Root/mm/hugetlb.c

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
22#include <asm/page.h>
23#include <asm/pgtable.h>
24#include <asm/io.h>
25
26#include <linux/hugetlb.h>
27#include <linux/node.h>
28#include "internal.h"
29
30const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
31static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
32unsigned long hugepages_treat_as_movable;
33
34static int max_hstate;
35unsigned int default_hstate_idx;
36struct hstate hstates[HUGE_MAX_HSTATE];
37
38__initdata LIST_HEAD(huge_boot_pages);
39
40/* for command line parsing */
41static struct hstate * __initdata parsed_hstate;
42static unsigned long __initdata default_hstate_max_huge_pages;
43static unsigned long __initdata default_hstate_size;
44
45#define for_each_hstate(h) \
46    for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
47
48/*
49 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
50 */
51static DEFINE_SPINLOCK(hugetlb_lock);
52
53/*
54 * Region tracking -- allows tracking of reservations and instantiated pages
55 * across the pages in a mapping.
56 *
57 * The region data structures are protected by a combination of the mmap_sem
58 * and the hugetlb_instantion_mutex. To access or modify a region the caller
59 * must either hold the mmap_sem for write, or the mmap_sem for read and
60 * the hugetlb_instantiation mutex:
61 *
62 * down_write(&mm->mmap_sem);
63 * or
64 * down_read(&mm->mmap_sem);
65 * mutex_lock(&hugetlb_instantiation_mutex);
66 */
67struct file_region {
68    struct list_head link;
69    long from;
70    long to;
71};
72
73static long region_add(struct list_head *head, long f, long t)
74{
75    struct file_region *rg, *nrg, *trg;
76
77    /* Locate the region we are either in or before. */
78    list_for_each_entry(rg, head, link)
79        if (f <= rg->to)
80            break;
81
82    /* Round our left edge to the current segment if it encloses us. */
83    if (f > rg->from)
84        f = rg->from;
85
86    /* Check for and consume any regions we now overlap with. */
87    nrg = rg;
88    list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
89        if (&rg->link == head)
90            break;
91        if (rg->from > t)
92            break;
93
94        /* If this area reaches higher then extend our area to
95         * include it completely. If this is not the first area
96         * which we intend to reuse, free it. */
97        if (rg->to > t)
98            t = rg->to;
99        if (rg != nrg) {
100            list_del(&rg->link);
101            kfree(rg);
102        }
103    }
104    nrg->from = f;
105    nrg->to = t;
106    return 0;
107}
108
109static long region_chg(struct list_head *head, long f, long t)
110{
111    struct file_region *rg, *nrg;
112    long chg = 0;
113
114    /* Locate the region we are before or in. */
115    list_for_each_entry(rg, head, link)
116        if (f <= rg->to)
117            break;
118
119    /* If we are below the current region then a new region is required.
120     * Subtle, allocate a new region at the position but make it zero
121     * size such that we can guarantee to record the reservation. */
122    if (&rg->link == head || t < rg->from) {
123        nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
124        if (!nrg)
125            return -ENOMEM;
126        nrg->from = f;
127        nrg->to = f;
128        INIT_LIST_HEAD(&nrg->link);
129        list_add(&nrg->link, rg->link.prev);
130
131        return t - f;
132    }
133
134    /* Round our left edge to the current segment if it encloses us. */
135    if (f > rg->from)
136        f = rg->from;
137    chg = t - f;
138
139    /* Check for and consume any regions we now overlap with. */
140    list_for_each_entry(rg, rg->link.prev, link) {
141        if (&rg->link == head)
142            break;
143        if (rg->from > t)
144            return chg;
145
146        /* We overlap with this area, if it extends futher than
147         * us then we must extend ourselves. Account for its
148         * existing reservation. */
149        if (rg->to > t) {
150            chg += rg->to - t;
151            t = rg->to;
152        }
153        chg -= rg->to - rg->from;
154    }
155    return chg;
156}
157
158static long region_truncate(struct list_head *head, long end)
159{
160    struct file_region *rg, *trg;
161    long chg = 0;
162
163    /* Locate the region we are either in or before. */
164    list_for_each_entry(rg, head, link)
165        if (end <= rg->to)
166            break;
167    if (&rg->link == head)
168        return 0;
169
170    /* If we are in the middle of a region then adjust it. */
171    if (end > rg->from) {
172        chg = rg->to - end;
173        rg->to = end;
174        rg = list_entry(rg->link.next, typeof(*rg), link);
175    }
176
177    /* Drop any remaining regions. */
178    list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
179        if (&rg->link == head)
180            break;
181        chg += rg->to - rg->from;
182        list_del(&rg->link);
183        kfree(rg);
184    }
185    return chg;
186}
187
188static long region_count(struct list_head *head, long f, long t)
189{
190    struct file_region *rg;
191    long chg = 0;
192
193    /* Locate each segment we overlap with, and count that overlap. */
194    list_for_each_entry(rg, head, link) {
195        int seg_from;
196        int seg_to;
197
198        if (rg->to <= f)
199            continue;
200        if (rg->from >= t)
201            break;
202
203        seg_from = max(rg->from, f);
204        seg_to = min(rg->to, t);
205
206        chg += seg_to - seg_from;
207    }
208
209    return chg;
210}
211
212/*
213 * Convert the address within this vma to the page offset within
214 * the mapping, in pagecache page units; huge pages here.
215 */
216static pgoff_t vma_hugecache_offset(struct hstate *h,
217            struct vm_area_struct *vma, unsigned long address)
218{
219    return ((address - vma->vm_start) >> huge_page_shift(h)) +
220            (vma->vm_pgoff >> huge_page_order(h));
221}
222
223/*
224 * Return the size of the pages allocated when backing a VMA. In the majority
225 * cases this will be same size as used by the page table entries.
226 */
227unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
228{
229    struct hstate *hstate;
230
231    if (!is_vm_hugetlb_page(vma))
232        return PAGE_SIZE;
233
234    hstate = hstate_vma(vma);
235
236    return 1UL << (hstate->order + PAGE_SHIFT);
237}
238EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
239
240/*
241 * Return the page size being used by the MMU to back a VMA. In the majority
242 * of cases, the page size used by the kernel matches the MMU size. On
243 * architectures where it differs, an architecture-specific version of this
244 * function is required.
245 */
246#ifndef vma_mmu_pagesize
247unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
248{
249    return vma_kernel_pagesize(vma);
250}
251#endif
252
253/*
254 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
255 * bits of the reservation map pointer, which are always clear due to
256 * alignment.
257 */
258#define HPAGE_RESV_OWNER (1UL << 0)
259#define HPAGE_RESV_UNMAPPED (1UL << 1)
260#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
261
262/*
263 * These helpers are used to track how many pages are reserved for
264 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
265 * is guaranteed to have their future faults succeed.
266 *
267 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
268 * the reserve counters are updated with the hugetlb_lock held. It is safe
269 * to reset the VMA at fork() time as it is not in use yet and there is no
270 * chance of the global counters getting corrupted as a result of the values.
271 *
272 * The private mapping reservation is represented in a subtly different
273 * manner to a shared mapping. A shared mapping has a region map associated
274 * with the underlying file, this region map represents the backing file
275 * pages which have ever had a reservation assigned which this persists even
276 * after the page is instantiated. A private mapping has a region map
277 * associated with the original mmap which is attached to all VMAs which
278 * reference it, this region map represents those offsets which have consumed
279 * reservation ie. where pages have been instantiated.
