto out_ptl;
4166		}
 
 
4167
4168	get_page(page);
4169
4170	if (flags & FAULT_FLAG_WRITE) {
4171		if (!huge_pte_write(entry)) {
4172			ret = hugetlb_cow(mm, vma, address, ptep,
4173					  pagecache_page, ptl);
4174			goto out_put_page;
4175		}
4176		entry = huge_pte_mkdirty(entry);
4177	}
4178	entry = pte_mkyoung(entry);
4179	if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4180						flags & FAULT_FLAG_WRITE))
4181		update_mmu_cache(vma, haddr, ptep);
4182out_put_page:
4183	if (page != pagecache_page)
4184		unlock_page(page);
4185	put_page(page);
4186out_ptl:
4187	spin_unlock(ptl);
4188
4189	if (pagecache_page) {
4190		unlock_page(pagecache_page);
4191		put_page(pagecache_page);
4192	}
 
 
 
 
4193out_mutex:
4194	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4195	/*
4196	 * Generally it's safe to hold refcount during waiting page lock. But
4197	 * here we just wait to defer the next page fault to avoid busy loop and
4198	 * the page is not used after unlocked before returning from the current
4199	 * page fault. So we are safe from accessing freed page, even if we wait
4200	 * here without taking refcount.
4201	 */
4202	if (need_wait_lock)
4203		wait_on_page_locked(page);
4204	return ret;
4205}
4206
4207/*
4208 * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4209 * modifications for huge pages.
4210 */
4211int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4212			    pte_t *dst_pte,
4213			    struct vm_area_struct *dst_vma,
4214			    unsigned long dst_addr,
4215			    unsigned long src_addr,
4216			    struct page **pagep)
4217{
4218	struct address_space *mapping;
4219	pgoff_t idx;
4220	unsigned long size;
4221	int vm_shared = dst_vma->vm_flags & VM_SHARED;
4222	struct hstate *h = hstate_vma(dst_vma);
4223	pte_t _dst_pte;
4224	spinlock_t *ptl;
4225	int ret;
4226	struct page *page;
4227
4228	if (!*pagep) {
4229		ret = -ENOMEM;
4230		page = alloc_huge_page(dst_vma, dst_addr, 0);
4231		if (IS_ERR(page))
4232			goto out;
4233
4234		ret = copy_huge_page_from_user(page,
4235						(const void __user *) src_addr,
4236						pages_per_huge_page(h), false);
4237
4238		/* fallback to copy_from_user outside mmap_sem */
4239		if (unlikely(ret)) {
4240			ret = -ENOENT;
4241			*pagep = page;
4242			/* don't free the page */
4243			goto out;
4244		}
4245	} else {
4246		page = *pagep;
4247		*pagep = NULL;
4248	}
4249
4250	/*
4251	 * The memory barrier inside __SetPageUptodate makes sure that
4252	 * preceding stores to the page contents become visible before
4253	 * the set_pte_at() write.
4254	 */
4255	__SetPageUptodate(page);
4256
4257	mapping = dst_vma->vm_file->f_mapping;
4258	idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4259
4260	/*
4261	 * If shared, add to page cache
4262	 */
4263	if (vm_shared) {
4264		size = i_size_read(mapping->host) >> huge_page_shift(h);
4265		ret = -EFAULT;
4266		if (idx >= size)
4267			goto out_release_nounlock;
4268
4269		/*
4270		 * Serialization between remove_inode_hugepages() and
4271		 * huge_add_to_page_cache() below happens through the
4272		 * hugetlb_fault_mutex_table that here must be hold by
4273		 * the caller.
4274		 */
4275		ret = huge_add_to_page_cache(page, mapping, idx);
4276		if (ret)
4277			goto out_release_nounlock;
4278	}
4279
4280	ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4281	spin_lock(ptl);
4282
4283	/*
4284	 * Recheck the i_size after holding PT lock to make sure not
4285	 * to leave any page mapped (as page_mapped()) beyond the end
4286	 * of the i_size (remove_inode_hugepages() is strict about
4287	 * enforcing that). If we bail out here, we'll also leave a
4288	 * page in the radix tree in the vm_shared case beyond the end
4289	 * of the i_size, but remove_inode_hugepages() will take care
4290	 * of it as soon as we drop the hugetlb_fault_mutex_table.
4291	 */
4292	size = i_size_read(mapping->host) >> huge_page_shift(h);
4293	ret = -EFAULT;
4294	if (idx >= size)
4295		goto out_release_unlock;
4296
4297	ret = -EEXIST;
4298	if (!huge_pte_none(huge_ptep_get(dst_pte)))
4299		goto out_release_unlock;
4300
4301	if (vm_shared) {
4302		page_dup_rmap(page, true);
4303	} else {
4304		ClearPagePrivate(page);
4305		hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4306	}
4307
4308	_dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4309	if (dst_vma->vm_flags & VM_WRITE)
4310		_dst_pte = huge_pte_mkdirty(_dst_pte);
4311	_dst_pte = pte_mkyoung(_dst_pte);
4312
4313	set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4314
4315	(void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4316					dst_vma->vm_flags & VM_WRITE);
4317	hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4318
4319	/* No need to invalidate - it was non-present before */
4320	update_mmu_cache(dst_vma, dst_addr, dst_pte);
4321
4322	spin_unlock(ptl);
4323	set_page_huge_active(page);
4324	if (vm_shared)
4325		unlock_page(page);
4326	ret = 0;
4327out:
4328	return ret;
4329out_release_unlock:
4330	spin_unlock(ptl);
4331	if (vm_shared)
4332		unlock_page(page);
4333out_release_nounlock:
4334	put_page(page);
4335	goto out;
4336}
4337
4338long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4339			 struct page **pages, struct vm_area_struct **vmas,
4340			 unsigned long *position, unsigned long *nr_pages,
4341			 long i, unsigned int flags, int *nonblocking)
4342{
4343	unsigned long pfn_offset;
4344	unsigned long vaddr = *position;
4345	unsigned long remainder = *nr_pages;
4346	struct hstate *h = hstate_vma(vma);
4347	int err = -EFAULT;
4348
 
4349	while (vaddr < vma->vm_end && remainder) {
4350		pte_t *pte;
4351		spinlock_t *ptl = NULL;
4352		int absent;
4353		struct page *page;
4354
4355		/*
4356		 * If we have a pending SIGKILL, don't keep faulting pages and
4357		 * potentially allocating memory.
4358		 */
4359		if (fatal_signal_pending(current)) {
4360			remainder = 0;
4361			break;
4362		}
4363
4364		/*
4365		 * Some archs (sparc64, sh*) have multiple pte_ts to
4366		 * each hugepage.  We have to make sure we get the
4367		 * first, for the page indexing below to work.
4368		 *
4369		 * Note that page table lock is not held when pte is null.
4370		 */
4371		pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4372				      huge_page_size(h));
4373		if (pte)
4374			ptl = huge_pte_lock(h, mm, pte);
4375		absent = !pte || huge_pte_none(huge_ptep_get(pte));
4376
4377		/*
4378		 * When coredumping, it suits get_dump_page if we just return
4379		 * an error where there's an empty slot with no huge pagecache
4380		 * to back it.  This way, we avoid allocating a hugepage, and
4381		 * the sparse dumpfile avoids allocating disk blocks, but its
4382		 * huge holes still show up with zeroes where they need to be.
4383		 */
4384		if (absent && (flags & FOLL_DUMP) &&
4385		    !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4386			if (pte)
4387				spin_unlock(ptl);
4388			remainder = 0;
4389			break;
4390		}
4391
4392		/*
4393		 * We need call hugetlb_fault for both hugepages under migration
4394		 * (in which case hugetlb_fault waits for the migration,) and
4395		 * hwpoisoned hugepages (in which case we need to prevent the
4396		 * caller from accessing to them.) In order to do this, we use
4397		 * here is_swap_pte instead of is_hugetlb_entry_migration and
4398		 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4399		 * both cases, and because we can't follow correct pages
4400		 * directly from any kind of swap entries.
4401		 */
4402		if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4403		    ((flags & FOLL_WRITE) &&
4404		      !huge_pte_write(huge_ptep_get(pte)))) {
4405			vm_fault_t ret;
4406			unsigned int fault_flags = 0;
4407
4408			if (pte)
4409				spin_unlock(ptl);
4410			if (flags & FOLL_WRITE)
4411				fault_flags |= FAULT_FLAG_WRITE;
4412			if (nonblocking)
4413				fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4414			if (flags & FOLL_NOWAIT)
4415				fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4416					FAULT_FLAG_RETRY_NOWAIT;
4417			if (flags & FOLL_TRIED) {
4418				VM_WARN_ON_ONCE(fault_flags &
4419						FAULT_FLAG_ALLOW_RETRY);
4420				fault_flags |= FAULT_FLAG_TRIED;
4421			}
4422			ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4423			if (ret & VM_FAULT_ERROR) {
4424				err = vm_fault_to_errno(ret, flags);
4425				remainder = 0;
4426				break;
4427			}
4428			if (ret & VM_FAULT_RETRY) {
4429				if (nonblocking &&
4430				    !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4431					*nonblocking = 0;
4432				*nr_pages = 0;
4433				/*
4434				 * VM_FAULT_RETRY must not return an
4435				 * error, it will return zero
4436				 * instead.
4437				 *
4438				 * No need to update "position" as the
4439				 * caller will not check it after
4440				 * *nr_pages is set to 0.
4441				 */
4442				return i;
4443			}
4444			continue;
4445		}
4446
4447		pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4448		page = pte_page(huge_ptep_get(pte));
4449
4450		/*
4451		 * Instead of doing 'try_get_page()' below in the same_page
4452		 * loop, just check the count once here.
4453		 */
4454		if (unlikely(page_count(page) <= 0)) {
4455			if (pages) {
4456				spin_unlock(ptl);
4457				remainder = 0;
4458				err = -ENOMEM;
4459				break;
4460			}
4461		}
4462same_page:
4463		if (pages) {
4464			pages[i] = mem_map_offset(page, pfn_offset);
4465			get_page(pages[i]);
4466		}
4467
4468		if (vmas)
4469			vmas[i] = vma;
4470
4471		vaddr += PAGE_SIZE;
4472		++pfn_offset;
4473		--remainder;
4474		++i;
4475		if (vaddr < vma->vm_end && remainder &&
4476				pfn_offset < pages_per_huge_page(h)) {
4477			/*
4478			 * We use pfn_offset to avoid touching the pageframes
4479			 * of this compound page.
4480			 */
4481			goto same_page;
4482		}
4483		spin_unlock(ptl);
4484	}
4485	*nr_pages = remainder;
4486	/*
4487	 * setting position is actually required only if remainder is
4488	 * not zero but it's faster not to add a "if (remainder)"
4489	 * branch.
4490	 */
4491	*position = vaddr;
4492
4493	return i ? i : err;
4494}
4495
4496#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4497/*
4498 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4499 * implement this.
4500 */
4501#define flush_hugetlb_tlb_range(vma, addr, end)	flush_tlb_range(vma, addr, end)
4502#endif
4503
4504unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4505		unsigned long address, unsigned long end, pgprot_t newprot)
4506{
4507	struct mm_struct *mm = vma->vm_mm;
4508	unsigned long start = address;
4509	pte_t *ptep;
4510	pte_t pte;
4511	struct hstate *h = hstate_vma(vma);
4512	unsigned long pages = 0;
4513	bool shared_pmd = false;
4514	struct mmu_notifier_range range;
4515
4516	/*
4517	 * In the case of shared PMDs, the area to flush could be beyond
4518	 * start/end.  Set range.start/range.end to cover the maximum possible
4519	 * range if PMD sharing is possible.
4520	 */
4521	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4522				0, vma, mm, start, end);
4523	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4524
4525	BUG_ON(address >= end);
4526	flush_cache_range(vma, range.start, range.end);
4527
4528	mmu_notifier_invalidate_range_start(&range);
4529	i_mmap_lock_write(vma->vm_file->f_mapping);
4530	for (; address < end; address += huge_page_size(h)) {
4531		spinlock_t *ptl;
4532		ptep = huge_pte_offset(mm, address, huge_page_size(h));
4533		if (!ptep)
4534			continue;
4535		ptl = huge_pte_lock(h, mm, ptep);
4536		if (huge_pmd_unshare(mm, &address, ptep)) {
4537			pages++;
4538			spin_unlock(ptl);
4539			shared_pmd = true;
4540			continue;
4541		}
4542		pte = huge_ptep_get(ptep);
4543		if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4544			spin_unlock(ptl);
4545			continue;
 
 
 
 
4546		}
4547		if (unlikely(is_hugetlb_entry_migration(pte))) {
4548			swp_entry_t entry = pte_to_swp_entry(pte);
4549
4550			if (is_write_migration_entry(entry)) {
4551				pte_t newpte;
4552
4553				make_migration_entry_read(&entry);
4554				newpte = swp_entry_to_pte(entry);
4555				set_huge_swap_pte_at(mm, address, ptep,
4556						     newpte, huge_page_size(h));
4557				pages++;
4558			}
4559			spin_unlock(ptl);
4560			continue;
4561		}
4562		if (!huge_pte_none(pte)) {
4563			pte_t old_pte;
4564
4565			old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4566			pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4567			pte = arch_make_huge_pte(pte, vma, NULL, 0);
4568			huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4569			pages++;
4570		}
4571		spin_unlock(ptl);
4572	}
 