280 */
281static unsigned long get_vma_private_data(struct vm_area_struct *vma)
282{
283    return (unsigned long)vma->vm_private_data;
284}
285
286static void set_vma_private_data(struct vm_area_struct *vma,
287                            unsigned long value)
288{
289    vma->vm_private_data = (void *)value;
290}
291
292struct resv_map {
293    struct kref refs;
294    struct list_head regions;
295};
296
297static struct resv_map *resv_map_alloc(void)
298{
299    struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
300    if (!resv_map)
301        return NULL;
302
303    kref_init(&resv_map->refs);
304    INIT_LIST_HEAD(&resv_map->regions);
305
306    return resv_map;
307}
308
309static void resv_map_release(struct kref *ref)
310{
311    struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
312
313    /* Clear out any active regions before we release the map. */
314    region_truncate(&resv_map->regions, 0);
315    kfree(resv_map);
316}
317
318static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
319{
320    VM_BUG_ON(!is_vm_hugetlb_page(vma));
321    if (!(vma->vm_flags & VM_MAYSHARE))
322        return (struct resv_map *)(get_vma_private_data(vma) &
323                            ~HPAGE_RESV_MASK);
324    return NULL;
325}
326
327static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
328{
329    VM_BUG_ON(!is_vm_hugetlb_page(vma));
330    VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
331
332    set_vma_private_data(vma, (get_vma_private_data(vma) &
333                HPAGE_RESV_MASK) | (unsigned long)map);
334}
335
336static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
337{
338    VM_BUG_ON(!is_vm_hugetlb_page(vma));
339    VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
340
341    set_vma_private_data(vma, get_vma_private_data(vma) | flags);
342}
343
344static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
345{
346    VM_BUG_ON(!is_vm_hugetlb_page(vma));
347
348    return (get_vma_private_data(vma) & flag) != 0;
349}
350
351/* Decrement the reserved pages in the hugepage pool by one */
352static void decrement_hugepage_resv_vma(struct hstate *h,
353            struct vm_area_struct *vma)
354{
355    if (vma->vm_flags & VM_NORESERVE)
356        return;
357
358    if (vma->vm_flags & VM_MAYSHARE) {
359        /* Shared mappings always use reserves */
360        h->resv_huge_pages--;
361    } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
362        /*
363         * Only the process that called mmap() has reserves for
364         * private mappings.
365         */
366        h->resv_huge_pages--;
367    }
368}
369
370/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
371void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
372{
373    VM_BUG_ON(!is_vm_hugetlb_page(vma));
374    if (!(vma->vm_flags & VM_MAYSHARE))
375        vma->vm_private_data = (void *)0;
376}
377
378/* Returns true if the VMA has associated reserve pages */
379static int vma_has_reserves(struct vm_area_struct *vma)
380{
381    if (vma->vm_flags & VM_MAYSHARE)
382        return 1;
383    if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
384        return 1;
385    return 0;
386}
387
388static void clear_gigantic_page(struct page *page,
389            unsigned long addr, unsigned long sz)
390{
391    int i;
392    struct page *p = page;
393
394    might_sleep();
395    for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
396        cond_resched();
397        clear_user_highpage(p, addr + i * PAGE_SIZE);
398    }
399}
400static void clear_huge_page(struct page *page,
401            unsigned long addr, unsigned long sz)
402{
403    int i;
404
405    if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) {
406        clear_gigantic_page(page, addr, sz);
407        return;
408    }
409
410    might_sleep();
411    for (i = 0; i < sz/PAGE_SIZE; i++) {
412        cond_resched();
413        clear_user_highpage(page + i, addr + i * PAGE_SIZE);
414    }
415}
416
417static void copy_gigantic_page(struct page *dst, struct page *src,
418               unsigned long addr, struct vm_area_struct *vma)
419{
420    int i;
421    struct hstate *h = hstate_vma(vma);
422    struct page *dst_base = dst;
423    struct page *src_base = src;
424    might_sleep();
425    for (i = 0; i < pages_per_huge_page(h); ) {
426        cond_resched();
427        copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
428
429        i++;
430        dst = mem_map_next(dst, dst_base, i);
431        src = mem_map_next(src, src_base, i);
432    }
433}
434static void copy_huge_page(struct page *dst, struct page *src,
435               unsigned long addr, struct vm_area_struct *vma)
436{
437    int i;
438    struct hstate *h = hstate_vma(vma);
439
440    if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
441        copy_gigantic_page(dst, src, addr, vma);
442        return;
443    }
444
445    might_sleep();
446    for (i = 0; i < pages_per_huge_page(h); i++) {
447        cond_resched();
448        copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
449    }
450}
451
452static void enqueue_huge_page(struct hstate *h, struct page *page)
453{
454    int nid = page_to_nid(page);
455    list_add(&page->lru, &h->hugepage_freelists[nid]);
456    h->free_huge_pages++;
457    h->free_huge_pages_node[nid]++;
458}
459
460static struct page *dequeue_huge_page_vma(struct hstate *h,
461                struct vm_area_struct *vma,
462                unsigned long address, int avoid_reserve)
463{
464    int nid;
465    struct page *page = NULL;
466    struct mempolicy *mpol;
467    nodemask_t *nodemask;
468    struct zonelist *zonelist;
469    struct zone *zone;
470    struct zoneref *z;
471
472    get_mems_allowed();
473    zonelist = huge_zonelist(vma, address,
474                    htlb_alloc_mask, &mpol, &nodemask);
475    /*
476     * A child process with MAP_PRIVATE mappings created by their parent
477     * have no page reserves. This check ensures that reservations are
478     * not "stolen". The child may still get SIGKILLed
479     */
480    if (!vma_has_reserves(vma) &&
481            h->free_huge_pages - h->resv_huge_pages == 0)
482        goto err;
483
484    /* If reserves cannot be used, ensure enough pages are in the pool */
485    if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
486        goto err;;
487
488    for_each_zone_zonelist_nodemask(zone, z, zonelist,
489                        MAX_NR_ZONES - 1, nodemask) {
490        nid = zone_to_nid(zone);
491        if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
492            !list_empty(&h->hugepage_freelists[nid])) {
493            page = list_entry(h->hugepage_freelists[nid].next,
494                      struct page, lru);
495            list_del(&page->lru);
496            h->free_huge_pages--;
497            h->free_huge_pages_node[nid]--;
498
499            if (!avoid_reserve)
500                decrement_hugepage_resv_vma(h, vma);
501
502            break;
503        }
504    }
505err:
506    mpol_cond_put(mpol);
507    put_mems_allowed();
508    return page;
509}
510
511static void update_and_free_page(struct hstate *h, struct page *page)
512{
513    int i;
514
515    VM_BUG_ON(h->order >= MAX_ORDER);
516
517    h->nr_huge_pages--;
518    h->nr_huge_pages_node[page_to_nid(page)]--;
519    for (i = 0; i < pages_per_huge_page(h); i++) {
520        page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
521                1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
522                1 << PG_private | 1<< PG_writeback);
523    }
524    set_compound_page_dtor(page, NULL);
525    set_page_refcounted(page);
526    arch_release_hugepage(page);
527    __free_pages(page, huge_page_order(h));
528}
529
530struct hstate *size_to_hstate(unsigned long size)
531{
532    struct hstate *h;
533
534    for_each_hstate(h) {
535        if (huge_page_size(h) == size)
536            return h;
537    }
538    return NULL;
539}
540
541static void free_huge_page(struct page *page)
542{
543    /*
544     * Can't pass hstate in here because it is called from the
545     * compound page destructor.