4573	/*
4574	 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4575	 * may have cleared our pud entry and done put_page on the page table:
4576	 * once we release i_mmap_rwsem, another task can do the final put_page
4577	 * and that page table be reused and filled with junk.  If we actually
4578	 * did unshare a page of pmds, flush the range corresponding to the pud.
4579	 */
4580	if (shared_pmd)
4581		flush_hugetlb_tlb_range(vma, range.start, range.end);
4582	else
4583		flush_hugetlb_tlb_range(vma, start, end);
4584	/*
4585	 * No need to call mmu_notifier_invalidate_range() we are downgrading
4586	 * page table protection not changing it to point to a new page.
4587	 *
4588	 * See Documentation/vm/mmu_notifier.rst
4589	 */
4590	i_mmap_unlock_write(vma->vm_file->f_mapping);
4591	mmu_notifier_invalidate_range_end(&range);
4592
4593	return pages << h->order;
4594}
4595
4596int hugetlb_reserve_pages(struct inode *inode,
4597					long from, long to,
4598					struct vm_area_struct *vma,
4599					vm_flags_t vm_flags)
4600{
4601	long ret, chg;
4602	struct hstate *h = hstate_inode(inode);
4603	struct hugepage_subpool *spool = subpool_inode(inode);
4604	struct resv_map *resv_map;
4605	long gbl_reserve;
4606
4607	/* This should never happen */
4608	if (from > to) {
4609		VM_WARN(1, "%s called with a negative range\n", __func__);
4610		return -EINVAL;
4611	}
4612
4613	/*
4614	 * Only apply hugepage reservation if asked. At fault time, an
4615	 * attempt will be made for VM_NORESERVE to allocate a page
4616	 * without using reserves
4617	 */
4618	if (vm_flags & VM_NORESERVE)
4619		return 0;
4620
4621	/*
4622	 * Shared mappings base their reservation on the number of pages that
4623	 * are already allocated on behalf of the file. Private mappings need
4624	 * to reserve the full area even if read-only as mprotect() may be
4625	 * called to make the mapping read-write. Assume !vma is a shm mapping
4626	 */
4627	if (!vma || vma->vm_flags & VM_MAYSHARE) {
4628		/*
4629		 * resv_map can not be NULL as hugetlb_reserve_pages is only
4630		 * called for inodes for which resv_maps were created (see
4631		 * hugetlbfs_get_inode).
4632		 */
4633		resv_map = inode_resv_map(inode);
4634
4635		chg = region_chg(resv_map, from, to);
4636
4637	} else {
4638		resv_map = resv_map_alloc();
4639		if (!resv_map)
4640			return -ENOMEM;
4641
4642		chg = to - from;
4643
4644		set_vma_resv_map(vma, resv_map);
4645		set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4646	}
4647
4648	if (chg < 0) {
4649		ret = chg;
4650		goto out_err;
4651	}
4652
4653	/*
4654	 * There must be enough pages in the subpool for the mapping. If
4655	 * the subpool has a minimum size, there may be some global
4656	 * reservations already in place (gbl_reserve).
4657	 */
4658	gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4659	if (gbl_reserve < 0) {
4660		ret = -ENOSPC;
4661		goto out_err;
4662	}
4663
4664	/*
4665	 * Check enough hugepages are available for the reservation.
4666	 * Hand the pages back to the subpool if there are not
4667	 */
4668	ret = hugetlb_acct_memory(h, gbl_reserve);
4669	if (ret < 0) {
4670		/* put back original number of pages, chg */
4671		(void)hugepage_subpool_put_pages(spool, chg);
4672		goto out_err;
4673	}
4674
4675	/*
4676	 * Account for the reservations made. Shared mappings record regions
4677	 * that have reservations as they are shared by multiple VMAs.
4678	 * When the last VMA disappears, the region map says how much
4679	 * the reservation was and the page cache tells how much of
4680	 * the reservation was consumed. Private mappings are per-VMA and
4681	 * only the consumed reservations are tracked. When the VMA
4682	 * disappears, the original reservation is the VMA size and the
4683	 * consumed reservations are stored in the map. Hence, nothing
4684	 * else has to be done for private mappings here
4685	 */
4686	if (!vma || vma->vm_flags & VM_MAYSHARE) {
4687		long add = region_add(resv_map, from, to);
4688
4689		if (unlikely(chg > add)) {
4690			/*
4691			 * pages in this range were added to the reserve
4692			 * map between region_chg and region_add.  This
4693			 * indicates a race with alloc_huge_page.  Adjust
4694			 * the subpool and reserve counts modified above
4695			 * based on the difference.
4696			 */
4697			long rsv_adjust;
4698
4699			rsv_adjust = hugepage_subpool_put_pages(spool,
4700								chg - add);
4701			hugetlb_acct_memory(h, -rsv_adjust);
4702		}
4703	}
4704	return 0;
4705out_err:
4706	if (!vma || vma->vm_flags & VM_MAYSHARE)
4707		/* Don't call region_abort if region_chg failed */
4708		if (chg >= 0)
4709			region_abort(resv_map, from, to);
4710	if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4711		kref_put(&resv_map->refs, resv_map_release);
4712	return ret;
4713}
4714
4715long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4716								long freed)
4717{
4718	struct hstate *h = hstate_inode(inode);
4719	struct resv_map *resv_map = inode_resv_map(inode);
4720	long chg = 0;
4721	struct hugepage_subpool *spool = subpool_inode(inode);
4722	long gbl_reserve;
4723
4724	/*
4725	 * Since this routine can be called in the evict inode path for all
4726	 * hugetlbfs inodes, resv_map could be NULL.
4727	 */
4728	if (resv_map) {
4729		chg = region_del(resv_map, start, end);
4730		/*
4731		 * region_del() can fail in the rare case where a region
4732		 * must be split and another region descriptor can not be
4733		 * allocated.  If end == LONG_MAX, it will not fail.
4734		 */
4735		if (chg < 0)
4736			return chg;
4737	}
4738
4739	spin_lock(&inode->i_lock);
4740	inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4741	spin_unlock(&inode->i_lock);
4742
4743	/*
4744	 * If the subpool has a minimum size, the number of global
4745	 * reservations to be released may be adjusted.
4746	 */
4747	gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4748	hugetlb_acct_memory(h, -gbl_reserve);
4749
4750	return 0;
4751}
4752
4753#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4754static unsigned long page_table_shareable(struct vm_area_struct *svma,
4755				struct vm_area_struct *vma,
4756				unsigned long addr, pgoff_t idx)
4757{
4758	unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4759				svma->vm_start;
4760	unsigned long sbase = saddr & PUD_MASK;
4761	unsigned long s_end = sbase + PUD_SIZE;
4762
4763	/* Allow segments to share if only one is marked locked */
4764	unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4765	unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4766
4767	/*
4768	 * match the virtual addresses, permission and the alignment of the
4769	 * page table page.
4770	 */
4771	if (pmd_index(addr) != pmd_index(saddr) ||
4772	    vm_flags != svm_flags ||
4773	    sbase < svma->vm_start || svma->vm_end < s_end)
4774		return 0;
4775
4776	return saddr;
4777}
4778
4779static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4780{
4781	unsigned long base = addr & PUD_MASK;
4782	unsigned long end = base + PUD_SIZE;
4783
4784	/*
4785	 * check on proper vm_flags and page table alignment
4786	 */
4787	if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4788		return true;
4789	return false;
4790}
4791
4792/*
4793 * Determine if start,end range within vma could be mapped by shared pmd.
4794 * If yes, adjust start and end to cover range associated with possible
4795 * shared pmd mappings.
4796 */
4797void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4798				unsigned long *start, unsigned long *end)
4799{
4800	unsigned long check_addr = *start;
4801
4802	if (!(vma->vm_flags & VM_MAYSHARE))
4803		return;
4804
4805	for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4806		unsigned long a_start = check_addr & PUD_MASK;
4807		unsigned long a_end = a_start + PUD_SIZE;
4808
4809		/*
4810		 * If sharing is possible, adjust start/end if necessary.
4811		 */
4812		if (range_in_vma(vma, a_start, a_end)) {
4813			if (a_start < *start)
4814				*start = a_start;
4815			if (a_end > *end)
4816				*end = a_end;
4817		}
4818	}
4819}
4820
4821/*
4822 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4823 * and returns the corresponding pte. While this is not necessary for the
4824 * !shared pmd case because we can allocate the pmd later as well, it makes the
4825 * code much cleaner. pmd allocation is essential for the shared case because
4826 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4827 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4828 * bad pmd for sharing.
4829 */
4830pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4831{
4832	struct vm_area_struct *vma = find_vma(mm, addr);
4833	struct address_space *mapping = vma->vm_file->f_mapping;
4834	pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4835			vma->vm_pgoff;
4836	struct vm_area_struct *svma;
4837	unsigned long saddr;
4838	pte_t *spte = NULL;
4839	pte_t *pte;
4840	spinlock_t *ptl;
4841
4842	if (!vma_shareable(vma, addr))
4843		return (pte_t *)pmd_alloc(mm, pud, addr);
4844
4845	i_mmap_lock_write(mapping);
4846	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4847		if (svma == vma)
4848			continue;
4849
4850		saddr = page_table_shareable(svma, vma, addr, idx);
4851		if (saddr) {
4852			spte = huge_pte_offset(svma->vm_mm, saddr,
4853					       vma_mmu_pagesize(svma));
4854			if (spte) {
4855				get_page(virt_to_page(spte));
4856				break;
4857			}
4858		}
4859	}
4860
4861	if (!spte)
4862		goto out;
4863
4864	ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4865	if (pud_none(*pud)) {
4866		pud_populate(mm, pud,
4867				(pmd_t *)((unsigned long)spte & PAGE_MASK));
4868		mm_inc_nr_pmds(mm);
4869	} else {
4870		put_page(virt_to_page(spte));
4871	}
4872	spin_unlock(ptl);
4873out:
4874	pte = (pte_t *)pmd_alloc(mm, pud, addr);
4875	i_mmap_unlock_write(mapping);
4876	return pte;
4877}
4878
4879/*
4880 * unmap huge page backed by shared pte.
4881 *
4882 * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4883 * indicated by page_count > 1, unmap is achieved by clearing pud and
4884 * decrementing the ref count. If count == 1, the pte page is not shared.
4885 *
4886 * called with page table lock held.
4887 *
4888 * returns: 1 successfully unmapped a shared pte page
4889 *	    0 the underlying pte page is not shared, or it is the last user
4890 */
4891int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4892{
4893	pgd_t *pgd = pgd_offset(mm, *addr);
4894	p4d_t *p4d = p4d_offset(pgd, *addr);
4895	pud_t *pud = pud_offset(p4d, *addr);
4896
4897	BUG_ON(page_count(virt_to_page(ptep)) == 0);
4898	if (page_count(virt_to_page(ptep)) == 1)
4899		return 0;
4900
4901	pud_clear(pud);
4902	put_page(virt_to_page(ptep));
4903	mm_dec_nr_pmds(mm);
4904	*addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4905	return 1;
4906}
4907#define want_pmd_share()	(1)
4908#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4909pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4910{
4911	return NULL;
4912}
 
 
4913
4914int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4915{
 
4916	return 0;
4917}
4918
4919void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4920				unsigned long *start, unsigned long *end)
4921{
4922}
4923#define want_pmd_share()	(0)
4924#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4925
4926#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4927pte_t *huge_pte_alloc(struct mm_struct *mm,
4928			unsigned long addr, unsigned long sz)
4929{
4930	pgd_t *pgd;
4931	p4d_t *p4d;
4932	pud_t *pud;
4933	pte_t *pte = NULL;
4934
4935	pgd = pgd_offset(mm, addr);
4936	p4d = p4d_alloc(mm, pgd, addr);
4937	if (!p4d)
4938		return NULL;
4939	pud = pud_alloc(mm, p4d, addr);
4940	if (pud) {
4941		if (sz == PUD_SIZE) {
4942			pte = (pte_t *)pud;
4943		} else {
4944			BUG_ON(sz != PMD_SIZE);
4945			if (want_pmd_share() && pud_none(*pud))
4946				pte = huge_pmd_share(mm, addr, pud);
4947			else
4948				pte = (pte_t *)pmd_alloc(mm, pud, addr);
4949		}
4950	}
4951	BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4952
4953	return pte;
4954}
4955
4956/*
4957 * huge_pte_offset() - Walk the page table to resolve the hugepage
4958 * entry at address @addr
4959 *
4960 * Return: Pointer to page table or swap entry (PUD or PMD) for
4961 * address @addr, or NULL if a p*d_none() entry is encountered and the
4962 * size @sz doesn't match the hugepage size at this level of the page
4963 * table.
4964 */
4965pte_t *huge_pte_offset(struct mm_struct *mm,
4966		       unsigned long addr, unsigned long sz)
4967{
4968	pgd_t *pgd;
4969	p4d_t *p4d;
4970	pud_t *pud;
4971	pmd_t *pmd;
4972
4973	pgd = pgd_offset(mm, addr);
4974	if (!pgd_present(*pgd))
4975		return NULL;
4976	p4d = p4d_offset(pgd, addr);
4977	if (!p4d_present(*p4d))
4978		return NULL;
4979
4980	pud = pud_offset(p4d, addr);
4981	if (sz != PUD_SIZE && pud_none(*pud))
4982		return NULL;
4983	/* hugepage or swap? */
4984	if (pud_huge(*pud) || !pud_present(*pud))
4985		return (pte_t *)pud;
4986
4987	pmd = pmd_offset(pud, addr);
4988	if (sz != PMD_SIZE && pmd_none(*pmd))
4989		return NULL;
4990	/* hugepage or swap? */
4991	if (pmd_huge(*pmd) || !pmd_present(*pmd))
4992		return (pte_t *)pmd;
4993
4994	return NULL;
4995}
4996
4997#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4998
4999/*
5000 * These functions are overwritable if your architecture needs its own
5001 * behavior.
5002 */
5003struct page * __weak
5004follow_huge_addr(struct mm_struct *mm, unsigned long address,
5005			      int write)
5006{
5007	return ERR_PTR(-EINVAL);
5008}
 
5009
5010struct page * __weak
5011follow_huge_pd(struct vm_area_struct *vma,
5012	       unsigned long address, hugepd_t hpd, int flags, int pdshift)
5013{
5014	WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5015	return NULL;
5016}
5017
5018struct page * __weak
5019follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5020		pmd_t *pmd, int flags)
5021{
5022	struct page *page = NULL;
5023	spinlock_t *ptl;
5024	pte_t pte;
5025retry:
5026	ptl = pmd_lockptr(mm, pmd);
5027	spin_lock(ptl);
5028	/*
5029	 * make sure that the address range covered by this pmd is not
5030	 * unmapped from other threads.
5031	 */
5032	if (!pmd_huge(*pmd))
5033		goto out;
5034	pte = huge_ptep_get((pte_t *)pmd);
5035	if (pte_present(pte)) {
5036		page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5037		if (flags & FOLL_GET)
5038			get_page(page);
5039	} else {
5040		if (is_hugetlb_entry_migration(pte)) {
5041			spin_unlock(ptl);
5042			__migration_entry_wait(mm, (pte_t *)pmd, ptl);
5043			goto retry;
5044		}
5045		/*
5046		 * hwpoisoned entry is treated as no_page_table in
5047		 * follow_page_mask().
5048		 */
5049	}
5050out:
5051	spin_unlock(ptl);
5052	return page;
5053}
5054
5055struct page * __weak
5056follow_huge_pud(struct mm_struct *mm, unsigned long address,
5057		pud_t *pud, int flags)
5058{
5059	if (flags & FOLL_GET)
5060		return NULL;
5061
5062	return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5063}
5064
5065struct page * __weak
5066follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5067{
5068	if (flags & FOLL_GET)
5069		return NULL;
5070
5071	return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5072}
5073
5074bool isolate_huge_page(struct page *page, struct list_head *list)
5075{
5076	bool ret = true;
5077
5078	VM_BUG_ON_PAGE(!PageHead(page), page);
5079	spin_lock(&hugetlb_lock);
5080	if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5081		ret = false;
5082		goto unlock;
5083	}
5084	clear_page_huge_active(page);
5085	list_move_tail(&page->lru, list);
5086unlock:
5087	spin_unlock(&hugetlb_lock);
5088	return ret;
5089}
5090
5091void putback_active_hugepage(struct page *page)
5092{
5093	VM_BUG_ON_PAGE(!PageHead(page), page);
5094	spin_lock(&hugetlb_lock);
5095	set_page_huge_active(page);
5096	list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5097	spin_unlock(&hugetlb_lock);
5098	put_page(page);
5099}
5100
5101void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5102{
5103	struct hstate *h = page_hstate(oldpage);
5104
5105	hugetlb_cgroup_migrate(oldpage, newpage);
5106	set_page_owner_migrate_reason(newpage, reason);
5107
5108	/*
5109	 * transfer temporary state of the new huge page. This is
5110	 * reverse to other transitions because the newpage is going to
5111	 * be final while the old one will be freed so it takes over
5112	 * the temporary status.
5113	 *
5114	 * Also note that we have to transfer the per-node surplus state
5115	 * here as well otherwise the global surplus count will not match
5116	 * the per-node's.
5117	 */
5118	if (PageHugeTemporary(newpage)) {
5119		int old_nid = page_to_nid(oldpage);
5120		int new_nid = page_to_nid(newpage);
5121
5122		SetPageHugeTemporary(oldpage);
5123		ClearPageHugeTemporary(newpage);
5124
5125		spin_lock(&hugetlb_lock);
5126		if (h->surplus_huge_pages_node[old_nid]) {
5127			h->surplus_huge_pages_node[old_nid]--;
5128			h->surplus_huge_pages_node[new_nid]++;
5129		}
5130		spin_unlock(&hugetlb_lock);
5131	}
5132}
v3.5.6
 
   1/*
   2 * Generic hugetlb support.
   3 * (C) William Irwin, April 2004
   4 */
   5#include <linux/list.h>
   6#include <linux/init.h>
   7#include <linux/module.h>
   8#include <linux/mm.h>
   9#include <linux/seq_file.h>
  10#include <linux/sysctl.h>
  11#include <linux/highmem.h>
  12#include <linux/mmu_notifier.h>
  13#include <linux/nodemask.h>
  14#include <linux/pagemap.h>
  15#include <linux/mempolicy.h>
 
  16#include <linux/cpuset.h>
  17#include <linux/mutex.h>
  18#include <linux/bootmem.h>
  19#include <linux/sysfs.h>
  20#include <linux/slab.h>
 
 
  21#include <linux/rmap.h>
 
  22#include <linux/swap.h>
  23#include <linux/swapops.h>
 
 
  24
  25#include <asm/page.h>
  26#include <asm/pgtable.h>
 
 
  27#include <linux/io.h>
  28
  29#include <linux/hugetlb.h>
 
  30#include <linux/node.h>
 
 
  31#include "internal.h"
  32
  33const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  34static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
  35unsigned long hugepages_treat_as_movable;
  36
  37static int max_hstate;
  38unsigned int default_hstate_idx;
  39struct hstate hstates[HUGE_MAX_HSTATE];
 
 
 
 
 
  40
  41__initdata LIST_HEAD(huge_boot_pages);
  42
  43/* for command line parsing */
  44static struct hstate * __initdata parsed_hstate;
  45static unsigned long __initdata default_hstate_max_huge_pages;
  46static unsigned long __initdata default_hstate_size;
 
  47
  48#define for_each_hstate(h) \
  49	for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
 
 
 
  50
  51/*
  52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
 
  53 */
  54static DEFINE_SPINLOCK(hugetlb_lock);
 
 
 
 
  55
  56static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
  57{
  58	bool free = (spool->count == 0) && (spool->used_hpages == 0);
  59
  60	spin_unlock(&spool->lock);
  61
  62	/* If no pages are used, and no other handles to the subpool
  63	 * remain, free the subpool the subpool remain */
  64	if (free)
 
 
 
 
  65		kfree(spool);
 
  66}
  67
  68struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
 
  69{
  70	struct hugepage_subpool *spool;
  71
  72	spool = kmalloc(sizeof(*spool), GFP_KERNEL);
  73	if (!spool)
  74		return NULL;
  75
  76	spin_lock_init(&spool->lock);
  77	spool->count = 1;
  78	spool->max_hpages = nr_blocks;
  79	spool->used_hpages = 0;
 
 
 
 
 
 
 
  80
  81	return spool;
  82}
  83
  84void hugepage_put_subpool(struct hugepage_subpool *spool)
  85{
  86	spin_lock(&spool->lock);
  87	BUG_ON(!spool->count);
  88	spool->count--;
  89	unlock_or_release_subpool(spool);
  90}
  91
  92static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
 
 
 
 
 
 
 
 
  93				      long delta)
  94{
  95	int ret = 0;
  96
  97	if (!spool)
  98		return 0;
  99
 100	spin_lock(&spool->lock);
 101	if ((spool->used_hpages + delta) <= spool->max_hpages) {
 102		spool->used_hpages += delta;
 103	} else {
 104		ret = -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 105	}
 
 
 106	spin_unlock(&spool->lock);
 107
 108	return ret;
 109}
 110
 111static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
 
 
 
 
 
 
 112				       long delta)
 113{
 
 
 114	if (!spool)
 115		return;
 116
 117	spin_lock(&spool->lock);
 118	spool->used_hpages -= delta;
 119	/* If hugetlbfs_put_super couldn't free spool due to
 120	* an outstanding quota reference, free it now. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 121	unlock_or_release_subpool(spool);
 
 
 122}
 123
 124static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
 125{
 126	return HUGETLBFS_SB(inode->i_sb)->spool;
 127}
 128
 129static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
 130{
 131	return subpool_inode(vma->vm_file->f_dentry->d_inode);
 132}
 133
 134/*
 135 * Region tracking -- allows tracking of reservations and instantiated pages
 136 *                    across the pages in a mapping.
 137 *
 138 * The region data structures are protected by a combination of the mmap_sem
 139 * and the hugetlb_instantion_mutex.  To access or modify a region the caller
 140 * must either hold the mmap_sem for write, or the mmap_sem for read and
 141 * the hugetlb_instantiation mutex:
 
 
 142 *
 143 *	down_write(&mm->mmap_sem);
 144 * or
 145 *	down_read(&mm->mmap_sem);
 146 *	mutex_lock(&hugetlb_instantiation_mutex);
 
 
 
 147 */
 148struct file_region {
 149	struct list_head link;
 150	long from;
 151	long to;
 152};
 153
 154static long region_add(struct list_head *head, long f, long t)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 155{
 
 156	struct file_region *rg, *nrg, *trg;
 
 157
 
 158	/* Locate the region we are either in or before. */
 159	list_for_each_entry(rg, head, link)
 160		if (f <= rg->to)
 161			break;
 162
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 163	/* Round our left edge to the current segment if it encloses us. */
 164	if (f > rg->from)
 165		f = rg->from;
 166
 167	/* Check for and consume any regions we now overlap with. */
 168	nrg = rg;
 169	list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 170		if (&rg->link == head)
 171			break;
 172		if (rg->from > t)
 173			break;
 174
 175		/* If this area reaches higher then extend our area to
 176		 * include it completely.  If this is not the first area
 177		 * which we intend to reuse, free it. */
 178		if (rg->to > t)
 179			t = rg->to;
 180		if (rg != nrg) {
 
 
 
 
 
 181			list_del(&rg->link);
 182			kfree(rg);
 183		}
 184	}
 
 
 185	nrg->from = f;
 
 186	nrg->to = t;
 187	return 0;
 
 
 
 
 
 188}
 189
 190static long region_chg(struct list_head *head, long f, long t)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 191{
 192	struct file_region *rg, *nrg;
 
 193	long chg = 0;
 194
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 195	/* Locate the region we are before or in. */
 196	list_for_each_entry(rg, head, link)
 197		if (f <= rg->to)
 198			break;
 199
 200	/* If we are below the current region then a new region is required.
 201	 * Subtle, allocate a new region at the position but make it zero
 202	 * size such that we can guarantee to record the reservation. */
 203	if (&rg->link == head || t < rg->from) {
 204		nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
 205		if (!nrg)
 206			return -ENOMEM;
 207		nrg->from = f;
 208		nrg->to   = f;
 209		INIT_LIST_HEAD(&nrg->link);
 
 
 
 
 
 
 
 210		list_add(&nrg->link, rg->link.prev);
 211
 212		return t - f;
 213	}
 214
 215	/* Round our left edge to the current segment if it encloses us. */
 216	if (f > rg->from)
 217		f = rg->from;
 218	chg = t - f;
 219
 220	/* Check for and consume any regions we now overlap with. */
 221	list_for_each_entry(rg, rg->link.prev, link) {
 222		if (&rg->link == head)
 223			break;
 224		if (rg->from > t)
 225			return chg;
 226
 227		/* We overlap with this area, if it extends further than
 228		 * us then we must extend ourselves.  Account for its
 229		 * existing reservation. */
 230		if (rg->to > t) {
 231			chg += rg->to - t;
 232			t = rg->to;
 233		}
 234		chg -= rg->to - rg->from;
 235	}
 
 
 
 
 
 
 
 
 236	return chg;
 237}
 238
 239static long region_truncate(struct list_head *head, long end)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 240{
 
 241	struct file_region *rg, *trg;
 242	long chg = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 243
 244	/* Locate the region we are either in or before. */
 245	list_for_each_entry(rg, head, link)
 246		if (end <= rg->to)
 247			break;
 248	if (&rg->link == head)
 249		return 0;
 
 
 
 
 
 
 250
 251	/* If we are in the middle of a region then adjust it. */
 252	if (end > rg->from) {
 253		chg = rg->to - end;
 254		rg->to = end;
 255		rg = list_entry(rg->link.next, typeof(*rg), link);
 
 
 256	}
 257
 258	/* Drop any remaining regions. */
 259	list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
 260		if (&rg->link == head)
 261			break;
 262		chg += rg->to - rg->from;
 263		list_del(&rg->link);
 264		kfree(rg);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 265	}
 266	return chg;
 267}
 268
 269static long region_count(struct list_head *head, long f, long t)
 
 
 
 
 270{
 
 271	struct file_region *rg;
 272	long chg = 0;
 273
 
 274	/* Locate each segment we overlap with, and count that overlap. */
 275	list_for_each_entry(rg, head, link) {
 276		long seg_from;
 277		long seg_to;
 278
 279		if (rg->to <= f)
 280			continue;
 281		if (rg->from >= t)
 282			break;
 283
 284		seg_from = max(rg->from, f);
 285		seg_to = min(rg->to, t);
 286
 287		chg += seg_to - seg_from;
 288	}
 
 289
 290	return chg;
 291}
 292
 293/*
 294 * Convert the address within this vma to the page offset within
 295 * the mapping, in pagecache page units; huge pages here.
 296 */
 297static pgoff_t vma_hugecache_offset(struct hstate *h,
 298			struct vm_area_struct *vma, unsigned long address)
 299{
 300	return ((address - vma->vm_start) >> huge_page_shift(h)) +
 301			(vma->vm_pgoff >> huge_page_order(h));
 302}
 303
 304pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
 305				     unsigned long address)
 306{
 307	return vma_hugecache_offset(hstate_vma(vma), vma, address);
 308}
 
 309
 310/*
 311 * Return the size of the pages allocated when backing a VMA. In the majority
 312 * cases this will be same size as used by the page table entries.
 313 */
 314unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
 315{
 316	struct hstate *hstate;
 317
 318	if (!is_vm_hugetlb_page(vma))
 319		return PAGE_SIZE;
 320
 321	hstate = hstate_vma(vma);
 322
 323	return 1UL << (hstate->order + PAGE_SHIFT);
 324}
 325EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
 326
 327/*
 328 * Return the page size being used by the MMU to back a VMA. In the majority
 329 * of cases, the page size used by the kernel matches the MMU size. On
 330 * architectures where it differs, an architecture-specific version of this
 331 * function is required.
 332 */
 333#ifndef vma_mmu_pagesize
 334unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
 335{
 336	return vma_kernel_pagesize(vma);
 337}
 338#endif
 339
 340/*
 341 * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
 342 * bits of the reservation map pointer, which are always clear due to
 343 * alignment.
 344 */
 345#define HPAGE_RESV_OWNER    (1UL << 0)
 346#define HPAGE_RESV_UNMAPPED (1UL << 1)
 347#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
 348
 349/*
 350 * These helpers are used to track how many pages are reserved for
 351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
 352 * is guaranteed to have their future faults succeed.
 353 *
 354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
 355 * the reserve counters are updated with the hugetlb_lock held. It is safe
 356 * to reset the VMA at fork() time as it is not in use yet and there is no
 357 * chance of the global counters getting corrupted as a result of the values.
 358 *
 359 * The private mapping reservation is represented in a subtly different
 360 * manner to a shared mapping.  A shared mapping has a region map associated
 361 * with the underlying file, this region map represents the backing file
 362 * pages which have ever had a reservation assigned which this persists even
 363 * after the page is instantiated.  A private mapping has a region map
 364 * associated with the original mmap which is attached to all VMAs which
 365 * reference it, this region map represents those offsets which have consumed
 366 * reservation ie. where pages have been instantiated.
 367 */
 368static unsigned long get_vma_private_data(struct vm_area_struct *vma)
 369{
 370	return (unsigned long)vma->vm_private_data;
 371}
 372
 373static void set_vma_private_data(struct vm_area_struct *vma,
 374							unsigned long value)
 375{
 376	vma->vm_private_data = (void *)value;
 377}
 378
 379struct resv_map {
 380	struct kref refs;
 381	struct list_head regions;
 382};
 383
 384static struct resv_map *resv_map_alloc(void)
 385{
 386	struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
 387	if (!resv_map)
 
 
 
 
 388		return NULL;
 
 389
 390	kref_init(&resv_map->refs);
 
 391	INIT_LIST_HEAD(&resv_map->regions);
 392
 
 
 
 
 
 
 393	return resv_map;
 394}
 395
 396static void resv_map_release(struct kref *ref)
 397{
 398	struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
 
 
 399
 400	/* Clear out any active regions before we release the map. */
 401	region_truncate(&resv_map->regions, 0);
 
 
 
 
 
 
 
 
 
 402	kfree(resv_map);
 403}
 404
 
 
 
 
 
 
 
 
 
 
 
 
 
 405static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
 406{
 407	VM_BUG_ON(!is_vm_hugetlb_page(vma));
 408	if (!(vma->vm_flags & VM_MAYSHARE))
 
 
 
 
 
 
 409		return (struct resv_map *)(get_vma_private_data(vma) &
 410							~HPAGE_RESV_MASK);
 411	return NULL;
 412}
 413
 414static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
 415{
 416	VM_BUG_ON(!is_vm_hugetlb_page(vma));
 417	VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 418
 419	set_vma_private_data(vma, (get_vma_private_data(vma) &
 420				HPAGE_RESV_MASK) | (unsigned long)map);
 421}
 422
 423static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
 424{
 425	VM_BUG_ON(!is_vm_hugetlb_page(vma));
 426	VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
 427
 428	set_vma_private_data(vma, get_vma_private_data(vma) | flags);
 429}
 430
 431static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
 432{
 433	VM_BUG_ON(!is_vm_hugetlb_page(vma));
 434
 435	return (get_vma_private_data(vma) & flag) != 0;
 436}
 437
 438/* Decrement the reserved pages in the hugepage pool by one */
 439static void decrement_hugepage_resv_vma(struct hstate *h,
 440			struct vm_area_struct *vma)
 441{
 442	if (vma->vm_flags & VM_NORESERVE)
 443		return;
 444
 445	if (vma->vm_flags & VM_MAYSHARE) {
 446		/* Shared mappings always use reserves */
 447		h->resv_huge_pages--;
 448	} else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 449		/*
 450		 * Only the process that called mmap() has reserves for
 451		 * private mappings.
 452		 */
 453		h->resv_huge_pages--;
 454	}
 455}
 456
 457/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
 458void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 459{
 460	VM_BUG_ON(!is_vm_hugetlb_page(vma));
 461	if (!(vma->vm_flags & VM_MAYSHARE))
 462		vma->vm_private_data = (void *)0;
 463}
 464
 465/* Returns true if the VMA has associated reserve pages */
 466static int vma_has_reserves(struct vm_area_struct *vma)
 467{
 468	if (vma->vm_flags & VM_MAYSHARE)
 469		return 1;
 470	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
 471		return 1;
 472	return 0;
 473}
 
 
 
 
 
 
 
 
 
 474
 475static void copy_gigantic_page(struct page *dst, struct page *src)
 476{
 477	int i;
 478	struct hstate *h = page_hstate(src);
 479	struct page *dst_base = dst;
 480	struct page *src_base = src;
 481
 482	for (i = 0; i < pages_per_huge_page(h); ) {
 483		cond_resched();
 484		copy_highpage(dst, src);
 485
 486		i++;
 487		dst = mem_map_next(dst, dst_base, i);
 488		src = mem_map_next(src, src_base, i);
 489	}
 490}
 491
 492void copy_huge_page(struct page *dst, struct page *src)
 493{
 494	int i;
 495	struct hstate *h = page_hstate(src);
 496
 497	if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
 498		copy_gigantic_page(dst, src);
 499		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 500	}
 501
 502	might_sleep();
 503	for (i = 0; i < pages_per_huge_page(h); i++) {
 504		cond_resched();
 505		copy_highpage(dst + i, src + i);
 506	}
 507}
 508
 509static void enqueue_huge_page(struct hstate *h, struct page *page)
 510{
 511	int nid = page_to_nid(page);
 512	list_add(&page->lru, &h->hugepage_freelists[nid]);
 513	h->free_huge_pages++;
 514	h->free_huge_pages_node[nid]++;
 515}
 516
 517static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
 518{
 519	struct page *page;
 520
 521	if (list_empty(&h->hugepage_freelists[nid]))
 
 
 
 
 
 
 
 522		return NULL;
 523	page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
 524	list_del(&page->lru);
 525	set_page_refcounted(page);
 526	h->free_huge_pages--;
 527	h->free_huge_pages_node[nid]--;
 528	return page;
 529}
 530
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 531static struct page *dequeue_huge_page_vma(struct hstate *h,
 532				struct vm_area_struct *vma,
 533				unsigned long address, int avoid_reserve)
 
 534{
 535	struct page *page = NULL;
 536	struct mempolicy *mpol;
 
 537	nodemask_t *nodemask;
 538	struct zonelist *zonelist;
 539	struct zone *zone;
 540	struct zoneref *z;
 541	unsigned int cpuset_mems_cookie;
 542
 543retry_cpuset:
 544	cpuset_mems_cookie = get_mems_allowed();
 545	zonelist = huge_zonelist(vma, address,
 546					htlb_alloc_mask, &mpol, &nodemask);
 547	/*
 548	 * A child process with MAP_PRIVATE mappings created by their parent
 549	 * have no page reserves. This check ensures that reservations are
 550	 * not "stolen". The child may still get SIGKILLed
 551	 */
 552	if (!vma_has_reserves(vma) &&
 553			h->free_huge_pages - h->resv_huge_pages == 0)
 554		goto err;
 555
 556	/* If reserves cannot be used, ensure enough pages are in the pool */
 557	if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
 558		goto err;
 559
 560	for_each_zone_zonelist_nodemask(zone, z, zonelist,
 561						MAX_NR_ZONES - 1, nodemask) {
 562		if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
 563			page = dequeue_huge_page_node(h, zone_to_nid(zone));
 564			if (page) {
 565				if (!avoid_reserve)
 566					decrement_hugepage_resv_vma(h, vma);
 567				break;
 568			}
 569		}
 570	}
 571
 572	mpol_cond_put(mpol);
 573	if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
 574		goto retry_cpuset;
 575	return page;
 576
 577err:
 578	mpol_cond_put(mpol);
 579	return NULL;
 580}
 581
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 582static void update_and_free_page(struct hstate *h, struct page *page)
 583{
 584	int i;
 585
 586	VM_BUG_ON(h->order >= MAX_ORDER);
 
 587
 588	h->nr_huge_pages--;
 589	h->nr_huge_pages_node[page_to_nid(page)]--;
 590	for (i = 0; i < pages_per_huge_page(h); i++) {
 591		page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
 592				1 << PG_referenced | 1 << PG_dirty |
 593				1 << PG_active | 1 << PG_reserved |
 594				1 << PG_private | 1 << PG_writeback);
 595	}
 596	set_compound_page_dtor(page, NULL);
 