546     */
547    struct hstate *h = page_hstate(page);
548    int nid = page_to_nid(page);
549    struct address_space *mapping;
550
551    mapping = (struct address_space *) page_private(page);
552    set_page_private(page, 0);
553    page->mapping = NULL;
554    BUG_ON(page_count(page));
555    INIT_LIST_HEAD(&page->lru);
556
557    spin_lock(&hugetlb_lock);
558    if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
559        update_and_free_page(h, page);
560        h->surplus_huge_pages--;
561        h->surplus_huge_pages_node[nid]--;
562    } else {
563        enqueue_huge_page(h, page);
564    }
565    spin_unlock(&hugetlb_lock);
566    if (mapping)
567        hugetlb_put_quota(mapping, 1);
568}
569
570static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
571{
572    set_compound_page_dtor(page, free_huge_page);
573    spin_lock(&hugetlb_lock);
574    h->nr_huge_pages++;
575    h->nr_huge_pages_node[nid]++;
576    spin_unlock(&hugetlb_lock);
577    put_page(page); /* free it into the hugepage allocator */
578}
579
580static void prep_compound_gigantic_page(struct page *page, unsigned long order)
581{
582    int i;
583    int nr_pages = 1 << order;
584    struct page *p = page + 1;
585
586    /* we rely on prep_new_huge_page to set the destructor */
587    set_compound_order(page, order);
588    __SetPageHead(page);
589    for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
590        __SetPageTail(p);
591        p->first_page = page;
592    }
593}
594
595int PageHuge(struct page *page)
596{
597    compound_page_dtor *dtor;
598
599    if (!PageCompound(page))
600        return 0;
601
602    page = compound_head(page);
603    dtor = get_compound_page_dtor(page);
604
605    return dtor == free_huge_page;
606}
607
608static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
609{
610    struct page *page;
611
612    if (h->order >= MAX_ORDER)
613        return NULL;
614
615    page = alloc_pages_exact_node(nid,
616        htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
617                        __GFP_REPEAT|__GFP_NOWARN,
618        huge_page_order(h));
619    if (page) {
620        if (arch_prepare_hugepage(page)) {
621            __free_pages(page, huge_page_order(h));
622            return NULL;
623        }
624        prep_new_huge_page(h, page, nid);
625    }
626
627    return page;
628}
629
630/*
631 * common helper functions for hstate_next_node_to_{alloc|free}.
632 * We may have allocated or freed a huge page based on a different
633 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
634 * be outside of *nodes_allowed. Ensure that we use an allowed
635 * node for alloc or free.
636 */
637static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
638{
639    nid = next_node(nid, *nodes_allowed);
640    if (nid == MAX_NUMNODES)
641        nid = first_node(*nodes_allowed);
642    VM_BUG_ON(nid >= MAX_NUMNODES);
643
644    return nid;
645}
646
647static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
648{
649    if (!node_isset(nid, *nodes_allowed))
650        nid = next_node_allowed(nid, nodes_allowed);
651    return nid;
652}
653
654/*
655 * returns the previously saved node ["this node"] from which to
656 * allocate a persistent huge page for the pool and advance the
657 * next node from which to allocate, handling wrap at end of node
658 * mask.
659 */
660static int hstate_next_node_to_alloc(struct hstate *h,
661                    nodemask_t *nodes_allowed)
662{
663    int nid;
664
665    VM_BUG_ON(!nodes_allowed);
666
667    nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
668    h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
669
670    return nid;
671}
672
673static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
674{
675    struct page *page;
676    int start_nid;
677    int next_nid;
678    int ret = 0;
679
680    start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
681    next_nid = start_nid;
682
683    do {
684        page = alloc_fresh_huge_page_node(h, next_nid);
685        if (page) {
686            ret = 1;
687            break;
688        }
689        next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
690    } while (next_nid != start_nid);
691
692    if (ret)
693        count_vm_event(HTLB_BUDDY_PGALLOC);
694    else
695        count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
696
697    return ret;
698}
699
700/*
701 * helper for free_pool_huge_page() - return the previously saved
702 * node ["this node"] from which to free a huge page. Advance the
703 * next node id whether or not we find a free huge page to free so
704 * that the next attempt to free addresses the next node.
705 */
706static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
707{
708    int nid;
709
710    VM_BUG_ON(!nodes_allowed);
711
712    nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
713    h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
714
715    return nid;
716}
717
718/*
719 * Free huge page from pool from next node to free.
720 * Attempt to keep persistent huge pages more or less
721 * balanced over allowed nodes.
722 * Called with hugetlb_lock locked.
723 */
724static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
725                             bool acct_surplus)
726{
727    int start_nid;
728    int next_nid;
729    int ret = 0;
730
731    start_nid = hstate_next_node_to_free(h, nodes_allowed);
732    next_nid = start_nid;
733
734    do {
735        /*
736         * If we're returning unused surplus pages, only examine
737         * nodes with surplus pages.
738         */
739        if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
740            !list_empty(&h->hugepage_freelists[next_nid])) {
741            struct page *page =
742                list_entry(h->hugepage_freelists[next_nid].next,
743                      struct page, lru);
744            list_del(&page->lru);
745            h->free_huge_pages--;
746            h->free_huge_pages_node[next_nid]--;
747            if (acct_surplus) {
748                h->surplus_huge_pages--;
749                h->surplus_huge_pages_node[next_nid]--;
750            }
751            update_and_free_page(h, page);
752            ret = 1;
753            break;
754        }
755        next_nid = hstate_next_node_to_free(h, nodes_allowed);
756    } while (next_nid != start_nid);
757
758    return ret;
759}
760
761static struct page *alloc_buddy_huge_page(struct hstate *h,
762            struct vm_area_struct *vma, unsigned long address)
763{
764    struct page *page;
765    unsigned int nid;
766
767    if (h->order >= MAX_ORDER)
768        return NULL;
769
770    /*
771     * Assume we will successfully allocate the surplus page to
772     * prevent racing processes from causing the surplus to exceed
773     * overcommit
774     *
775     * This however introduces a different race, where a process B
776     * tries to grow the static hugepage pool while alloc_pages() is
777     * called by process A. B will only examine the per-node
778     * counters in determining if surplus huge pages can be
779     * converted to normal huge pages in adjust_pool_surplus(). A
780     * won't be able to increment the per-node counter, until the
781     * lock is dropped by B, but B doesn't drop hugetlb_lock until
782     * no more huge pages can be converted from surplus to normal
783     * state (and doesn't try to convert again). Thus, we have a
784     * case where a surplus huge page exists, the pool is grown, and
785     * the surplus huge page still exists after, even though it
786     * should just have been converted to a normal huge page. This
787     * does not leak memory, though, as the hugepage will be freed
788     * once it is out of use. It also does not allow the counters to
789     * go out of whack in adjust_pool_surplus() as we don't modify
790     * the node values until we've gotten the hugepage and only the
791     * per-node value is checked there.
792     */
793    spin_lock(&hugetlb_lock);
794    if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
795        spin_unlock(&hugetlb_lock);
796        return NULL;
797    } else {
798        h->nr_huge_pages++;
799        h->surplus_huge_pages++;
800    }
801    spin_unlock(&hugetlb_lock);
802
803    page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
804                    __GFP_REPEAT|__GFP_NOWARN,
805                    huge_page_order(h));
806
807    if (page && arch_prepare_hugepage(page)) {
808        __free_pages(page, huge_page_order(h));
809        return NULL;
810    }
811
812    spin_lock(&hugetlb_lock);
813    if (page) {
814        /*
815         * This page is now managed by the hugetlb allocator and has
816         * no users -- drop the buddy allocator's reference.
817         */
818        put_page_testzero(page);
819        VM_BUG_ON(page_count(page));
820        nid = page_to_nid(page);
821        set_compound_page_dtor(page, free_huge_page);
822        /*
823         * We incremented the global counters already
824         */
825        h->nr_huge_pages_node[nid]++;
826        h->surplus_huge_pages_node[nid]++;
827        __count_vm_event(HTLB_BUDDY_PGALLOC);
828    } else {
829        h->nr_huge_pages--;
830        h->surplus_huge_pages--;
831        __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
832    }
833    spin_unlock(&hugetlb_lock);
834
835    return page;
836}
837
838/*
839 * Increase the hugetlb pool such that it can accomodate a reservation
840 * of size 'delta'.