 597	set_page_refcounted(page);
 598	arch_release_hugepage(page);
 599	__free_pages(page, huge_page_order(h));
 
 
 
 
 600}
 601
 602struct hstate *size_to_hstate(unsigned long size)
 603{
 604	struct hstate *h;
 605
 606	for_each_hstate(h) {
 607		if (huge_page_size(h) == size)
 608			return h;
 609	}
 610	return NULL;
 611}
 612
 613static void free_huge_page(struct page *page)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 614{
 615	/*
 616	 * Can't pass hstate in here because it is called from the
 617	 * compound page destructor.
 618	 */
 619	struct hstate *h = page_hstate(page);
 620	int nid = page_to_nid(page);
 621	struct hugepage_subpool *spool =
 622		(struct hugepage_subpool *)page_private(page);
 
 
 
 
 623
 624	set_page_private(page, 0);
 625	page->mapping = NULL;
 626	BUG_ON(page_count(page));
 627	BUG_ON(page_mapcount(page));
 628	INIT_LIST_HEAD(&page->lru);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 629
 630	spin_lock(&hugetlb_lock);
 631	if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
 
 
 
 
 
 
 
 
 
 
 
 
 632		update_and_free_page(h, page);
 633		h->surplus_huge_pages--;
 634		h->surplus_huge_pages_node[nid]--;
 635	} else {
 
 636		enqueue_huge_page(h, page);
 637	}
 638	spin_unlock(&hugetlb_lock);
 639	hugepage_subpool_put_pages(spool, 1);
 640}
 641
 642static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
 643{
 644	set_compound_page_dtor(page, free_huge_page);
 
 645	spin_lock(&hugetlb_lock);
 
 646	h->nr_huge_pages++;
 647	h->nr_huge_pages_node[nid]++;
 648	spin_unlock(&hugetlb_lock);
 649	put_page(page); /* free it into the hugepage allocator */
 650}
 651
 652static void prep_compound_gigantic_page(struct page *page, unsigned long order)
 653{
 654	int i;
 655	int nr_pages = 1 << order;
 656	struct page *p = page + 1;
 657
 658	/* we rely on prep_new_huge_page to set the destructor */
 659	set_compound_order(page, order);
 
 660	__SetPageHead(page);
 661	for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
 662		__SetPageTail(p);
 
 
 
 
 
 
 
 
 
 
 
 
 663		set_page_count(p, 0);
 664		p->first_page = page;
 665	}
 
 666}
 667
 
 
 
 
 
 668int PageHuge(struct page *page)
 669{
 670	compound_page_dtor *dtor;
 671
 672	if (!PageCompound(page))
 673		return 0;
 674
 675	page = compound_head(page);
 676	dtor = get_compound_page_dtor(page);
 677
 678	return dtor == free_huge_page;
 679}
 680EXPORT_SYMBOL_GPL(PageHuge);
 681
 682static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
 
 
 
 
 683{
 684	struct page *page;
 
 685
 686	if (h->order >= MAX_ORDER)
 687		return NULL;
 688
 689	page = alloc_pages_exact_node(nid,
 690		htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 691						__GFP_REPEAT|__GFP_NOWARN,
 692		huge_page_order(h));
 693	if (page) {
 694		if (arch_prepare_hugepage(page)) {
 695			__free_pages(page, huge_page_order(h));
 696			return NULL;
 697		}
 698		prep_new_huge_page(h, page, nid);
 699	}
 700
 701	return page;
 702}
 703
 704/*
 705 * common helper functions for hstate_next_node_to_{alloc|free}.
 706 * We may have allocated or freed a huge page based on a different
 707 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
 708 * be outside of *nodes_allowed.  Ensure that we use an allowed
 709 * node for alloc or free.
 710 */
 711static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
 712{
 713	nid = next_node(nid, *nodes_allowed);
 714	if (nid == MAX_NUMNODES)
 715		nid = first_node(*nodes_allowed);
 716	VM_BUG_ON(nid >= MAX_NUMNODES);
 717
 718	return nid;
 719}
 720
 721static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
 
 
 722{
 723	if (!node_isset(nid, *nodes_allowed))
 724		nid = next_node_allowed(nid, nodes_allowed);
 725	return nid;
 726}
 727
 728/*
 729 * returns the previously saved node ["this node"] from which to
 730 * allocate a persistent huge page for the pool and advance the
 731 * next node from which to allocate, handling wrap at end of node
 732 * mask.
 733 */
 734static int hstate_next_node_to_alloc(struct hstate *h,
 735					nodemask_t *nodes_allowed)
 736{
 737	int nid;
 
 
 
 
 
 
 
 
 
 738
 739	VM_BUG_ON(!nodes_allowed);
 
 
 
 
 
 
 740
 741	nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
 742	h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
 
 
 
 
 
 743
 744	return nid;
 745}
 746
 747static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
 
 
 
 
 
 
 748{
 749	struct page *page;
 750	int start_nid;
 751	int next_nid;
 752	int ret = 0;
 753
 754	start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
 755	next_nid = start_nid;
 756
 757	do {
 758		page = alloc_fresh_huge_page_node(h, next_nid);
 759		if (page) {
 760			ret = 1;
 761			break;
 762		}
 763		next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
 764	} while (next_nid != start_nid);
 765
 766	if (ret)
 767		count_vm_event(HTLB_BUDDY_PGALLOC);
 768	else
 769		count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 770
 771	return ret;
 772}
 773
 774/*
 775 * helper for free_pool_huge_page() - return the previously saved
 776 * node ["this node"] from which to free a huge page.  Advance the
 777 * next node id whether or not we find a free huge page to free so
 778 * that the next attempt to free addresses the next node.
 779 */
 780static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
 
 781{
 782	int nid;
 
 
 
 
 
 
 
 
 
 783
 784	VM_BUG_ON(!nodes_allowed);
 
 785
 786	nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
 787	h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
 788
 789	return nid;
 790}
 791
 792/*
 793 * Free huge page from pool from next node to free.
 794 * Attempt to keep persistent huge pages more or less
 795 * balanced over allowed nodes.
 796 * Called with hugetlb_lock locked.
 797 */
 798static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
 799							 bool acct_surplus)
 800{
 801	int start_nid;
 802	int next_nid;
 803	int ret = 0;
 804
 805	start_nid = hstate_next_node_to_free(h, nodes_allowed);
 806	next_nid = start_nid;
 807
 808	do {
 809		/*
 810		 * If we're returning unused surplus pages, only examine
 811		 * nodes with surplus pages.
 812		 */
 813		if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
 814		    !list_empty(&h->hugepage_freelists[next_nid])) {
 815			struct page *page =
 816				list_entry(h->hugepage_freelists[next_nid].next,
 817					  struct page, lru);
 818			list_del(&page->lru);
 819			h->free_huge_pages--;
 820			h->free_huge_pages_node[next_nid]--;
 821			if (acct_surplus) {
 822				h->surplus_huge_pages--;
 823				h->surplus_huge_pages_node[next_nid]--;
 824			}
 825			update_and_free_page(h, page);
 826			ret = 1;
 827			break;
 828		}
 829		next_nid = hstate_next_node_to_free(h, nodes_allowed);
 830	} while (next_nid != start_nid);
 831
 832	return ret;
 833}
 834
 835static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 836{
 
 837	struct page *page;
 838	unsigned int r_nid;
 839
 840	if (h->order >= MAX_ORDER)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 841		return NULL;
 842
 
 843	/*
 844	 * Assume we will successfully allocate the surplus page to
 845	 * prevent racing processes from causing the surplus to exceed
 846	 * overcommit
 847	 *
 848	 * This however introduces a different race, where a process B
 849	 * tries to grow the static hugepage pool while alloc_pages() is
 850	 * called by process A. B will only examine the per-node
 851	 * counters in determining if surplus huge pages can be
 852	 * converted to normal huge pages in adjust_pool_surplus(). A
 853	 * won't be able to increment the per-node counter, until the
 854	 * lock is dropped by B, but B doesn't drop hugetlb_lock until
 855	 * no more huge pages can be converted from surplus to normal
 856	 * state (and doesn't try to convert again). Thus, we have a
 857	 * case where a surplus huge page exists, the pool is grown, and
 858	 * the surplus huge page still exists after, even though it
 859	 * should just have been converted to a normal huge page. This
 860	 * does not leak memory, though, as the hugepage will be freed
 861	 * once it is out of use. It also does not allow the counters to
 862	 * go out of whack in adjust_pool_surplus() as we don't modify
 863	 * the node values until we've gotten the hugepage and only the
 864	 * per-node value is checked there.
 865	 */
 866	spin_lock(&hugetlb_lock);
 867	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
 
 868		spin_unlock(&hugetlb_lock);
 
 869		return NULL;
 870	} else {
 871		h->nr_huge_pages++;
 872		h->surplus_huge_pages++;
 
 873	}
 
 
 874	spin_unlock(&hugetlb_lock);
 875
 876	if (nid == NUMA_NO_NODE)
 877		page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
 878				   __GFP_REPEAT|__GFP_NOWARN,
 879				   huge_page_order(h));
 880	else
 881		page = alloc_pages_exact_node(nid,
 882			htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
 883			__GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
 
 
 884
 885	if (page && arch_prepare_hugepage(page)) {
 886		__free_pages(page, huge_page_order(h));
 887		page = NULL;
 888	}
 889
 890	spin_lock(&hugetlb_lock);
 891	if (page) {
 892		r_nid = page_to_nid(page);
 893		set_compound_page_dtor(page, free_huge_page);
 894		/*
 895		 * We incremented the global counters already
 896		 */
 897		h->nr_huge_pages_node[r_nid]++;
 898		h->surplus_huge_pages_node[r_nid]++;
 899		__count_vm_event(HTLB_BUDDY_PGALLOC);
 900	} else {
 901		h->nr_huge_pages--;
 902		h->surplus_huge_pages--;
 903		__count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 904	}
 905	spin_unlock(&hugetlb_lock);
 906
 907	return page;
 908}
 909
 910/*
 911 * This allocation function is useful in the context where vma is irrelevant.
 912 * E.g. soft-offlining uses this function because it only cares physical
 913 * address of error page.
 914 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 915struct page *alloc_huge_page_node(struct hstate *h, int nid)
 916{
 917	struct page *page;
 
 
 
 
 918
 919	spin_lock(&hugetlb_lock);
 920	page = dequeue_huge_page_node(h, nid);
 
 921	spin_unlock(&hugetlb_lock);
 922
 923	if (!page)
 924		page = alloc_buddy_huge_page(h, nid);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 925
 926	return page;
 927}
 928
 929/*
 930 * Increase the hugetlb pool such that it can accommodate a reservation
 931 * of size 'delta'.
 932 */
 933static int gather_surplus_pages(struct hstate *h, int delta)
 934{
 935	struct list_head surplus_list;
 936	struct page *page, *tmp;
 937	int ret, i;
 938	int needed, allocated;
 939	bool alloc_ok = true;
 940
 941	needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
 942	if (needed <= 0) {
 943		h->resv_huge_pages += delta;
 944		return 0;
 945	}
 946
 947	allocated = 0;
 948	INIT_LIST_HEAD(&surplus_list);
 949
 950	ret = -ENOMEM;
 951retry:
 952	spin_unlock(&hugetlb_lock);
 953	for (i = 0; i < needed; i++) {
 954		page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
 
 955		if (!page) {
 956			alloc_ok = false;
 957			break;
 958		}
 959		list_add(&page->lru, &surplus_list);
 
 960	}
 961	allocated += i;
 962
 963	/*
 964	 * After retaking hugetlb_lock, we need to recalculate 'needed'
 965	 * because either resv_huge_pages or free_huge_pages may have changed.
 966	 */
 967	spin_lock(&hugetlb_lock);
 968	needed = (h->resv_huge_pages + delta) -
 969			(h->free_huge_pages + allocated);
 970	if (needed > 0) {
 971		if (alloc_ok)
 972			goto retry;
 973		/*
 974		 * We were not able to allocate enough pages to
 975		 * satisfy the entire reservation so we free what
 976		 * we've allocated so far.
 977		 */
 978		goto free;
 979	}
 980	/*
 981	 * The surplus_list now contains _at_least_ the number of extra pages
 982	 * needed to accommodate the reservation.  Add the appropriate number
 983	 * of pages to the hugetlb pool and free the extras back to the buddy
 984	 * allocator.  Commit the entire reservation here to prevent another
 985	 * process from stealing the pages as they are added to the pool but
 986	 * before they are reserved.
 987	 */
 988	needed += allocated;
 989	h->resv_huge_pages += delta;
 990	ret = 0;
 991
 992	/* Free the needed pages to the hugetlb pool */
 993	list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
 994		if ((--needed) < 0)
 995			break;
 996		list_del(&page->lru);
 997		/*
 998		 * This page is now managed by the hugetlb allocator and has
 999		 * no users -- drop the buddy allocator's reference.
1000		 */
1001		put_page_testzero(page);
1002		VM_BUG_ON(page_count(page));
1003		enqueue_huge_page(h, page);
1004	}
1005free:
1006	spin_unlock(&hugetlb_lock);
1007
1008	/* Free unnecessary surplus pages to the buddy allocator */
1009	if (!list_empty(&surplus_list)) {
1010		list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1011			list_del(&page->lru);
1012			put_page(page);
1013		}
1014	}
1015	spin_lock(&hugetlb_lock);
1016
1017	return ret;
1018}
1019
1020/*
1021 * When releasing a hugetlb pool reservation, any surplus pages that were
1022 * allocated to satisfy the reservation must be explicitly freed if they were
1023 * never used.
1024 * Called with hugetlb_lock held.
 
 
 
 
 
 
 
 
1025 */
1026static void return_unused_surplus_pages(struct hstate *h,
1027					unsigned long unused_resv_pages)
1028{
1029	unsigned long nr_pages;
1030
1031	/* Uncommit the reservation */
1032	h->resv_huge_pages -= unused_resv_pages;
1033
1034	/* Cannot return gigantic pages currently */
1035	if (h->order >= MAX_ORDER)
1036		return;
1037
 
 
 
 
1038	nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1039
1040	/*
1041	 * We want to release as many surplus pages as possible, spread
1042	 * evenly across all nodes with memory. Iterate across these nodes
1043	 * until we can no longer free unreserved surplus pages. This occurs
1044	 * when the nodes with surplus pages have no free pages.
1045	 * free_pool_huge_page() will balance the the freed pages across the
1046	 * on-line nodes with memory and will handle the hstate accounting.
 
 
 
 
1047	 */
1048	while (nr_pages--) {
1049		if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1050			break;
 
 
 
1051	}
 
 
 
 
1052}
1053
 
1054/*
1055 * Determine if the huge page at addr within the vma has an associated
1056 * reservation.  Where it does not we will need to logically increase
1057 * reservation and actually increase subpool usage before an allocation
1058 * can occur.  Where any new reservation would be required the
1059 * reservation change is prepared, but not committed.  Once the page
1060 * has been allocated from the subpool and instantiated the change should
1061 * be committed via vma_commit_reservation.  No action is required on
1062 * failure.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1063 */
1064static long vma_needs_reservation(struct hstate *h,
1065			struct vm_area_struct *vma, unsigned long addr)
 
 
 
 
 
 
 
1066{
1067	struct address_space *mapping = vma->vm_file->f_mapping;
1068	struct inode *inode = mapping->host;
 
1069
1070	if (vma->vm_flags & VM_MAYSHARE) {
1071		pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1072		return region_chg(&inode->i_mapping->private_list,
1073							idx, idx + 1);
1074
1075	} else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1076		return 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1077
1078	} else  {
1079		long err;
1080		pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1081		struct resv_map *reservations = vma_resv_map(vma);
1082
1083		err = region_chg(&reservations->regions, idx, idx + 1);
1084		if (err < 0)
1085			return err;
1086		return 0;
 
 
 
 
 
 
 
 
 
 
 
1087	}
 
 
1088}
1089static void vma_commit_reservation(struct hstate *h,
 
 
 
 
 
 
 
1090			struct vm_area_struct *vma, unsigned long addr)
1091{
1092	struct address_space *mapping = vma->vm_file->f_mapping;
1093	struct inode *inode = mapping->host;
 
 
 
 
 
 
1094
1095	if (vma->vm_flags & VM_MAYSHARE) {
1096		pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1097		region_add(&inode->i_mapping->private_list, idx, idx + 1);
 
 
1098
1099	} else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1100		pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1101		struct resv_map *reservations = vma_resv_map(vma);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1102
1103		/* Mark this page used in the map. */
1104		region_add(&reservations->regions, idx, idx + 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1105	}
1106}
1107
1108static struct page *alloc_huge_page(struct vm_area_struct *vma,
1109				    unsigned long addr, int avoid_reserve)
1110{
1111	struct hugepage_subpool *spool = subpool_vma(vma);
1112	struct hstate *h = hstate_vma(vma);
1113	struct page *page;
1114	long chg;
 
 
 
 
 
 
 
 
 
 
 
 
 
1115
1116	/*
1117	 * Processes that did not create the mapping will have no
1118	 * reserves and will not have accounted against subpool
1119	 * limit. Check that the subpool limit can be made before
1120	 * satisfying the allocation MAP_NORESERVE mappings may also
1121	 * need pages and subpool limit allocated allocated if no reserve
1122	 * mapping overlaps.
1123	 */
1124	chg = vma_needs_reservation(h, vma, addr);
1125	if (chg < 0)
1126		return ERR_PTR(-VM_FAULT_OOM);
1127	if (chg)
1128		if (hugepage_subpool_get_pages(spool, chg))
1129			return ERR_PTR(-VM_FAULT_SIGBUS);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1130
1131	spin_lock(&hugetlb_lock);
1132	page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1133	spin_unlock(&hugetlb_lock);
1134
 
 
 
1135	if (!page) {
1136		page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1137		if (!page) {
1138			hugepage_subpool_put_pages(spool, chg);
1139			return ERR_PTR(-VM_FAULT_SIGBUS);
 
 
 
1140		}
 
 
 
1141	}
 
 
1142
1143	set_page_private(page, (unsigned long)spool);
1144
1145	vma_commit_reservation(h, vma, addr);
 
 
 
 
 
 
 
 
 
 
 
1146
 
 
 
1147	return page;
 
 
 
 
 
 
 
 
1148}
1149
1150int __weak alloc_bootmem_huge_page(struct hstate *h)
 
 
1151{
1152	struct huge_bootmem_page *m;
1153	int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1154
1155	while (nr_nodes) {
1156		void *addr;
1157
1158		addr = __alloc_bootmem_node_nopanic(
1159				NODE_DATA(hstate_next_node_to_alloc(h,
1160						&node_states[N_HIGH_MEMORY])),
1161				huge_page_size(h), huge_page_size(h), 0);
1162
1163		if (addr) {
1164			/*
1165			 * Use the beginning of the huge page to store the
1166			 * huge_bootmem_page struct (until gather_bootmem
1167			 * puts them into the mem_map).
1168			 */
1169			m = addr;
1170			goto found;
1171		}
1172		nr_nodes--;
1173	}
1174	return 0;
1175
1176found:
1177	BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1178	/* Put them into a private list first because mem_map is not up yet */
 
1179	list_add(&m->list, &huge_boot_pages);
1180	m->hstate = h;
1181	return 1;
1182}
1183
1184static void prep_compound_huge_page(struct page *page, int order)
 
1185{
1186	if (unlikely(order > (MAX_ORDER - 1)))
1187		prep_compound_gigantic_page(page, order);
1188	else
1189		prep_compound_page(page, order);
1190}
1191
1192/* Put bootmem huge pages into the standard lists after mem_map is up */
1193static void __init gather_bootmem_prealloc(void)
1194{
1195	struct huge_bootmem_page *m;
1196
1197	list_for_each_entry(m, &huge_boot_pages, list) {
 
1198		struct hstate *h = m->hstate;
1199		struct page *page;
1200
1201#ifdef CONFIG_HIGHMEM
1202		page = pfn_to_page(m->phys >> PAGE_SHIFT);
1203		free_bootmem_late((unsigned long)m,
1204				  sizeof(struct huge_bootmem_page));
1205#else
1206		page = virt_to_page(m);
1207#endif
1208		__ClearPageReserved(page);
1209		WARN_ON(page_count(page) != 1);
1210		prep_compound_huge_page(page, h->order);
 
1211		prep_new_huge_page(h, page, page_to_nid(page));
 
 
1212		/*
1213		 * If we had gigantic hugepages allocated at boot time, we need
1214		 * to restore the 'stolen' pages to totalram_pages in order to
1215		 * fix confusing memory reports from free(1) and another
1216		 * side-effects, like CommitLimit going negative.
1217		 */
1218		if (h->order > (MAX_ORDER - 1))
1219			totalram_pages += 1 << h->order;
 
1220	}
1221}
1222
1223static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1224{
1225	unsigned long i;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1226
1227	for (i = 0; i < h->max_huge_pages; ++i) {
1228		if (h->order >= MAX_ORDER) {
1229			if (!alloc_bootmem_huge_page(h))
1230				break;
1231		} else if (!alloc_fresh_huge_page(h,
1232					 &node_states[N_HIGH_MEMORY]))
 
1233			break;
 
 
 
 
 
 
 
 
 
1234	}
1235	h->max_huge_pages = i;
 
1236}
1237
1238static void __init hugetlb_init_hstates(void)
1239{
1240	struct hstate *h;
1241
1242	for_each_hstate(h) {
 
 
 
1243		/* oversize hugepages were init'ed in early boot */
1244		if (h->order < MAX_ORDER)
1245			hugetlb_hstate_alloc_pages(h);
1246	}
1247}
1248
1249static char * __init memfmt(char *buf, unsigned long n)
1250{
1251	if (n >= (1UL << 30))
1252		sprintf(buf, "%lu GB", n >> 30);
1253	else if (n >= (1UL << 20))
1254		sprintf(buf, "%lu MB", n >> 20);
1255	else
1256		sprintf(buf, "%lu KB", n >> 10);
1257	return buf;
1258}
1259
1260static void __init report_hugepages(void)
1261{
1262	struct hstate *h;
1263
1264	for_each_hstate(h) {
1265		char buf[32];
1266		printk(KERN_INFO "HugeTLB registered %s page size, "
1267				 "pre-allocated %ld pages\n",
1268			memfmt(buf, huge_page_size(h)),
1269			h->free_huge_pages);
1270	}
1271}
1272
1273#ifdef CONFIG_HIGHMEM
1274static void try_to_free_low(struct hstate *h, unsigned long count,
1275						nodemask_t *nodes_allowed)
1276{
1277	int i;
1278
1279	if (h->order >= MAX_ORDER)
1280		return;
1281
1282	for_each_node_mask(i, *nodes_allowed) {
1283		struct page *page, *next;
1284		struct list_head *freel = &h->hugepage_freelists[i];
1285		list_for_each_entry_safe(page, next, freel, lru) {
1286			if (count >= h->nr_huge_pages)
1287				return;
1288			if (PageHighMem(page))
1289				continue;
1290			list_del(&page->lru);
1291			update_and_free_page(h, page);
1292			h->free_huge_pages--;
1293			h->free_huge_pages_node[page_to_nid(page)]--;
1294		}
1295	}
1296}
1297#else
1298static inline void try_to_free_low(struct hstate *h, unsigned long count,
1299						nodemask_t *nodes_allowed)
1300{
1301}
1302#endif
1303
1304/*
1305 * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1306 * balanced by operating on them in a round-robin fashion.
1307 * Returns 1 if an adjustment was made.
1308 */
1309static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1310				int delta)
1311{
1312	int start_nid, next_nid;
1313	int ret = 0;
1314
1315	VM_BUG_ON(delta != -1 && delta != 1);
1316
1317	if (delta < 0)
1318		start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1319	else
1320		start_nid = hstate_next_node_to_free(h, nodes_allowed);
1321	next_nid = start_nid;
1322
1323	do {
1324		int nid = next_nid;
1325		if (delta < 0)  {
1326			/*
1327			 * To shrink on this node, there must be a surplus page
1328			 */
1329			if (!h->surplus_huge_pages_node[nid]) {
1330				next_nid = hstate_next_node_to_alloc(h,
1331								nodes_allowed);
1332				continue;
1333			}
1334		}
1335		if (delta > 0) {
1336			/*
1337			 * Surplus cannot exceed the total number of pages
1338			 */
1339			if (h->surplus_huge_pages_node[nid] >=
1340						h->nr_huge_pages_node[nid]) {
1341				next_nid = hstate_next_node_to_free(h,
1342								nodes_allowed);
1343				continue;
1344			}
1345		}
 
 
1346
1347		h->surplus_huge_pages += delta;
1348		h->surplus_huge_pages_node[nid] += delta;
1349		ret = 1;
1350		break;
1351	} while (next_nid != start_nid);
1352
1353	return ret;
1354}
1355
1356#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1357static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1358						nodemask_t *nodes_allowed)
1359{
1360	unsigned long min_count, ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1361
1362	if (h->order >= MAX_ORDER)
1363		return h->max_huge_pages;
 
 
 
 
 
 
 
 
 
 
 
 
 
1364
1365	/*
1366	 * Increase the pool size
1367	 * First take pages out of surplus state.  Then make up the
1368	 * remaining difference by allocating fresh huge pages.
1369	 *
1370	 * We might race with alloc_buddy_huge_page() here and be unable
1371	 * to convert a surplus huge page to a normal huge page. That is
1372	 * not critical, though, it just means the overall size of the
1373	 * pool might be one hugepage larger than it needs to be, but
1374	 * within all the constraints specified by the sysctls.
1375	 */
1376	spin_lock(&hugetlb_lock);
1377	while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1378		if (!adjust_pool_surplus(h, nodes_allowed, -1))
1379			break;
1380	}
1381
1382	while (count > persistent_huge_pages(h)) {
1383		/*
1384		 * If this allocation races such that we no longer need the
1385		 * page, free_huge_page will handle it by freeing the page
1386		 * and reducing the surplus.
1387		 */
1388		spin_unlock(&hugetlb_lock);
1389		ret = alloc_fresh_huge_page(h, nodes_allowed);
 
 
 
 
 
1390		spin_lock(&hugetlb_lock);
1391		if (!ret)
1392			goto out;
1393
1394		/* Bail for signals. Probably ctrl-c from user */
1395		if (signal_pending(current))
1396			goto out;
1397	}
1398
1399	/*
1400	 * Decrease the pool size
1401	 * First return free pages to the buddy allocator (being careful
1402	 * to keep enough around to satisfy reservations).  Then place
1403	 * pages into surplus state as needed so the pool will shrink
1404	 * to the desired size as pages become free.
1405	 *
1406	 * By placing pages into the surplus state independent of the
1407	 * overcommit value, we are allowing the surplus pool size to
1408	 * exceed overcommit. There are few sane options here. Since
1409	 * alloc_buddy_huge_page() is checking the global counter,
1410	 * though, we'll note that we're not allowed to exceed surplus
1411	 * and won't grow the pool anywhere else. Not until one of the
1412	 * sysctls are changed, or the surplus pages go out of use.
1413	 */
1414	min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1415	min_count = max(count, min_count);
1416	try_to_free_low(h, min_count, nodes_allowed);
1417	while (min_count < persistent_huge_pages(h)) {
1418		if (!free_pool_huge_page(h, nodes_allowed, 0))
1419			break;
 
1420	}
1421	while (count < persistent_huge_pages(h)) {
1422		if (!adjust_pool_surplus(h, nodes_allowed, 1))
1423			break;
1424	}
1425out:
1426	ret = persistent_huge_pages(h);
1427	spin_unlock(&hugetlb_lock);
1428	return ret;
 
 
 
1429}
1430
1431#define HSTATE_ATTR_RO(_name) \
1432	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1433
1434#define HSTATE_ATTR(_name) \
1435	static struct kobj_attribute _name##_attr = \
1436		__ATTR(_name, 0644, _name##_show, _name##_store)
1437
1438static struct kobject *hugepages_kobj;
1439static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1440
1441static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1442
1443static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1444{
1445	int i;
1446
1447	for (i = 0; i < HUGE_MAX_HSTATE; i++)
1448		if (hstate_kobjs[i] == kobj) {
1449			if (nidp)
1450				*nidp = NUMA_NO_NODE;
1451			return &hstates[i];
1452		}
1453
1454	return kobj_to_node_hstate(kobj, nidp);
1455}
1456
1457static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1458					struct kobj_attribute *attr, char *buf)
1459{
1460	struct hstate *h;
1461	unsigned long nr_huge_pages;
1462	int nid;
1463
1464	h = kobj_to_hstate(kobj, &nid);
1465	if (nid == NUMA_NO_NODE)
1466		nr_huge_pages = h->nr_huge_pages;
1467	else
1468		nr_huge_pages = h->nr_huge_pages_node[nid];
1469
1470	return sprintf(buf, "%lu\n", nr_huge_pages);
1471}
1472
1473static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1474			struct kobject *kobj, struct kobj_attribute *attr,
1475			const char *buf, size_t len)
1476{
1477	int err;
1478	int nid;
1479	unsigned long count;
1480	struct hstate *h;
1481	NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1482
1483	err = strict_strtoul(buf, 10, &count);
1484	if (err)
1485		goto out;
1486
1487	h = kobj_to_hstate(kobj, &nid);
1488	if (h->order >= MAX_ORDER) {
1489		err = -EINVAL;
1490		goto out;
1491	}
1492
1493	if (nid == NUMA_NO_NODE) {
1494		/*
1495		 * global hstate attribute
1496		 */
1497		if (!(obey_mempolicy &&
1498				init_nodemask_of_mempolicy(nodes_allowed))) {
1499			NODEMASK_FREE(nodes_allowed);
1500			nodes_allowed = &node_states[N_HIGH_MEMORY];
1501		}
1502	} else if (nodes_allowed) {
1503		/*
1504		 * per node hstate attribute: adjust count to global,
1505		 * but restrict alloc/free to the specified node.
1506		 */
1507		count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1508		init_nodemask_of_node(nodes_allowed, nid);
1509	} else
1510		nodes_allowed = &node_states[N_HIGH_MEMORY];
 
 
 
 
1511
1512	h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
 
 
 
 
 
 
 
1513
1514	if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1515		NODEMASK_FREE(nodes_allowed);
 