841 */
842static int gather_surplus_pages(struct hstate *h, int delta)
843{
844    struct list_head surplus_list;
845    struct page *page, *tmp;
846    int ret, i;
847    int needed, allocated;
848
849    needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
850    if (needed <= 0) {
851        h->resv_huge_pages += delta;
852        return 0;
853    }
854
855    allocated = 0;
856    INIT_LIST_HEAD(&surplus_list);
857
858    ret = -ENOMEM;
859retry:
860    spin_unlock(&hugetlb_lock);
861    for (i = 0; i < needed; i++) {
862        page = alloc_buddy_huge_page(h, NULL, 0);
863        if (!page) {
864            /*
865             * We were not able to allocate enough pages to
866             * satisfy the entire reservation so we free what
867             * we've allocated so far.
868             */
869            spin_lock(&hugetlb_lock);
870            needed = 0;
871            goto free;
872        }
873
874        list_add(&page->lru, &surplus_list);
875    }
876    allocated += needed;
877
878    /*
879     * After retaking hugetlb_lock, we need to recalculate 'needed'
880     * because either resv_huge_pages or free_huge_pages may have changed.
881     */
882    spin_lock(&hugetlb_lock);
883    needed = (h->resv_huge_pages + delta) -
884            (h->free_huge_pages + allocated);
885    if (needed > 0)
886        goto retry;
887
888    /*
889     * The surplus_list now contains _at_least_ the number of extra pages
890     * needed to accomodate the reservation. Add the appropriate number
891     * of pages to the hugetlb pool and free the extras back to the buddy
892     * allocator. Commit the entire reservation here to prevent another
893     * process from stealing the pages as they are added to the pool but
894     * before they are reserved.
895     */
896    needed += allocated;
897    h->resv_huge_pages += delta;
898    ret = 0;
899free:
900    /* Free the needed pages to the hugetlb pool */
901    list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
902        if ((--needed) < 0)
903            break;
904        list_del(&page->lru);
905        enqueue_huge_page(h, page);
906    }
907
908    /* Free unnecessary surplus pages to the buddy allocator */
909    if (!list_empty(&surplus_list)) {
910        spin_unlock(&hugetlb_lock);
911        list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
912            list_del(&page->lru);
913            /*
914             * The page has a reference count of zero already, so
915             * call free_huge_page directly instead of using
916             * put_page. This must be done with hugetlb_lock
917             * unlocked which is safe because free_huge_page takes
918             * hugetlb_lock before deciding how to free the page.
919             */
920            free_huge_page(page);
921        }
922        spin_lock(&hugetlb_lock);
923    }
924
925    return ret;
926}
927
928/*
929 * When releasing a hugetlb pool reservation, any surplus pages that were
930 * allocated to satisfy the reservation must be explicitly freed if they were
931 * never used.
932 * Called with hugetlb_lock held.
933 */
934static void return_unused_surplus_pages(struct hstate *h,
935                    unsigned long unused_resv_pages)
936{
937    unsigned long nr_pages;
938
939    /* Uncommit the reservation */
940    h->resv_huge_pages -= unused_resv_pages;
941
942    /* Cannot return gigantic pages currently */
943    if (h->order >= MAX_ORDER)
944        return;
945
946    nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
947
948    /*
949     * We want to release as many surplus pages as possible, spread
950     * evenly across all nodes with memory. Iterate across these nodes
951     * until we can no longer free unreserved surplus pages. This occurs
952     * when the nodes with surplus pages have no free pages.
953     * free_pool_huge_page() will balance the the freed pages across the
954     * on-line nodes with memory and will handle the hstate accounting.
955     */
956    while (nr_pages--) {
957        if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
958            break;
959    }
960}
961
962/*
963 * Determine if the huge page at addr within the vma has an associated
964 * reservation. Where it does not we will need to logically increase
965 * reservation and actually increase quota before an allocation can occur.
966 * Where any new reservation would be required the reservation change is
967 * prepared, but not committed. Once the page has been quota'd allocated
968 * an instantiated the change should be committed via vma_commit_reservation.
969 * No action is required on failure.
970 */
971static long vma_needs_reservation(struct hstate *h,
972            struct vm_area_struct *vma, unsigned long addr)
973{
974    struct address_space *mapping = vma->vm_file->f_mapping;
975    struct inode *inode = mapping->host;
976
977    if (vma->vm_flags & VM_MAYSHARE) {
978        pgoff_t idx = vma_hugecache_offset(h, vma, addr);
979        return region_chg(&inode->i_mapping->private_list,
980                            idx, idx + 1);
981
982    } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
983        return 1;
984
985    } else {
986        long err;
987        pgoff_t idx = vma_hugecache_offset(h, vma, addr);
988        struct resv_map *reservations = vma_resv_map(vma);
989
990        err = region_chg(&reservations->regions, idx, idx + 1);
991        if (err < 0)
992            return err;
993        return 0;
994    }
995}
996static void vma_commit_reservation(struct hstate *h,
997            struct vm_area_struct *vma, unsigned long addr)
998{
999    struct address_space *mapping = vma->vm_file->f_mapping;
1000    struct inode *inode = mapping->host;
1001
1002    if (vma->vm_flags & VM_MAYSHARE) {
1003        pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1004        region_add(&inode->i_mapping->private_list, idx, idx + 1);
1005
1006    } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1007        pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1008        struct resv_map *reservations = vma_resv_map(vma);
1009
1010        /* Mark this page used in the map. */
1011        region_add(&reservations->regions, idx, idx + 1);
1012    }
1013}
1014
1015static struct page *alloc_huge_page(struct vm_area_struct *vma,
1016                    unsigned long addr, int avoid_reserve)
1017{
1018    struct hstate *h = hstate_vma(vma);
1019    struct page *page;
1020    struct address_space *mapping = vma->vm_file->f_mapping;
1021    struct inode *inode = mapping->host;
1022    long chg;
1023
1024    /*
1025     * Processes that did not create the mapping will have no reserves and
1026     * will not have accounted against quota. Check that the quota can be
1027     * made before satisfying the allocation
1028     * MAP_NORESERVE mappings may also need pages and quota allocated
1029     * if no reserve mapping overlaps.
1030     */
1031    chg = vma_needs_reservation(h, vma, addr);
1032    if (chg < 0)
1033        return ERR_PTR(chg);
1034    if (chg)
1035        if (hugetlb_get_quota(inode->i_mapping, chg))
1036            return ERR_PTR(-ENOSPC);
1037
1038    spin_lock(&hugetlb_lock);
1039    page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1040    spin_unlock(&hugetlb_lock);
1041
1042    if (!page) {
1043        page = alloc_buddy_huge_page(h, vma, addr);
1044        if (!page) {
1045            hugetlb_put_quota(inode->i_mapping, chg);
1046            return ERR_PTR(-VM_FAULT_SIGBUS);
1047        }
1048    }
1049
1050    set_page_refcounted(page);
1051    set_page_private(page, (unsigned long) mapping);
1052
1053    vma_commit_reservation(h, vma, addr);
1054
1055    return page;
1056}
1057
1058int __weak alloc_bootmem_huge_page(struct hstate *h)
1059{
1060    struct huge_bootmem_page *m;
1061    int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1062
1063    while (nr_nodes) {
1064        void *addr;
1065
1066        addr = __alloc_bootmem_node_nopanic(
1067                NODE_DATA(hstate_next_node_to_alloc(h,
1068                        &node_states[N_HIGH_MEMORY])),
1069                huge_page_size(h), huge_page_size(h), 0);
1070
1071        if (addr) {
1072            /*
1073             * Use the beginning of the huge page to store the
1074             * huge_bootmem_page struct (until gather_bootmem
1075             * puts them into the mem_map).