1516
1517	return len;
1518out:
1519	NODEMASK_FREE(nodes_allowed);
1520	return err;
1521}
1522
1523static ssize_t nr_hugepages_show(struct kobject *kobj,
1524				       struct kobj_attribute *attr, char *buf)
1525{
1526	return nr_hugepages_show_common(kobj, attr, buf);
1527}
1528
1529static ssize_t nr_hugepages_store(struct kobject *kobj,
1530	       struct kobj_attribute *attr, const char *buf, size_t len)
1531{
1532	return nr_hugepages_store_common(false, kobj, attr, buf, len);
1533}
1534HSTATE_ATTR(nr_hugepages);
1535
1536#ifdef CONFIG_NUMA
1537
1538/*
1539 * hstate attribute for optionally mempolicy-based constraint on persistent
1540 * huge page alloc/free.
1541 */
1542static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1543				       struct kobj_attribute *attr, char *buf)
1544{
1545	return nr_hugepages_show_common(kobj, attr, buf);
1546}
1547
1548static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1549	       struct kobj_attribute *attr, const char *buf, size_t len)
1550{
1551	return nr_hugepages_store_common(true, kobj, attr, buf, len);
1552}
1553HSTATE_ATTR(nr_hugepages_mempolicy);
1554#endif
1555
1556
1557static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1558					struct kobj_attribute *attr, char *buf)
1559{
1560	struct hstate *h = kobj_to_hstate(kobj, NULL);
1561	return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1562}
1563
1564static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1565		struct kobj_attribute *attr, const char *buf, size_t count)
1566{
1567	int err;
1568	unsigned long input;
1569	struct hstate *h = kobj_to_hstate(kobj, NULL);
1570
1571	if (h->order >= MAX_ORDER)
1572		return -EINVAL;
1573
1574	err = strict_strtoul(buf, 10, &input);
1575	if (err)
1576		return err;
1577
1578	spin_lock(&hugetlb_lock);
1579	h->nr_overcommit_huge_pages = input;
1580	spin_unlock(&hugetlb_lock);
1581
1582	return count;
1583}
1584HSTATE_ATTR(nr_overcommit_hugepages);
1585
1586static ssize_t free_hugepages_show(struct kobject *kobj,
1587					struct kobj_attribute *attr, char *buf)
1588{
1589	struct hstate *h;
1590	unsigned long free_huge_pages;
1591	int nid;
1592
1593	h = kobj_to_hstate(kobj, &nid);
1594	if (nid == NUMA_NO_NODE)
1595		free_huge_pages = h->free_huge_pages;
1596	else
1597		free_huge_pages = h->free_huge_pages_node[nid];
1598
1599	return sprintf(buf, "%lu\n", free_huge_pages);
1600}
1601HSTATE_ATTR_RO(free_hugepages);
1602
1603static ssize_t resv_hugepages_show(struct kobject *kobj,
1604					struct kobj_attribute *attr, char *buf)
1605{
1606	struct hstate *h = kobj_to_hstate(kobj, NULL);
1607	return sprintf(buf, "%lu\n", h->resv_huge_pages);
1608}
1609HSTATE_ATTR_RO(resv_hugepages);
1610
1611static ssize_t surplus_hugepages_show(struct kobject *kobj,
1612					struct kobj_attribute *attr, char *buf)
1613{
1614	struct hstate *h;
1615	unsigned long surplus_huge_pages;
1616	int nid;
1617
1618	h = kobj_to_hstate(kobj, &nid);
1619	if (nid == NUMA_NO_NODE)
1620		surplus_huge_pages = h->surplus_huge_pages;
1621	else
1622		surplus_huge_pages = h->surplus_huge_pages_node[nid];
1623
1624	return sprintf(buf, "%lu\n", surplus_huge_pages);
1625}
1626HSTATE_ATTR_RO(surplus_hugepages);
1627
1628static struct attribute *hstate_attrs[] = {
1629	&nr_hugepages_attr.attr,
1630	&nr_overcommit_hugepages_attr.attr,
1631	&free_hugepages_attr.attr,
1632	&resv_hugepages_attr.attr,
1633	&surplus_hugepages_attr.attr,
1634#ifdef CONFIG_NUMA
1635	&nr_hugepages_mempolicy_attr.attr,
1636#endif
1637	NULL,
1638};
1639
1640static struct attribute_group hstate_attr_group = {
1641	.attrs = hstate_attrs,
1642};
1643
1644static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1645				    struct kobject **hstate_kobjs,
1646				    struct attribute_group *hstate_attr_group)
1647{
1648	int retval;
1649	int hi = h - hstates;
1650
1651	hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1652	if (!hstate_kobjs[hi])
1653		return -ENOMEM;
1654
1655	retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1656	if (retval)
1657		kobject_put(hstate_kobjs[hi]);
1658
1659	return retval;
1660}
1661
1662static void __init hugetlb_sysfs_init(void)
1663{
1664	struct hstate *h;
1665	int err;
1666
1667	hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1668	if (!hugepages_kobj)
1669		return;
1670
1671	for_each_hstate(h) {
1672		err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1673					 hstate_kobjs, &hstate_attr_group);
1674		if (err)
1675			printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1676								h->name);
1677	}
1678}
1679
1680#ifdef CONFIG_NUMA
1681
1682/*
1683 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1684 * with node devices in node_devices[] using a parallel array.  The array
1685 * index of a node device or _hstate == node id.
1686 * This is here to avoid any static dependency of the node device driver, in
1687 * the base kernel, on the hugetlb module.
1688 */
1689struct node_hstate {
1690	struct kobject		*hugepages_kobj;
1691	struct kobject		*hstate_kobjs[HUGE_MAX_HSTATE];
1692};
1693struct node_hstate node_hstates[MAX_NUMNODES];
1694
1695/*
1696 * A subset of global hstate attributes for node devices
1697 */
1698static struct attribute *per_node_hstate_attrs[] = {
1699	&nr_hugepages_attr.attr,
1700	&free_hugepages_attr.attr,
1701	&surplus_hugepages_attr.attr,
1702	NULL,
1703};
1704
1705static struct attribute_group per_node_hstate_attr_group = {
1706	.attrs = per_node_hstate_attrs,
1707};
1708
1709/*
1710 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1711 * Returns node id via non-NULL nidp.
1712 */
1713static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1714{
1715	int nid;
1716
1717	for (nid = 0; nid < nr_node_ids; nid++) {
1718		struct node_hstate *nhs = &node_hstates[nid];
1719		int i;
1720		for (i = 0; i < HUGE_MAX_HSTATE; i++)
1721			if (nhs->hstate_kobjs[i] == kobj) {
1722				if (nidp)
1723					*nidp = nid;
1724				return &hstates[i];
1725			}
1726	}
1727
1728	BUG();
1729	return NULL;
1730}
1731
1732/*
1733 * Unregister hstate attributes from a single node device.
1734 * No-op if no hstate attributes attached.
1735 */
1736void hugetlb_unregister_node(struct node *node)
1737{
1738	struct hstate *h;
1739	struct node_hstate *nhs = &node_hstates[node->dev.id];
1740
1741	if (!nhs->hugepages_kobj)
1742		return;		/* no hstate attributes */
1743
1744	for_each_hstate(h)
1745		if (nhs->hstate_kobjs[h - hstates]) {
1746			kobject_put(nhs->hstate_kobjs[h - hstates]);
1747			nhs->hstate_kobjs[h - hstates] = NULL;
 
1748		}
 
1749
1750	kobject_put(nhs->hugepages_kobj);
1751	nhs->hugepages_kobj = NULL;
1752}
1753
1754/*
1755 * hugetlb module exit:  unregister hstate attributes from node devices
1756 * that have them.
1757 */
1758static void hugetlb_unregister_all_nodes(void)
1759{
1760	int nid;
1761
1762	/*
1763	 * disable node device registrations.
1764	 */
1765	register_hugetlbfs_with_node(NULL, NULL);
1766
1767	/*
1768	 * remove hstate attributes from any nodes that have them.
1769	 */
1770	for (nid = 0; nid < nr_node_ids; nid++)
1771		hugetlb_unregister_node(&node_devices[nid]);
1772}
1773
1774/*
1775 * Register hstate attributes for a single node device.
1776 * No-op if attributes already registered.
1777 */
1778void hugetlb_register_node(struct node *node)
1779{
1780	struct hstate *h;
1781	struct node_hstate *nhs = &node_hstates[node->dev.id];
1782	int err;
1783
1784	if (nhs->hugepages_kobj)
1785		return;		/* already allocated */
1786
1787	nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1788							&node->dev.kobj);
1789	if (!nhs->hugepages_kobj)
1790		return;
1791
1792	for_each_hstate(h) {
1793		err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1794						nhs->hstate_kobjs,
1795						&per_node_hstate_attr_group);
1796		if (err) {
1797			printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1798					" for node %d\n",
1799						h->name, node->dev.id);
1800			hugetlb_unregister_node(node);
1801			break;
1802		}
1803	}
1804}
1805
1806/*
1807 * hugetlb init time:  register hstate attributes for all registered node
1808 * devices of nodes that have memory.  All on-line nodes should have
1809 * registered their associated device by this time.
1810 */
1811static void hugetlb_register_all_nodes(void)
1812{
1813	int nid;
1814
1815	for_each_node_state(nid, N_HIGH_MEMORY) {
1816		struct node *node = &node_devices[nid];
1817		if (node->dev.id == nid)
1818			hugetlb_register_node(node);
1819	}
1820
1821	/*
1822	 * Let the node device driver know we're here so it can
1823	 * [un]register hstate attributes on node hotplug.
1824	 */
1825	register_hugetlbfs_with_node(hugetlb_register_node,
1826				     hugetlb_unregister_node);
1827}
1828#else	/* !CONFIG_NUMA */
1829
1830static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1831{
1832	BUG();
1833	if (nidp)
1834		*nidp = -1;
1835	return NULL;
1836}
1837
1838static void hugetlb_unregister_all_nodes(void) { }
1839
1840static void hugetlb_register_all_nodes(void) { }
1841
1842#endif
1843
1844static void __exit hugetlb_exit(void)
1845{
1846	struct hstate *h;
1847
1848	hugetlb_unregister_all_nodes();
1849
1850	for_each_hstate(h) {
1851		kobject_put(hstate_kobjs[h - hstates]);
1852	}
1853
1854	kobject_put(hugepages_kobj);
1855}
1856module_exit(hugetlb_exit);
1857
1858static int __init hugetlb_init(void)
1859{
1860	/* Some platform decide whether they support huge pages at boot
1861	 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1862	 * there is no such support
1863	 */
1864	if (HPAGE_SHIFT == 0)
1865		return 0;
1866
1867	if (!size_to_hstate(default_hstate_size)) {
 
 
 
 
 
1868		default_hstate_size = HPAGE_SIZE;
1869		if (!size_to_hstate(default_hstate_size))
1870			hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1871	}
1872	default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1873	if (default_hstate_max_huge_pages)
1874		default_hstate.max_huge_pages = default_hstate_max_huge_pages;
 
 
1875
1876	hugetlb_init_hstates();
1877
1878	gather_bootmem_prealloc();
1879
1880	report_hugepages();
1881
1882	hugetlb_sysfs_init();
 
 
1883
1884	hugetlb_register_all_nodes();
 
 
 
 
 
 
 
 
1885
 
 
1886	return 0;
1887}
1888module_init(hugetlb_init);
1889
1890/* Should be called on processing a hugepagesz=... option */
1891void __init hugetlb_add_hstate(unsigned order)
 
 
 
 
 
1892{
1893	struct hstate *h;
1894	unsigned long i;
1895
1896	if (size_to_hstate(PAGE_SIZE << order)) {
1897		printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1898		return;
1899	}
1900	BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1901	BUG_ON(order == 0);
1902	h = &hstates[max_hstate++];
1903	h->order = order;
1904	h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1905	h->nr_huge_pages = 0;
1906	h->free_huge_pages = 0;
1907	for (i = 0; i < MAX_NUMNODES; ++i)
1908		INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1909	h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1910	h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
 
1911	snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1912					huge_page_size(h)/1024);
1913
1914	parsed_hstate = h;
1915}
1916
1917static int __init hugetlb_nrpages_setup(char *s)
1918{
1919	unsigned long *mhp;
1920	static unsigned long *last_mhp;
1921
 
 
 
 
 
 
1922	/*
1923	 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1924	 * so this hugepages= parameter goes to the "default hstate".
1925	 */
1926	if (!max_hstate)
1927		mhp = &default_hstate_max_huge_pages;
1928	else
1929		mhp = &parsed_hstate->max_huge_pages;
1930
1931	if (mhp == last_mhp) {
1932		printk(KERN_WARNING "hugepages= specified twice without "
1933			"interleaving hugepagesz=, ignoring\n");
1934		return 1;
1935	}
1936
1937	if (sscanf(s, "%lu", mhp) <= 0)
1938		*mhp = 0;
1939
1940	/*
1941	 * Global state is always initialized later in hugetlb_init.
1942	 * But we need to allocate >= MAX_ORDER hstates here early to still
1943	 * use the bootmem allocator.
1944	 */
1945	if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1946		hugetlb_hstate_alloc_pages(parsed_hstate);
1947
1948	last_mhp = mhp;
1949
1950	return 1;
1951}
1952__setup("hugepages=", hugetlb_nrpages_setup);
1953
1954static int __init hugetlb_default_setup(char *s)
1955{
1956	default_hstate_size = memparse(s, &s);
1957	return 1;
1958}
1959__setup("default_hugepagesz=", hugetlb_default_setup);
1960
1961static unsigned int cpuset_mems_nr(unsigned int *array)
1962{
1963	int node;
1964	unsigned int nr = 0;
1965
1966	for_each_node_mask(node, cpuset_current_mems_allowed)
1967		nr += array[node];
1968
1969	return nr;
1970}
1971
1972#ifdef CONFIG_SYSCTL
1973static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1974			 struct ctl_table *table, int write,
1975			 void __user *buffer, size_t *length, loff_t *ppos)
1976{
1977	struct hstate *h = &default_hstate;
1978	unsigned long tmp;
1979	int ret;
1980
1981	tmp = h->max_huge_pages;
1982
1983	if (write && h->order >= MAX_ORDER)
1984		return -EINVAL;
1985
1986	table->data = &tmp;
1987	table->maxlen = sizeof(unsigned long);
1988	ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1989	if (ret)
1990		goto out;
1991
1992	if (write) {
1993		NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1994						GFP_KERNEL | __GFP_NORETRY);
1995		if (!(obey_mempolicy &&
1996			       init_nodemask_of_mempolicy(nodes_allowed))) {
1997			NODEMASK_FREE(nodes_allowed);
1998			nodes_allowed = &node_states[N_HIGH_MEMORY];
1999		}
2000		h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2001
2002		if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2003			NODEMASK_FREE(nodes_allowed);
2004	}
2005out:
2006	return ret;
2007}
2008
2009int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2010			  void __user *buffer, size_t *length, loff_t *ppos)
2011{
2012
2013	return hugetlb_sysctl_handler_common(false, table, write,
2014							buffer, length, ppos);
2015}
2016
2017#ifdef CONFIG_NUMA
2018int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2019			  void __user *buffer, size_t *length, loff_t *ppos)
2020{
2021	return hugetlb_sysctl_handler_common(true, table, write,
2022							buffer, length, ppos);
2023}
2024#endif /* CONFIG_NUMA */
2025
2026int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2027			void __user *buffer,
2028			size_t *length, loff_t *ppos)
2029{
2030	proc_dointvec(table, write, buffer, length, ppos);
2031	if (hugepages_treat_as_movable)
2032		htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2033	else
2034		htlb_alloc_mask = GFP_HIGHUSER;
2035	return 0;
2036}
2037
2038int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2039			void __user *buffer,
2040			size_t *length, loff_t *ppos)
2041{
2042	struct hstate *h = &default_hstate;
2043	unsigned long tmp;
2044	int ret;
2045
 
 
 
2046	tmp = h->nr_overcommit_huge_pages;
2047
2048	if (write && h->order >= MAX_ORDER)
2049		return -EINVAL;
2050
2051	table->data = &tmp;
2052	table->maxlen = sizeof(unsigned long);
2053	ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2054	if (ret)
2055		goto out;
2056
2057	if (write) {
2058		spin_lock(&hugetlb_lock);
2059		h->nr_overcommit_huge_pages = tmp;
2060		spin_unlock(&hugetlb_lock);
2061	}
2062out:
2063	return ret;
2064}
2065
2066#endif /* CONFIG_SYSCTL */
2067
2068void hugetlb_report_meminfo(struct seq_file *m)
2069{
2070	struct hstate *h = &default_hstate;
2071	seq_printf(m,
2072			"HugePages_Total:   %5lu\n"
2073			"HugePages_Free:    %5lu\n"
2074			"HugePages_Rsvd:    %5lu\n"
2075			"HugePages_Surp:    %5lu\n"
2076			"Hugepagesize:   %8lu kB\n",
2077			h->nr_huge_pages,
2078			h->free_huge_pages,
2079			h->resv_huge_pages,
2080			h->surplus_huge_pages,
2081			1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2082}
2083
2084int hugetlb_report_node_meminfo(int nid, char *buf)
2085{
2086	struct hstate *h = &default_hstate;
 
 
2087	return sprintf(buf,
2088		"Node %d HugePages_Total: %5u\n"
2089		"Node %d HugePages_Free:  %5u\n"
2090		"Node %d HugePages_Surp:  %5u\n",
2091		nid, h->nr_huge_pages_node[nid],
2092		nid, h->free_huge_pages_node[nid],
2093		nid, h->surplus_huge_pages_node[nid]);
2094}
2095
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2096/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2097unsigned long hugetlb_total_pages(void)
2098{
2099	struct hstate *h = &default_hstate;
2100	return h->nr_huge_pages * pages_per_huge_page(h);
 
 
 
 
2101}
2102
2103static int hugetlb_acct_memory(struct hstate *h, long delta)
2104{
2105	int ret = -ENOMEM;
2106
2107	spin_lock(&hugetlb_lock);
2108	/*
2109	 * When cpuset is configured, it breaks the strict hugetlb page
2110	 * reservation as the accounting is done on a global variable. Such
2111	 * reservation is completely rubbish in the presence of cpuset because
2112	 * the reservation is not checked against page availability for the
2113	 * current cpuset. Application can still potentially OOM'ed by kernel
2114	 * with lack of free htlb page in cpuset that the task is in.
2115	 * Attempt to enforce strict accounting with cpuset is almost
2116	 * impossible (or too ugly) because cpuset is too fluid that
2117	 * task or memory node can be dynamically moved between cpusets.
2118	 *
2119	 * The change of semantics for shared hugetlb mapping with cpuset is
2120	 * undesirable. However, in order to preserve some of the semantics,
2121	 * we fall back to check against current free page availability as
2122	 * a best attempt and hopefully to minimize the impact of changing
2123	 * semantics that cpuset has.
2124	 */
2125	if (delta > 0) {
2126		if (gather_surplus_pages(h, delta) < 0)
2127			goto out;
2128
2129		if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2130			return_unused_surplus_pages(h, delta);
2131			goto out;
2132		}
2133	}
2134
2135	ret = 0;
2136	if (delta < 0)
2137		return_unused_surplus_pages(h, (unsigned long) -delta);
2138
2139out:
2140	spin_unlock(&hugetlb_lock);
2141	return ret;
2142}
2143
2144static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2145{
2146	struct resv_map *reservations = vma_resv_map(vma);
2147
2148	/*
2149	 * This new VMA should share its siblings reservation map if present.
2150	 * The VMA will only ever have a valid reservation map pointer where
2151	 * it is being copied for another still existing VMA.  As that VMA
2152	 * has a reference to the reservation map it cannot disappear until
2153	 * after this open call completes.  It is therefore safe to take a
2154	 * new reference here without additional locking.
2155	 */
2156	if (reservations)
2157		kref_get(&reservations->refs);
2158}
2159
2160static void resv_map_put(struct vm_area_struct *vma)
2161{
2162	struct resv_map *reservations = vma_resv_map(vma);
 
 
 
 
2163
2164	if (!reservations)
2165		return;
2166	kref_put(&reservations->refs, resv_map_release);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2167}
2168
2169static void hugetlb_vm_op_close(struct vm_area_struct *vma)
 
 
 
 
 
 
 
2170{
2171	struct hstate *h = hstate_vma(vma);
2172	struct resv_map *reservations = vma_resv_map(vma);
2173	struct hugepage_subpool *spool = subpool_vma(vma);
2174	unsigned long reserve;
2175	unsigned long start;
2176	unsigned long end;
2177
2178	if (reservations) {
2179		start = vma_hugecache_offset(h, vma, vma->vm_start);
2180		end = vma_hugecache_offset(h, vma, vma->vm_end);
2181
2182		reserve = (end - start) -
2183			region_count(&reservations->regions, start, end);
2184
2185		resv_map_put(vma);
2186
2187		if (reserve) {
2188			hugetlb_acct_memory(h, -reserve);
2189			hugepage_subpool_put_pages(spool, reserve);
2190		}
2191	}
2192}
2193
2194/*
2195 * We cannot handle pagefaults against hugetlb pages at all.  They cause
2196 * handle_mm_fault() to try to instantiate regular-sized pages in the
2197 * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2198 * this far.
2199 */
2200static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2201{
2202	BUG();
2203	return 0;
2204}
2205
 
 
 
 
 
 
 
2206const struct vm_operations_struct hugetlb_vm_ops = {
2207	.fault = hugetlb_vm_op_fault,
2208	.open = hugetlb_vm_op_open,
2209	.close = hugetlb_vm_op_close,
 
 
2210};
2211
2212static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2213				int writable)
2214{
2215	pte_t entry;
2216
2217	if (writable) {
2218		entry =
2219		    pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2220	} else {
2221		entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 
2222	}
2223	entry = pte_mkyoung(entry);
2224	entry = pte_mkhuge(entry);
2225	entry = arch_make_huge_pte(entry, vma, page, writable);
2226
2227	return entry;
2228}
2229
2230static void set_huge_ptep_writable(struct vm_area_struct *vma,
2231				   unsigned long address, pte_t *ptep)
2232{
2233	pte_t entry;
2234
2235	entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2236	if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2237		update_mmu_cache(vma, address, ptep);
2238}
2239
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2240
2241int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2242			    struct vm_area_struct *vma)
2243{
2244	pte_t *src_pte, *dst_pte, entry;
2245	struct page *ptepage;
2246	unsigned long addr;
2247	int cow;
2248	struct hstate *h = hstate_vma(vma);
2249	unsigned long sz = huge_page_size(h);
 
 
2250
2251	cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2252
 
 
 
 
 
 
 
2253	for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2254		src_pte = huge_pte_offset(src, addr);
 
2255		if (!src_pte)
2256			continue;
2257		dst_pte = huge_pte_alloc(dst, addr, sz);
2258		if (!dst_pte)
2259			goto nomem;
 
 
2260
2261		/* If the pagetables are shared don't copy or take references */
2262		if (dst_pte == src_pte)
 
 
 
 
 
 
 
 
 
2263			continue;
2264
2265		spin_lock(&dst->page_table_lock);
2266		spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2267		if (!huge_pte_none(huge_ptep_get(src_pte))) {
2268			if (cow)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2269				huge_ptep_set_wrprotect(src, addr, src_pte);
 
2270			entry = huge_ptep_get(src_pte);
2271			ptepage = pte_page(entry);
2272			get_page(ptepage);
2273			page_dup_rmap(ptepage);
2274			set_huge_pte_at(dst, addr, dst_pte, entry);
 
2275		}
2276		spin_unlock(&src->page_table_lock);
2277		spin_unlock(&dst->page_table_lock);
2278	}
2279	return 0;
2280
2281nomem:
2282	return -ENOMEM;
2283}
2284
2285static int is_hugetlb_entry_migration(pte_t pte)
2286{
2287	swp_entry_t swp;
2288
2289	if (huge_pte_none(pte) || pte_present(pte))
2290		return 0;
2291	swp = pte_to_swp_entry(pte);
2292	if (non_swap_entry(swp) && is_migration_entry(swp))
2293		return 1;
2294	else
2295		return 0;
2296}
2297
2298static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2299{
2300	swp_entry_t swp;
2301
2302	if (huge_pte_none(pte) || pte_present(pte))
2303		return 0;
2304	swp = pte_to_swp_entry(pte);
2305	if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2306		return 1;
2307	else
2308		return 0;
2309}
2310
2311void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2312			    unsigned long end, struct page *ref_page)
 
2313{
2314	struct mm_struct *mm = vma->vm_mm;
2315	unsigned long address;
2316	pte_t *ptep;
2317	pte_t pte;
 
2318	struct page *page;
2319	struct page *tmp;
2320	struct hstate *h = hstate_vma(vma);
2321	unsigned long sz = huge_page_size(h);
2322
2323	/*
2324	 * A page gathering list, protected by per file i_mmap_mutex. The
2325	 * lock is used to avoid list corruption from multiple unmapping
2326	 * of the same page since we are using page->lru.
2327	 */
2328	LIST_HEAD(page_list);
2329
2330	WARN_ON(!is_vm_hugetlb_page(vma));
2331	BUG_ON(start & ~huge_page_mask(h));
2332	BUG_ON(end & ~huge_page_mask(h));
2333
2334	mmu_notifier_invalidate_range_start(mm, start, end);
2335	spin_lock(&mm->page_table_lock);
2336	for (address = start; address < end; address += sz) {
2337		ptep = huge_pte_offset(mm, address);
 
 
 
 
 
 
 
 
 
 
 
 
 
2338		if (!ptep)
2339			continue;
2340
2341		if (huge_pmd_unshare(mm, &address, ptep))
 
 
 
 
 
 
2342			continue;
 
2343
2344		pte = huge_ptep_get(ptep);
2345		if (huge_pte_none(pte))
 
2346			continue;
 
2347
2348		/*
2349		 * HWPoisoned hugepage is already unmapped and dropped reference
 
2350		 */
2351		if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
 
 
2352			continue;
 
2353
2354		page = pte_page(pte);
2355		/*
2356		 * If a reference page is supplied, it is because a specific
2357		 * page is being unmapped, not a range. Ensure the page we
2358		 * are about to unmap is the actual page of interest.
2359		 */
2360		if (ref_page) {
2361			if (page != ref_page)
 
2362				continue;
2363
2364			/*
2365			 * Mark the VMA as having unmapped its page so that
2366			 * future faults in this VMA will fail rather than
2367			 * looking like data was lost
2368			 */
2369			set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2370		}
2371
2372		pte = huge_ptep_get_and_clear(mm, address, ptep);
2373		if (pte_dirty(pte))
 
2374			set_page_dirty(page);
2375		list_add(&page->lru, &page_list);
2376
2377		/* Bail out after unmapping reference page if supplied */
 
 
 
 
 
 
 
2378		if (ref_page)
2379			break;
2380	}
2381	flush_tlb_range(vma, start, end);
2382	spin_unlock(&mm->page_table_lock);
2383	mmu_notifier_invalidate_range_end(mm, start, end);
2384	list_for_each_entry_safe(page, tmp, &page_list, lru) {
2385		page_remove_rmap(page);
2386		list_del(&page->lru);
2387		put_page(page);
2388	}
2389}
2390
2391void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 
2392			  unsigned long end, struct page *ref_page)
2393{
2394	mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2395	__unmap_hugepage_range(vma, start, end, ref_page);
2396	/*
2397	 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2398	 * test will fail on a vma being torn down, and not grab a page table
2399	 * on its way out.  We're lucky that the flag has such an appropriate
2400	 * name, and can in fact be safely cleared here. We could clear it
2401	 * before the __unmap_hugepage_range above, but all that's necessary
2402	 * is to clear it before releasing the i_mmap_mutex below.
2403	 *
2404	 * This works because in the contexts this is called, the VMA is
2405	 * going to be destroyed. It is not vunerable to madvise(DONTNEED)
2406	 * because madvise is not supported on hugetlbfs. The same applies
2407	 * for direct IO. unmap_hugepage_range() is only being called just
2408	 * before free_pgtables() so clearing VM_MAYSHARE will not cause
2409	 * surprises later.
2410	 */
2411	vma->vm_flags &= ~VM_MAYSHARE;
2412	mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2413}
2414
2415/*
2416 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2417 * mappping it owns the reserve page for. The intention is to unmap the page
2418 * from other VMAs and let the children be SIGKILLed if they are faulting the
2419 * same region.
2420 */
2421static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2422				struct page *page, unsigned long address)
2423{
2424	struct hstate *h = hstate_vma(vma);
2425	struct vm_area_struct *iter_vma;
2426	struct address_space *mapping;
2427	struct prio_tree_iter iter;
2428	pgoff_t pgoff;
2429
2430	/*
2431	 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2432	 * from page cache lookup which is in HPAGE_SIZE units.
2433	 */
2434	address = address & huge_page_mask(h);
2435	pgoff = vma_hugecache_offset(h, vma, address);
2436	mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
 
2437
2438	/*
2439	 * Take the mapping lock for the duration of the table walk. As
2440	 * this mapping should be shared between all the VMAs,
2441	 * __unmap_hugepage_range() is called as the lock is already held
2442	 */
2443	mutex_lock(&mapping->i_mmap_mutex);
2444	vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2445		/* Do not unmap the current VMA */
2446		if (iter_vma == vma)
2447			continue;
2448
2449		/*
 
 
 
 
 
 
 
 
2450		 * Unmap the page from other VMAs without their own reserves.
2451		 * They get marked to be SIGKILLed if they fault in these
2452		 * areas. This is because a future no-page fault on this VMA
2453		 * could insert a zeroed page instead of the data existing
2454		 * from the time of fork. This would look like data corruption
2455		 */
2456		if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2457			__unmap_hugepage_range(iter_vma,
2458				address, address + huge_page_size(h),
2459				page);
2460	}
2461	mutex_unlock(&mapping->i_mmap_mutex);
2462
2463	return 1;
2464}
2465
2466/*
2467 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2468 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2469 * cannot race with other handlers or page migration.
2470 * Keep the pte_same checks anyway to make transition from the mutex easier.
2471 */
2472static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2473			unsigned long address, pte_t *ptep, pte_t pte,
2474			struct page *pagecache_page)
2475{
 
2476	struct hstate *h = hstate_vma(vma);
2477	struct page *old_page, *new_page;
2478	int avoidcopy;
2479	int outside_reserve = 0;
 
 
 
2480
 
2481	old_page = pte_page(pte);
2482
2483retry_avoidcopy:
2484	/* If no-one else is actually using this page, avoid the copy
2485	 * and just make the page writable */
2486	avoidcopy = (page_mapcount(old_page) == 1);
2487	if (avoidcopy) {
2488		if (PageAnon(old_page))
2489			page_move_anon_rmap(old_page, vma, address);
2490		set_huge_ptep_writable(vma, address, ptep);
2491		return 0;
2492	}
2493
2494	/*
2495	 * If the process that created a MAP_PRIVATE mapping is about to
2496	 * perform a COW due to a shared page count, attempt to satisfy
2497	 * the allocation without using the existing reserves. The pagecache
2498	 * page is used to determine if the reserve at this address was
2499	 * consumed or not. If reserves were used, a partial faulted mapping
2500	 * at the time of fork() could consume its reserves on COW instead
2501	 * of the full address range.
2502	 */
2503	if (!(vma->vm_flags & VM_MAYSHARE) &&
2504			is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2505			old_page != pagecache_page)
2506		outside_reserve = 1;
2507
2508	page_cache_get(old_page);
2509
2510	/* Drop page_table_lock as buddy allocator may be called */
2511	spin_unlock(&mm->page_table_lock);
2512	new_page = alloc_huge_page(vma, address, outside_reserve);
 
 
 
2513
2514	if (IS_ERR(new_page)) {
2515		page_cache_release(old_page);
2516
2517		/*
2518		 * If a process owning a MAP_PRIVATE mapping fails to COW,
2519		 * it is due to references held by a child and an insufficient
2520		 * huge page pool. To guarantee the original mappers
2521		 * reliability, unmap the page from child processes. The child
2522		 * may get SIGKILLed if it later faults.
2523		 */
2524		if (outside_reserve) {
 
2525			BUG_ON(huge_pte_none(pte));
2526			if (unmap_ref_private(mm, vma, old_page, address)) {
2527				BUG_ON(huge_pte_none(pte));
2528				spin_lock(&mm->page_table_lock);
2529				ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2530				if (likely(pte_same(huge_ptep_get(ptep), pte)))
2531					goto retry_avoidcopy;
2532				/*
2533				 * race occurs while re-acquiring page_table_lock, and
2534				 * our job is done.
2535				 */
2536				return 0;
2537			}
2538			WARN_ON_ONCE(1);
2539		}
2540
2541		/* Caller expects lock to be held */
2542		spin_lock(&mm->page_table_lock);
2543		return -PTR_ERR(new_page);
2544	}
2545
2546	/*
2547	 * When the original hugepage is shared one, it does not have
2548	 * anon_vma prepared.
2549	 */
2550	if (unlikely(anon_vma_prepare(vma))) {
2551		page_cache_release(new_page);
2552		page_cache_release(old_page);
2553		/* Caller expects lock to be held */
2554		spin_lock(&mm->page_table_lock);
2555		return VM_FAULT_OOM;
2556	}
2557
2558	copy_user_huge_page(new_page, old_page, address, vma,
2559			    pages_per_huge_page(h));
2560	__SetPageUptodate(new_page);
2561
 
 
 
 
2562	/*
2563	 * Retake the page_table_lock to check for racing updates
2564	 * before the page tables are altered
2565	 */
2566	spin_lock(&mm->page_table_lock);
2567	ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2568	if (likely(pte_same(huge_ptep_get(ptep), pte))) {
 
 
2569		/* Break COW */
2570		mmu_notifier_invalidate_range_start(mm,
2571			address & huge_page_mask(h),
2572			(address & huge_page_mask(h)) + huge_page_size(h));
2573		huge_ptep_clear_flush(vma, address, ptep);
2574		set_huge_pte_at(mm, address, ptep,
2575				make_huge_pte(vma, new_page, 1));
2576		page_remove_rmap(old_page);
2577		hugepage_add_new_anon_rmap(new_page, vma, address);
 
2578		/* Make the old page be freed below */
2579		new_page = old_page;
2580		mmu_notifier_invalidate_range_end(mm,
2581			address & huge_page_mask(h),
2582			(address & huge_page_mask(h)) + huge_page_size(h));
2583	}
2584	page_cache_release(new_page);
2585	page_cache_release(old_page);
2586	return 0;
 
 
 
 
 
 
 
2587}
2588
2589/* Return the pagecache page at a given address within a VMA */
2590static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2591			struct vm_area_struct *vma, unsigned long address)
2592{
2593	struct address_space *mapping;
2594	pgoff_t idx;
2595
2596	mapping = vma->vm_file->f_mapping;
2597	idx = vma_hugecache_offset(h, vma, address);
2598
2599	return find_lock_page(mapping, idx);
2600}
2601
2602/*
2603 * Return whether there is a pagecache page to back given address within VMA.
2604 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2605 */
2606static bool hugetlbfs_pagecache_present(struct hstate *h,
2607			struct vm_area_struct *vma, unsigned long address)
2608{
2609	struct address_space *mapping;
2610	pgoff_t idx;
2611	struct page *page;
2612
2613	mapping = vma->vm_file->f_mapping;
2614	idx = vma_hugecache_offset(h, vma, address);
2615
2616	page = find_get_page(mapping, idx);
2617	if (page)
2618		put_page(page);
2619	return page != NULL;
2620}
2621
2622static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2623			unsigned long address, pte_t *ptep, unsigned int flags)
2624{
2625	struct hstate *h = hstate_vma(vma);
2626	int ret = VM_FAULT_SIGBUS;
2627	int anon_rmap = 0;
2628	pgoff_t idx;
2629	unsigned long size;
2630	struct page *page;
2631	struct address_space *mapping;
2632	pte_t new_pte;
 
 
 