1076             */
1077            m = addr;
1078            goto found;
1079        }
1080        nr_nodes--;
1081    }
1082    return 0;
1083
1084found:
1085    BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1086    /* Put them into a private list first because mem_map is not up yet */
1087    list_add(&m->list, &huge_boot_pages);
1088    m->hstate = h;
1089    return 1;
1090}
1091
1092static void prep_compound_huge_page(struct page *page, int order)
1093{
1094    if (unlikely(order > (MAX_ORDER - 1)))
1095        prep_compound_gigantic_page(page, order);
1096    else
1097        prep_compound_page(page, order);
1098}
1099
1100/* Put bootmem huge pages into the standard lists after mem_map is up */
1101static void __init gather_bootmem_prealloc(void)
1102{
1103    struct huge_bootmem_page *m;
1104
1105    list_for_each_entry(m, &huge_boot_pages, list) {
1106        struct page *page = virt_to_page(m);
1107        struct hstate *h = m->hstate;
1108        __ClearPageReserved(page);
1109        WARN_ON(page_count(page) != 1);
1110        prep_compound_huge_page(page, h->order);
1111        prep_new_huge_page(h, page, page_to_nid(page));
1112    }
1113}
1114
1115static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1116{
1117    unsigned long i;
1118
1119    for (i = 0; i < h->max_huge_pages; ++i) {
1120        if (h->order >= MAX_ORDER) {
1121            if (!alloc_bootmem_huge_page(h))
1122                break;
1123        } else if (!alloc_fresh_huge_page(h,
1124                     &node_states[N_HIGH_MEMORY]))
1125            break;
1126    }
1127    h->max_huge_pages = i;
1128}
1129
1130static void __init hugetlb_init_hstates(void)
1131{
1132    struct hstate *h;
1133
1134    for_each_hstate(h) {
1135        /* oversize hugepages were init'ed in early boot */
1136        if (h->order < MAX_ORDER)
1137            hugetlb_hstate_alloc_pages(h);
1138    }
1139}
1140
1141static char * __init memfmt(char *buf, unsigned long n)
1142{
1143    if (n >= (1UL << 30))
1144        sprintf(buf, "%lu GB", n >> 30);
1145    else if (n >= (1UL << 20))
1146        sprintf(buf, "%lu MB", n >> 20);
1147    else
1148        sprintf(buf, "%lu KB", n >> 10);
1149    return buf;
1150}
1151
1152static void __init report_hugepages(void)
1153{
1154    struct hstate *h;
1155
1156    for_each_hstate(h) {
1157        char buf[32];
1158        printk(KERN_INFO "HugeTLB registered %s page size, "
1159                 "pre-allocated %ld pages\n",
1160            memfmt(buf, huge_page_size(h)),
1161            h->free_huge_pages);
1162    }
1163}
1164
1165#ifdef CONFIG_HIGHMEM
1166static void try_to_free_low(struct hstate *h, unsigned long count,
1167                        nodemask_t *nodes_allowed)
1168{
1169    int i;
1170
1171    if (h->order >= MAX_ORDER)
1172        return;
1173
1174    for_each_node_mask(i, *nodes_allowed) {
1175        struct page *page, *next;
1176        struct list_head *freel = &h->hugepage_freelists[i];
1177        list_for_each_entry_safe(page, next, freel, lru) {
1178            if (count >= h->nr_huge_pages)
1179                return;
1180            if (PageHighMem(page))
1181                continue;
1182            list_del(&page->lru);
1183            update_and_free_page(h, page);
1184            h->free_huge_pages--;
1185            h->free_huge_pages_node[page_to_nid(page)]--;
1186        }
1187    }
1188}
1189#else
1190static inline void try_to_free_low(struct hstate *h, unsigned long count,
1191                        nodemask_t *nodes_allowed)
1192{
1193}
1194#endif
1195
1196/*
1197 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1198 * balanced by operating on them in a round-robin fashion.
1199 * Returns 1 if an adjustment was made.
1200 */
1201static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1202                int delta)
1203{
1204    int start_nid, next_nid;
1205    int ret = 0;
1206
1207    VM_BUG_ON(delta != -1 && delta != 1);
1208
1209    if (delta < 0)
1210        start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1211    else
1212        start_nid = hstate_next_node_to_free(h, nodes_allowed);
1213    next_nid = start_nid;
1214
1215    do {
1216        int nid = next_nid;
1217        if (delta < 0) {
1218            /*
1219             * To shrink on this node, there must be a surplus page
1220             */
1221            if (!h->surplus_huge_pages_node[nid]) {
1222                next_nid = hstate_next_node_to_alloc(h,
1223                                nodes_allowed);
1224                continue;
1225            }
1226        }
1227        if (delta > 0) {
1228            /*
1229             * Surplus cannot exceed the total number of pages
1230             */
1231            if (h->surplus_huge_pages_node[nid] >=
1232                        h->nr_huge_pages_node[nid]) {
1233                next_nid = hstate_next_node_to_free(h,
1234                                nodes_allowed);
1235                continue;
1236            }
1237        }
1238
1239        h->surplus_huge_pages += delta;
1240        h->surplus_huge_pages_node[nid] += delta;
1241        ret = 1;
1242        break;
1243    } while (next_nid != start_nid);
1244
1245    return ret;
1246}
1247
1248#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1249static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1250                        nodemask_t *nodes_allowed)
1251{
1252    unsigned long min_count, ret;
1253
1254    if (h->order >= MAX_ORDER)
1255        return h->max_huge_pages;
1256
1257    /*
1258     * Increase the pool size
1259     * First take pages out of surplus state. Then make up the
1260     * remaining difference by allocating fresh huge pages.
1261     *
1262     * We might race with alloc_buddy_huge_page() here and be unable
1263     * to convert a surplus huge page to a normal huge page. That is
1264     * not critical, though, it just means the overall size of the
1265     * pool might be one hugepage larger than it needs to be, but
1266     * within all the constraints specified by the sysctls.
1267     */
1268    spin_lock(&hugetlb_lock);
1269    while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1270        if (!adjust_pool_surplus(h, nodes_allowed, -1))
1271            break;
1272    }
1273
1274    while (count > persistent_huge_pages(h)) {
1275        /*
1276         * If this allocation races such that we no longer need the
1277         * page, free_huge_page will handle it by freeing the page
1278         * and reducing the surplus.
1279         */
1280        spin_unlock(&hugetlb_lock);
1281        ret = alloc_fresh_huge_page(h, nodes_allowed);
1282        spin_lock(&hugetlb_lock);
1283        if (!ret)
1284            goto out;
1285
1286        /* Bail for signals. Probably ctrl-c from user */
1287        if (signal_pending(current))
1288            goto out;
1289    }
1290
1291    /*
1292     * Decrease the pool size
1293     * First return free pages to the buddy allocator (being careful
1294     * to keep enough around to satisfy reservations). Then place
1295     * pages into surplus state as needed so the pool will shrink
1296     * to the desired size as pages become free.
1297     *
1298     * By placing pages into the surplus state independent of the
1299     * overcommit value, we are allowing the surplus pool size to
1300     * exceed overcommit. There are few sane options here. Since
1301     * alloc_buddy_huge_page() is checking the global counter,
1302     * though, we'll note that we're not allowed to exceed surplus
1303     * and won't grow the pool anywhere else. Not until one of the
1304     * sysctls are changed, or the surplus pages go out of use.
1305     */
1306    min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1307    min_count = max(count, min_count);
1308    try_to_free_low(h, min_count, nodes_allowed);
1309    while (min_count < persistent_huge_pages(h)) {
1310        if (!free_pool_huge_page(h, nodes_allowed, 0))
1311            break;
1312    }
1313    while (count < persistent_huge_pages(h)) {
1314        if (!adjust_pool_surplus(h, nodes_allowed, 1))
1315            break;
1316    }
1317out:
1318    ret = persistent_huge_pages(h);
1319    spin_unlock(&hugetlb_lock);
1320    return ret;
1321}
1322
1323#define HSTATE_ATTR_RO(_name) \
1324    static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1325
1326#define HSTATE_ATTR(_name) \
1327    static struct kobj_attribute _name##_attr = \
1328        __ATTR(_name, 0644, _name##_show, _name##_store)
1329
1330static struct kobject *hugepages_kobj;
1331static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1332
1333static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1334
1335static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1336{
1337    int i;
1338
1339    for (i = 0; i < HUGE_MAX_HSTATE; i++)
1340        if (hstate_kobjs[i] == kobj) {
1341            if (nidp)
1342                *nidp = NUMA_NO_NODE;
1343            return &hstates[i];
1344        }
1345
1346    return kobj_to_node_hstate(kobj, nidp);
1347}
1348
1349static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1350                    struct kobj_attribute *attr, char *buf)
1351{
1352    struct hstate *h;
1353    unsigned long nr_huge_pages;
1354    int nid;
1355
1356    h = kobj_to_hstate(kobj, &nid);
1357    if (nid == NUMA_NO_NODE)
1358        nr_huge_pages = h->nr_huge_pages;
1359    else
1360        nr_huge_pages = h->nr_huge_pages_node[nid];
1361
1362    return sprintf(buf, "%lu\n", nr_huge_pages);
1363}
1364static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1365            struct kobject *kobj, struct kobj_attribute *attr,
1366            const char *buf, size_t len)
1367{
1368    int err;
1369    int nid;
1370    unsigned long count;
1371    struct hstate *h;
1372    NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1373
1374    err = strict_strtoul(buf, 10, &count);
1375    if (err)
1376        return 0;
1377
1378    h = kobj_to_hstate(kobj, &nid);
1379    if (nid == NUMA_NO_NODE) {
1380        /*
1381         * global hstate attribute
1382         */
1383        if (!(obey_mempolicy &&
1384                init_nodemask_of_mempolicy(nodes_allowed))) {
1385            NODEMASK_FREE(nodes_allowed);
1386            nodes_allowed = &node_states[N_HIGH_MEMORY];
1387        }
1388    } else if (nodes_allowed) {
1389        /*
1390         * per node hstate attribute: adjust count to global,
1391         * but restrict alloc/free to the specified node.
1392         */
1393        count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1394        init_nodemask_of_node(nodes_allowed, nid);
1395    } else
1396        nodes_allowed = &node_states[N_HIGH_MEMORY];
1397
1398    h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1399
1400    if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1401        NODEMASK_FREE(nodes_allowed);
1402
1403    return len;
1404}
1405
1406static ssize_t nr_hugepages_show(struct kobject *kobj,
1407                       struct kobj_attribute *attr, char *buf)
1408{
1409    return nr_hugepages_show_common(kobj, attr, buf);
1410}
1411
1412static ssize_t nr_hugepages_store(struct kobject *kobj,
1413           struct kobj_attribute *attr, const char *buf, size_t len)
1414{
1415    return nr_hugepages_store_common(false, kobj, attr, buf, len);
1416}
1417HSTATE_ATTR(nr_hugepages);
1418
1419#ifdef CONFIG_NUMA
1420
1421/*
1422 * hstate attribute for optionally mempolicy-based constraint on persistent
1423 * huge page alloc/free.
1424 */
1425static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1426                       struct kobj_attribute *attr, char *buf)
1427{
1428    return nr_hugepages_show_common(kobj, attr, buf);
1429}
1430
1431static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1432           struct kobj_attribute *attr, const char *buf, size_t len)
1433{
1434    return nr_hugepages_store_common(true, kobj, attr, buf, len);
1435}
1436HSTATE_ATTR(nr_hugepages_mempolicy);
1437#endif
1438
1439
1440static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1441                    struct kobj_attribute *attr, char *buf)
1442{
1443    struct hstate *h = kobj_to_hstate(kobj, NULL);
1444    return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1445}
1446static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1447        struct kobj_attribute *attr, const char *buf, size_t count)
1448{
1449    int err;
1450    unsigned long input;
1451    struct hstate *h = kobj_to_hstate(kobj, NULL);
1452
1453    err = strict_strtoul(buf, 10, &input);
1454    if (err)
1455        return 0;
1456
1457    spin_lock(&hugetlb_lock);
1458    h->nr_overcommit_huge_pages = input;
1459    spin_unlock(&hugetlb_lock);
1460
1461    return count;
1462}
1463HSTATE_ATTR(nr_overcommit_hugepages);
1464
1465static ssize_t free_hugepages_show(struct kobject *kobj,
1466                    struct kobj_attribute *attr, char *buf)
1467{
1468    struct hstate *h;
1469    unsigned long free_huge_pages;
1470    int nid;
1471
1472    h = kobj_to_hstate(kobj, &nid);
1473    if (nid == NUMA_NO_NODE)
1474        free_huge_pages = h->free_huge_pages;
1475    else
1476        free_huge_pages = h->free_huge_pages_node[nid];
1477
1478    return sprintf(buf, "%lu\n", free_huge_pages);
1479}
1480HSTATE_ATTR_RO(free_hugepages);
1481
1482static ssize_t resv_hugepages_show(struct kobject *kobj,
1483                    struct kobj_attribute *attr, char *buf)
1484{
1485    struct hstate *h = kobj_to_hstate(kobj, NULL);
1486    return sprintf(buf, "%lu\n", h->resv_huge_pages);
1487}
1488HSTATE_ATTR_RO(resv_hugepages);
1489
1490static ssize_t surplus_hugepages_show(struct kobject *kobj,
1491                    struct kobj_attribute *attr, char *buf)
1492{
1493    struct hstate *h;
1494    unsigned long surplus_huge_pages;
1495    int nid;
1496
1497    h = kobj_to_hstate(kobj, &nid);
1498    if (nid == NUMA_NO_NODE)
1499        surplus_huge_pages = h->surplus_huge_pages;
1500    else
1501        surplus_huge_pages = h->surplus_huge_pages_node[nid];
1502
1503    return sprintf(buf, "%lu\n", surplus_huge_pages);
1504}
1505HSTATE_ATTR_RO(surplus_hugepages);
1506
1507static struct attribute *hstate_attrs[] = {
1508    &nr_hugepages_attr.attr,
1509    &nr_overcommit_hugepages_attr.attr,
1510    &free_hugepages_attr.attr,
1511    &resv_hugepages_attr.attr,
1512    &surplus_hugepages_attr.attr,
1513#ifdef CONFIG_NUMA
1514    &nr_hugepages_mempolicy_attr.attr,
1515#endif
1516    NULL,
1517};
1518
1519static struct attribute_group hstate_attr_group = {
1520    .attrs = hstate_attrs,
1521};
1522
1523static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1524                    struct kobject **hstate_kobjs,
1525                    struct attribute_group *hstate_attr_group)
1526{
1527    int retval;
1528    int hi = h - hstates;
1529
1530    hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1531    if (!hstate_kobjs[hi])
1532        return -ENOMEM;
1533
1534    retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1535    if (retval)
1536        kobject_put(hstate_kobjs[hi]);
1537
1538    return retval;
1539}
1540
1541static void __init hugetlb_sysfs_init(void)
1542{
1543    struct hstate *h;
1544    int err;
1545
1546    hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1547    if (!hugepages_kobj)
1548        return;
1549
1550    for_each_hstate(h) {
1551        err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1552                     hstate_kobjs, &hstate_attr_group);
1553        if (err)
1554            printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1555                                h->name);
1556    }
1557}
1558
1559#ifdef CONFIG_NUMA
1560
1561/*
1562 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1563 * with node sysdevs in node_devices[] using a parallel array. The array
1564 * index of a node sysdev or _hstate == node id.
1565 * This is here to avoid any static dependency of the node sysdev driver, in
1566 * the base kernel, on the hugetlb module.
1567 */
1568struct node_hstate {
1569    struct kobject *hugepages_kobj;
1570    struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1571};
1572struct node_hstate node_hstates[MAX_NUMNODES];
1573
1574/*
1575 * A subset of global hstate attributes for node sysdevs
1576 */
1577static struct attribute *per_node_hstate_attrs[] = {
1578    &nr_hugepages_attr.attr,
1579    &free_hugepages_attr.attr,
1580    &surplus_hugepages_attr.attr,
1581    NULL,
1582};
1583
1584static struct attribute_group per_node_hstate_attr_group = {
1585    .attrs = per_node_hstate_attrs,
1586};
1587
1588/*
1589 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1590 * Returns node id via non-NULL nidp.