2633
2634	/*
2635	 * Currently, we are forced to kill the process in the event the
2636	 * original mapper has unmapped pages from the child due to a failed
2637	 * COW. Warn that such a situation has occurred as it may not be obvious
2638	 */
2639	if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2640		printk(KERN_WARNING
2641			"PID %d killed due to inadequate hugepage pool\n",
2642			current->pid);
2643		return ret;
2644	}
2645
2646	mapping = vma->vm_file->f_mapping;
2647	idx = vma_hugecache_offset(h, vma, address);
2648
2649	/*
2650	 * Use page lock to guard against racing truncation
2651	 * before we get page_table_lock.
2652	 */
2653retry:
2654	page = find_lock_page(mapping, idx);
2655	if (!page) {
2656		size = i_size_read(mapping->host) >> huge_page_shift(h);
2657		if (idx >= size)
2658			goto out;
2659		page = alloc_huge_page(vma, address, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2660		if (IS_ERR(page)) {
2661			ret = -PTR_ERR(page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2662			goto out;
2663		}
2664		clear_huge_page(page, address, pages_per_huge_page(h));
2665		__SetPageUptodate(page);
 
2666
2667		if (vma->vm_flags & VM_MAYSHARE) {
2668			int err;
2669			struct inode *inode = mapping->host;
2670
2671			err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2672			if (err) {
2673				put_page(page);
2674				if (err == -EEXIST)
2675					goto retry;
2676				goto out;
2677			}
2678
2679			spin_lock(&inode->i_lock);
2680			inode->i_blocks += blocks_per_huge_page(h);
2681			spin_unlock(&inode->i_lock);
2682		} else {
2683			lock_page(page);
2684			if (unlikely(anon_vma_prepare(vma))) {
2685				ret = VM_FAULT_OOM;
2686				goto backout_unlocked;
2687			}
2688			anon_rmap = 1;
2689		}
2690	} else {
2691		/*
2692		 * If memory error occurs between mmap() and fault, some process
2693		 * don't have hwpoisoned swap entry for errored virtual address.
2694		 * So we need to block hugepage fault by PG_hwpoison bit check.
2695		 */
2696		if (unlikely(PageHWPoison(page))) {
2697			ret = VM_FAULT_HWPOISON |
2698			      VM_FAULT_SET_HINDEX(h - hstates);
2699			goto backout_unlocked;
2700		}
2701	}
2702
2703	/*
2704	 * If we are going to COW a private mapping later, we examine the
2705	 * pending reservations for this page now. This will ensure that
2706	 * any allocations necessary to record that reservation occur outside
2707	 * the spinlock.
2708	 */
2709	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2710		if (vma_needs_reservation(h, vma, address) < 0) {
2711			ret = VM_FAULT_OOM;
2712			goto backout_unlocked;
2713		}
 
 
 
2714
2715	spin_lock(&mm->page_table_lock);
2716	size = i_size_read(mapping->host) >> huge_page_shift(h);
2717	if (idx >= size)
2718		goto backout;
2719
2720	ret = 0;
2721	if (!huge_pte_none(huge_ptep_get(ptep)))
2722		goto backout;
2723
2724	if (anon_rmap)
2725		hugepage_add_new_anon_rmap(page, vma, address);
2726	else
2727		page_dup_rmap(page);
 
2728	new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2729				&& (vma->vm_flags & VM_SHARED)));
2730	set_huge_pte_at(mm, address, ptep, new_pte);
2731
 
2732	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2733		/* Optimization, do the COW without a second fault */
2734		ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2735	}
2736
2737	spin_unlock(&mm->page_table_lock);
 
 
 
 
 
 
 
 
 
2738	unlock_page(page);
2739out:
2740	return ret;
2741
2742backout:
2743	spin_unlock(&mm->page_table_lock);
2744backout_unlocked:
2745	unlock_page(page);
 
2746	put_page(page);
2747	goto out;
2748}
2749
2750int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2751			unsigned long address, unsigned int flags)
2752{
2753	pte_t *ptep;
2754	pte_t entry;
2755	int ret;
 
 
2756	struct page *page = NULL;
2757	struct page *pagecache_page = NULL;
2758	static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2759	struct hstate *h = hstate_vma(vma);
 
 
 
2760
2761	address &= huge_page_mask(h);
2762
2763	ptep = huge_pte_offset(mm, address);
2764	if (ptep) {
2765		entry = huge_ptep_get(ptep);
2766		if (unlikely(is_hugetlb_entry_migration(entry))) {
2767			migration_entry_wait(mm, (pmd_t *)ptep, address);
2768			return 0;
2769		} else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2770			return VM_FAULT_HWPOISON_LARGE |
2771			       VM_FAULT_SET_HINDEX(h - hstates);
 
 
 
 
2772	}
2773
2774	ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2775	if (!ptep)
2776		return VM_FAULT_OOM;
2777
2778	/*
2779	 * Serialize hugepage allocation and instantiation, so that we don't
2780	 * get spurious allocation failures if two CPUs race to instantiate
2781	 * the same page in the page cache.
2782	 */
2783	mutex_lock(&hugetlb_instantiation_mutex);
 
 
2784	entry = huge_ptep_get(ptep);
2785	if (huge_pte_none(entry)) {
2786		ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2787		goto out_mutex;
2788	}
2789
2790	ret = 0;
2791
2792	/*
 
 
 
 
 
 
 
 
 
 
2793	 * If we are going to COW the mapping later, we examine the pending
2794	 * reservations for this page now. This will ensure that any
2795	 * allocations necessary to record that reservation occur outside the
2796	 * spinlock. For private mappings, we also lookup the pagecache
2797	 * page now as it is used to determine if a reservation has been
2798	 * consumed.
2799	 */
2800	if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2801		if (vma_needs_reservation(h, vma, address) < 0) {
2802			ret = VM_FAULT_OOM;
2803			goto out_mutex;
2804		}
 
 
2805
2806		if (!(vma->vm_flags & VM_MAYSHARE))
2807			pagecache_page = hugetlbfs_pagecache_page(h,
2808								vma, address);
2809	}
2810
 
 
 
 
 
 
2811	/*
2812	 * hugetlb_cow() requires page locks of pte_page(entry) and
2813	 * pagecache_page, so here we need take the former one
2814	 * when page != pagecache_page or !pagecache_page.
2815	 * Note that locking order is always pagecache_page -> page,
2816	 * so no worry about deadlock.
2817	 */
2818	page = pte_page(entry);
2819	get_page(page);
2820	if (page != pagecache_page)
2821		lock_page(page);
2822
2823	spin_lock(&mm->page_table_lock);
2824	/* Check for a racing update before calling hugetlb_cow */
2825	if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2826		goto out_page_table_lock;
2827
 
2828
2829	if (flags & FAULT_FLAG_WRITE) {
2830		if (!pte_write(entry)) {
2831			ret = hugetlb_cow(mm, vma, address, ptep, entry,
2832							pagecache_page);
2833			goto out_page_table_lock;
2834		}
2835		entry = pte_mkdirty(entry);
2836	}
2837	entry = pte_mkyoung(entry);
2838	if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2839						flags & FAULT_FLAG_WRITE))
2840		update_mmu_cache(vma, address, ptep);
2841
2842out_page_table_lock:
2843	spin_unlock(&mm->page_table_lock);
 
 
 
2844
2845	if (pagecache_page) {
2846		unlock_page(pagecache_page);
2847		put_page(pagecache_page);
2848	}
2849	if (page != pagecache_page)
2850		unlock_page(page);
2851	put_page(page);
2852
2853out_mutex:
2854	mutex_unlock(&hugetlb_instantiation_mutex);
2855
 
 
 
 
 
 
 
 
2856	return ret;
2857}
2858
2859/* Can be overriden by architectures */
2860__attribute__((weak)) struct page *
2861follow_huge_pud(struct mm_struct *mm, unsigned long address,
2862	       pud_t *pud, int write)
 
 
 
 
 
 
2863{
2864	BUG();
2865	return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2866}
2867
2868int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2869			struct page **pages, struct vm_area_struct **vmas,
2870			unsigned long *position, int *length, int i,
2871			unsigned int flags)
2872{
2873	unsigned long pfn_offset;
2874	unsigned long vaddr = *position;
2875	int remainder = *length;
2876	struct hstate *h = hstate_vma(vma);
 
2877
2878	spin_lock(&mm->page_table_lock);
2879	while (vaddr < vma->vm_end && remainder) {
2880		pte_t *pte;
 
2881		int absent;
2882		struct page *page;
2883
2884		/*
 
 
 
 
 
 
 
 
 
2885		 * Some archs (sparc64, sh*) have multiple pte_ts to
2886		 * each hugepage.  We have to make sure we get the
2887		 * first, for the page indexing below to work.
 
 
2888		 */
2889		pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
 
 
 
2890		absent = !pte || huge_pte_none(huge_ptep_get(pte));
2891
2892		/*
2893		 * When coredumping, it suits get_dump_page if we just return
2894		 * an error where there's an empty slot with no huge pagecache
2895		 * to back it.  This way, we avoid allocating a hugepage, and
2896		 * the sparse dumpfile avoids allocating disk blocks, but its
2897		 * huge holes still show up with zeroes where they need to be.
2898		 */
2899		if (absent && (flags & FOLL_DUMP) &&
2900		    !hugetlbfs_pagecache_present(h, vma, vaddr)) {
 
 
2901			remainder = 0;
2902			break;
2903		}
2904
2905		if (absent ||
2906		    ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2907			int ret;
2908
2909			spin_unlock(&mm->page_table_lock);
2910			ret = hugetlb_fault(mm, vma, vaddr,
2911				(flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2912			spin_lock(&mm->page_table_lock);
2913			if (!(ret & VM_FAULT_ERROR))
2914				continue;
2915
2916			remainder = 0;
2917			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2918		}
2919
2920		pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2921		page = pte_page(huge_ptep_get(pte));
 
 
 
 
 
 
 
 
 
 
 
 
 
2922same_page:
2923		if (pages) {
2924			pages[i] = mem_map_offset(page, pfn_offset);
2925			get_page(pages[i]);
2926		}
2927
2928		if (vmas)
2929			vmas[i] = vma;
2930
2931		vaddr += PAGE_SIZE;
2932		++pfn_offset;
2933		--remainder;
2934		++i;
2935		if (vaddr < vma->vm_end && remainder &&
2936				pfn_offset < pages_per_huge_page(h)) {
2937			/*
2938			 * We use pfn_offset to avoid touching the pageframes
2939			 * of this compound page.
2940			 */
2941			goto same_page;
2942		}
 
2943	}
2944	spin_unlock(&mm->page_table_lock);
2945	*length = remainder;
 
 
 
 
2946	*position = vaddr;
2947
2948	return i ? i : -EFAULT;
2949}
2950
2951void hugetlb_change_protection(struct vm_area_struct *vma,
 
 
 
 
 
 
 
 
2952		unsigned long address, unsigned long end, pgprot_t newprot)
2953{
2954	struct mm_struct *mm = vma->vm_mm;
2955	unsigned long start = address;
2956	pte_t *ptep;
2957	pte_t pte;
2958	struct hstate *h = hstate_vma(vma);
 
 
 
 
 
 
 
 
 
 
 
 
2959
2960	BUG_ON(address >= end);
2961	flush_cache_range(vma, address, end);
2962
2963	mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2964	spin_lock(&mm->page_table_lock);
2965	for (; address < end; address += huge_page_size(h)) {
2966		ptep = huge_pte_offset(mm, address);
 
2967		if (!ptep)
2968			continue;
2969		if (huge_pmd_unshare(mm, &address, ptep))
 
 
 
 
 
 
 
 
 
2970			continue;
2971		if (!huge_pte_none(huge_ptep_get(ptep))) {
2972			pte = huge_ptep_get_and_clear(mm, address, ptep);
2973			pte = pte_mkhuge(pte_modify(pte, newprot));
2974			set_huge_pte_at(mm, address, ptep, pte);
2975		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2976	}
2977	spin_unlock(&mm->page_table_lock);
2978	/*
2979	 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
2980	 * may have cleared our pud entry and done put_page on the page table:
2981	 * once we release i_mmap_mutex, another task can do the final put_page
2982	 * and that page table be reused and filled with junk.
 
2983	 */
2984	flush_tlb_range(vma, start, end);
2985	mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
2986}
2987
2988int hugetlb_reserve_pages(struct inode *inode,
2989					long from, long to,
2990					struct vm_area_struct *vma,
2991					vm_flags_t vm_flags)
2992{
2993	long ret, chg;
2994	struct hstate *h = hstate_inode(inode);
2995	struct hugepage_subpool *spool = subpool_inode(inode);
 
 
 
 
 
 
 
 
2996
2997	/*
2998	 * Only apply hugepage reservation if asked. At fault time, an
2999	 * attempt will be made for VM_NORESERVE to allocate a page
3000	 * without using reserves
3001	 */
3002	if (vm_flags & VM_NORESERVE)
3003		return 0;
3004
3005	/*
3006	 * Shared mappings base their reservation on the number of pages that
3007	 * are already allocated on behalf of the file. Private mappings need
3008	 * to reserve the full area even if read-only as mprotect() may be
3009	 * called to make the mapping read-write. Assume !vma is a shm mapping
3010	 */
3011	if (!vma || vma->vm_flags & VM_MAYSHARE)
3012		chg = region_chg(&inode->i_mapping->private_list, from, to);
3013	else {
3014		struct resv_map *resv_map = resv_map_alloc();
 
 
 
 
 
 
 
 
3015		if (!resv_map)
3016			return -ENOMEM;
3017
3018		chg = to - from;
3019
3020		set_vma_resv_map(vma, resv_map);
3021		set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3022	}
3023
3024	if (chg < 0) {
3025		ret = chg;
3026		goto out_err;
3027	}
3028
3029	/* There must be enough pages in the subpool for the mapping */
3030	if (hugepage_subpool_get_pages(spool, chg)) {
 
 
 
 
 
3031		ret = -ENOSPC;
3032		goto out_err;
3033	}
3034
3035	/*
3036	 * Check enough hugepages are available for the reservation.
3037	 * Hand the pages back to the subpool if there are not
3038	 */
3039	ret = hugetlb_acct_memory(h, chg);
3040	if (ret < 0) {
3041		hugepage_subpool_put_pages(spool, chg);
 
3042		goto out_err;
3043	}
3044
3045	/*
3046	 * Account for the reservations made. Shared mappings record regions
3047	 * that have reservations as they are shared by multiple VMAs.
3048	 * When the last VMA disappears, the region map says how much
3049	 * the reservation was and the page cache tells how much of
3050	 * the reservation was consumed. Private mappings are per-VMA and
3051	 * only the consumed reservations are tracked. When the VMA
3052	 * disappears, the original reservation is the VMA size and the
3053	 * consumed reservations are stored in the map. Hence, nothing
3054	 * else has to be done for private mappings here
3055	 */
3056	if (!vma || vma->vm_flags & VM_MAYSHARE)
3057		region_add(&inode->i_mapping->private_list, from, to);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3058	return 0;
3059out_err:
3060	if (vma)
3061		resv_map_put(vma);
 
 
 
 
3062	return ret;
3063}
3064
3065void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
 
3066{
3067	struct hstate *h = hstate_inode(inode);
3068	long chg = region_truncate(&inode->i_mapping->private_list, offset);
 
3069	struct hugepage_subpool *spool = subpool_inode(inode);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3070
3071	spin_lock(&inode->i_lock);
3072	inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3073	spin_unlock(&inode->i_lock);
3074
3075	hugepage_subpool_put_pages(spool, (chg - freed));
3076	hugetlb_acct_memory(h, -(chg - freed));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3077}
3078
3079#ifdef CONFIG_MEMORY_FAILURE
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3080
3081/* Should be called in hugetlb_lock */
3082static int is_hugepage_on_freelist(struct page *hpage)
 
 
 
 
 
 
 
3083{
3084	struct page *page;
3085	struct page *tmp;
3086	struct hstate *h = page_hstate(hpage);
3087	int nid = page_to_nid(hpage);
3088
3089	list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3090		if (page == hpage)
3091			return 1;
3092	return 0;
3093}
3094
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3095/*
3096 * This function is called from memory failure code.
3097 * Assume the caller holds page lock of the head page.
3098 */
3099int dequeue_hwpoisoned_huge_page(struct page *hpage)
 
 
3100{
3101	struct hstate *h = page_hstate(hpage);
3102	int nid = page_to_nid(hpage);
3103	int ret = -EBUSY;
3104
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3105	spin_lock(&hugetlb_lock);
3106	if (is_hugepage_on_freelist(hpage)) {
3107		list_del(&hpage->lru);
3108		set_page_refcounted(hpage);
3109		h->free_huge_pages--;
3110		h->free_huge_pages_node[nid]--;
3111		ret = 0;
3112	}
3113	spin_unlock(&hugetlb_lock);
3114	return ret;
3115}
3116#endif