1591 */
1592static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1593{
1594    int nid;
1595
1596    for (nid = 0; nid < nr_node_ids; nid++) {
1597        struct node_hstate *nhs = &node_hstates[nid];
1598        int i;
1599        for (i = 0; i < HUGE_MAX_HSTATE; i++)
1600            if (nhs->hstate_kobjs[i] == kobj) {
1601                if (nidp)
1602                    *nidp = nid;
1603                return &hstates[i];
1604            }
1605    }
1606
1607    BUG();
1608    return NULL;
1609}
1610
1611/*
1612 * Unregister hstate attributes from a single node sysdev.
1613 * No-op if no hstate attributes attached.
1614 */
1615void hugetlb_unregister_node(struct node *node)
1616{
1617    struct hstate *h;
1618    struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1619
1620    if (!nhs->hugepages_kobj)
1621        return; /* no hstate attributes */
1622
1623    for_each_hstate(h)
1624        if (nhs->hstate_kobjs[h - hstates]) {
1625            kobject_put(nhs->hstate_kobjs[h - hstates]);
1626            nhs->hstate_kobjs[h - hstates] = NULL;
1627        }
1628
1629    kobject_put(nhs->hugepages_kobj);
1630    nhs->hugepages_kobj = NULL;
1631}
1632
1633/*
1634 * hugetlb module exit: unregister hstate attributes from node sysdevs
1635 * that have them.
1636 */
1637static void hugetlb_unregister_all_nodes(void)
1638{
1639    int nid;
1640
1641    /*
1642     * disable node sysdev registrations.
1643     */
1644    register_hugetlbfs_with_node(NULL, NULL);
1645
1646    /*
1647     * remove hstate attributes from any nodes that have them.
1648     */
1649    for (nid = 0; nid < nr_node_ids; nid++)
1650        hugetlb_unregister_node(&node_devices[nid]);
1651}
1652
1653/*
1654 * Register hstate attributes for a single node sysdev.
1655 * No-op if attributes already registered.
1656 */
1657void hugetlb_register_node(struct node *node)
1658{
1659    struct hstate *h;
1660    struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1661    int err;
1662
1663    if (nhs->hugepages_kobj)
1664        return; /* already allocated */
1665
1666    nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1667                            &node->sysdev.kobj);
1668    if (!nhs->hugepages_kobj)
1669        return;
1670
1671    for_each_hstate(h) {
1672        err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1673                        nhs->hstate_kobjs,
1674                        &per_node_hstate_attr_group);
1675        if (err) {
1676            printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1677                    " for node %d\n",
1678                        h->name, node->sysdev.id);
1679            hugetlb_unregister_node(node);
1680            break;
1681        }
1682    }
1683}
1684
1685/*
1686 * hugetlb init time: register hstate attributes for all registered node
1687 * sysdevs of nodes that have memory. All on-line nodes should have
1688 * registered their associated sysdev by this time.
1689 */
1690static void hugetlb_register_all_nodes(void)
1691{
1692    int nid;
1693
1694    for_each_node_state(nid, N_HIGH_MEMORY) {
1695        struct node *node = &node_devices[nid];
1696        if (node->sysdev.id == nid)
1697            hugetlb_register_node(node);
1698    }
1699
1700    /*
1701     * Let the node sysdev driver know we're here so it can
1702     * [un]register hstate attributes on node hotplug.
1703     */
1704    register_hugetlbfs_with_node(hugetlb_register_node,
1705                     hugetlb_unregister_node);
1706}
1707#else /* !CONFIG_NUMA */
1708
1709static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1710{
1711    BUG();
1712    if (nidp)
1713        *nidp = -1;
1714    return NULL;
1715}
1716
1717static void hugetlb_unregister_all_nodes(void) { }
1718
1719static void hugetlb_register_all_nodes(void) { }
1720
1721#endif
1722
1723static void __exit hugetlb_exit(void)
1724{
1725    struct hstate *h;
1726
1727    hugetlb_unregister_all_nodes();
1728
1729    for_each_hstate(h) {
1730        kobject_put(hstate_kobjs[h - hstates]);
1731    }
1732
1733    kobject_put(hugepages_kobj);
1734}
1735module_exit(hugetlb_exit);
1736
1737static int __init hugetlb_init(void)
1738{
1739    /* Some platform decide whether they support huge pages at boot
1740     * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1741     * there is no such support
1742     */
1743    if (HPAGE_SHIFT == 0)
1744        return 0;
1745
1746    if (!size_to_hstate(default_hstate_size)) {
1747        default_hstate_size = HPAGE_SIZE;
1748        if (!size_to_hstate(default_hstate_size))
1749            hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1750    }
1751    default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1752    if (default_hstate_max_huge_pages)
1753        default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1754
1755    hugetlb_init_hstates();
1756
1757    gather_bootmem_prealloc();
1758
1759    report_hugepages();
1760
1761    hugetlb_sysfs_init();
1762
1763    hugetlb_register_all_nodes();
1764
1765    return 0;
1766}
1767module_init(hugetlb_init);
1768
1769/* Should be called on processing a hugepagesz=... option */
1770void __init hugetlb_add_hstate(unsigned order)
1771{
1772    struct hstate *h;
1773    unsigned long i;
1774
1775    if (size_to_hstate(PAGE_SIZE << order)) {
1776        printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1777        return;
1778    }
1779    BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1780    BUG_ON(order == 0);
1781    h = &hstates[max_hstate++];
1782    h->order = order;
1783    h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1784    h->nr_huge_pages = 0;
1785    h->free_huge_pages = 0;
1786    for (i = 0; i < MAX_NUMNODES; ++i)
1787        INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1788    h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1789    h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1790    snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1791                    huge_page_size(h)/1024);
1792
1793    parsed_hstate = h;
1794}
1795
1796static int __init hugetlb_nrpages_setup(char *s)
1797{
1798    unsigned long *mhp;
1799    static unsigned long *last_mhp;
1800
1801    /*
1802     * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1803     * so this hugepages= parameter goes to the "default hstate".
1804     */
1805    if (!max_hstate)
1806        mhp = &default_hstate_max_huge_pages;
1807    else
1808        mhp = &parsed_hstate->max_huge_pages;
1809
1810    if (mhp == last_mhp) {
1811        printk(KERN_WARNING "hugepages= specified twice without "
1812            "interleaving hugepagesz=, ignoring\n");
1813        return 1;
1814    }
1815
1816    if (sscanf(s, "%lu", mhp) <= 0)
1817        *mhp = 0;
1818
1819    /*
1820     * Global state is always initialized later in hugetlb_init.
1821     * But we need to allocate >= MAX_ORDER hstates here early to still
1822     * use the bootmem allocator.
1823     */
1824    if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1825        hugetlb_hstate_alloc_pages(parsed_hstate);
1826
1827    last_mhp = mhp;
1828
1829    return 1;
1830}
1831__setup("hugepages=", hugetlb_nrpages_setup);
1832
1833static int __init hugetlb_default_setup(char *s)
1834{
1835    default_hstate_size = memparse(s, &s);
1836    return 1;
1837}
1838__setup("default_hugepagesz=", hugetlb_default_setup);
1839
1840static unsigned int cpuset_mems_nr(unsigned int *array)
1841{
1842    int node;
1843    unsigned int nr = 0;
1844
1845    for_each_node_mask(node, cpuset_current_mems_allowed)
1846        nr += array[node];
1847
1848    return nr;
1849}
1850
1851#ifdef CONFIG_SYSCTL
1852static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1853             struct ctl_table *table, int write,
1854             void __user *buffer, size_t *length, loff_t *ppos)
1855{
1856    struct hstate *h = &default_hstate;
1857    unsigned long tmp;
1858
1859    if (!write)
1860        tmp = h->max_huge_pages;
1861
1862    table->data = &tmp;
1863    table->maxlen = sizeof(unsigned long);
1864    proc_doulongvec_minmax(table, write, buffer, length, ppos);
1865
1866    if (write) {
1867        NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1868                        GFP_KERNEL | __GFP_NORETRY);
1869        if (!(obey_mempolicy &&
1870                   init_nodemask_of_mempolicy(nodes_allowed))) {
1871            NODEMASK_FREE(nodes_allowed);
1872            nodes_allowed = &node_states[N_HIGH_MEMORY];
1873        }
1874        h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1875
1876        if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1877            NODEMASK_FREE(nodes_allowed);
1878    }
1879
1880    return 0;
1881}
1882
1883int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1884              void __user *buffer, size_t *length, loff_t *ppos)
1885{
1886
1887    return hugetlb_sysctl_handler_common(false, table, write,
1888                            buffer, length, ppos);
1889}
1890
1891#ifdef CONFIG_NUMA
1892int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1893              void __user *buffer, size_t *length, loff_t *ppos)
1894{
1895    return hugetlb_sysctl_handler_common(true, table, write,
1896                            buffer, length, ppos);
1897}
1898#endif /* CONFIG_NUMA */
1899
1900int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1901            void __user *buffer,
1902            size_t *length, loff_t *ppos)
1903{
1904    proc_dointvec(table, write, buffer, length, ppos);
1905    if (hugepages_treat_as_movable)
1906        htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1907    else
1908        htlb_alloc_mask = GFP_HIGHUSER;
1909    return 0;
1910}
1911
1912int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1913            void __user *buffer,
1914            size_t *length, loff_t *ppos)
1915{
1916    struct hstate *h = &default_hstate;
1917    unsigned long tmp;
1918
1919    if (!write)
1920        tmp = h->nr_overcommit_huge_pages;
1921
1922    table->data = &tmp;
1923    table->maxlen = sizeof(unsigned long);
1924    proc_doulongvec_minmax(table, write, buffer, length, ppos);
1925
1926    if (write) {
1927        spin_lock(&hugetlb_lock);
1928        h->nr_overcommit_huge_pages = tmp;
1929        spin_unlock(&hugetlb_lock);
1930    }
1931
1932    return 0;
1933}
1934
1935#endif /* CONFIG_SYSCTL */
1936
1937void hugetlb_report_meminfo(struct seq_file *m)
1938{
1939    struct hstate *h = &default_hstate;
1940    seq_printf(m,
1941            "HugePages_Total: %5lu\n"
1942            "HugePages_Free: %5lu\n"
1943            "HugePages_Rsvd: %5lu\n"
1944            "HugePages_Surp: %5lu\n"
1945            "Hugepagesize: %8lu kB\n",
1946            h->nr_huge_pages,
1947            h->free_huge_pages,
1948            h->resv_huge_pages,
1949            h->surplus_huge_pages,
1950            1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1951}
1952
1953int hugetlb_report_node_meminfo(int nid, char *buf)
1954{
1955    struct hstate *h = &default_hstate;
1956    return sprintf(buf,
1957        "Node %d HugePages_Total: %5u\n"
1958        "Node %d HugePages_Free: %5u\n"
1959        "Node %d HugePages_Surp: %5u\n",
1960        nid, h->nr_huge_pages_node[nid],
1961        nid, h->free_huge_pages_node[nid],
1962        nid, h->surplus_huge_pages_node[nid]);
1963}
1964
1965/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1966unsigned long hugetlb_total_pages(void)
1967{
1968    struct hstate *h = &default_hstate;
1969    return h->nr_huge_pages * pages_per_huge_page(h);
1970}
1971
1972static int hugetlb_acct_memory(struct hstate *h, long delta)
1973{
1974    int ret = -ENOMEM;
1975
1976    spin_lock(&hugetlb_lock);
1977    /*
1978     * When cpuset is configured, it breaks the strict hugetlb page
1979     * reservation as the accounting is done on a global variable. Such
1980     * reservation is completely rubbish in the presence of cpuset because
1981     * the reservation is not checked against page availability for the
1982     * current cpuset. Application can still potentially OOM'ed by kernel
1983     * with lack of free htlb page in cpuset that the task is in.
1984     * Attempt to enforce strict accounting with cpuset is almost
1985     * impossible (or too ugly) because cpuset is too fluid that
1986     * task or memory node can be dynamically moved between cpusets.
1987     *
1988     * The change of semantics for shared hugetlb mapping with cpuset is
1989     * undesirable. However, in order to preserve some of the semantics,
1990     * we fall back to check against current free page availability as
1991     * a best attempt and hopefully to minimize the impact of changing
1992     * semantics that cpuset has.
1993     */
1994    if (delta > 0) {
1995        if (gather_surplus_pages(h, delta) < 0)
1996            goto out;
1997
1998        if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1999            return_unused_surplus_pages(h, delta);
2000            goto out;
2001        }
2002    }
2003
2004    ret = 0;
2005    if (delta < 0)
2006        return_unused_surplus_pages(h, (unsigned long) -delta);
2007
2008out:
2009    spin_unlock(&hugetlb_lock);
2010    return ret;
2011}
2012
2013static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2014{
2015    struct resv_map *reservations = vma_resv_map(vma);
2016
2017    /*
2018     * This new VMA should share its siblings reservation map if present.
2019     * The VMA will only ever have a valid reservation map pointer where
2020     * it is being copied for another still existing VMA. As that VMA
2021     * has a reference to the reservation map it cannot dissappear until
2022     * after this open call completes. It is therefore safe to take a
2023     * new reference here without additional locking.
2024     */
2025    if (reservations)
2026        kref_get(&reservations->refs);
2027}
2028
2029static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2030{
2031    struct hstate *h = hstate_vma(vma);
2032    struct resv_map *reservations = vma_resv_map(vma);
2033    unsigned long reserve;
2034    unsigned long start;
2035    unsigned long end;
2036
2037    if (reservations) {
2038        start = vma_hugecache_offset(h, vma, vma->vm_start);
2039        end = vma_hugecache_offset(h, vma, vma->vm_end);
2040
2041        reserve = (end - start) -
2042            region_count(&reservations->regions, start, end);
2043
2044        kref_put(&reservations->refs, resv_map_release);
2045
2046        if (reserve) {
2047            hugetlb_acct_memory(h, -reserve);
2048            hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2049        }
2050    }
2051}
2052
2053/*
2054 * We cannot handle pagefaults against hugetlb pages at all. They cause
2055 * handle_mm_fault() to try to instantiate regular-sized pages in the
2056 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2057 * this far.
2058 */
2059static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2060{
2061    BUG();
2062    return 0;
2063}
2064
2065const struct vm_operations_struct hugetlb_vm_ops = {
2066    .fault = hugetlb_vm_op_fault,
2067    .open = hugetlb_vm_op_open,
2068    .close = hugetlb_vm_op_close,
2069};
2070
2071static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2072                int writable)
2073{
2074    pte_t entry;
2075
2076    if (writable) {
2077        entry =
2078            pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2079    } else {
2080        entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2081    }
2082    entry = pte_mkyoung(entry);
2083    entry = pte_mkhuge(entry);
2084
2085    return entry;
2086}
2087
2088static void set_huge_ptep_writable(struct vm_area_struct *vma,
2089                   unsigned long address, pte_t *ptep)
2090{
2091    pte_t entry;
2092
2093    entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2094    if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2095        update_mmu_cache(vma, address, ptep);
2096    }
2097}
2098
2099
2100int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2101                struct vm_area_struct *vma)
2102{
2103    pte_t *src_pte, *dst_pte, entry;
2104    struct page *ptepage;
2105    unsigned long addr;
2106    int cow;
2107    struct hstate *h = hstate_vma(vma);
2108    unsigned long sz = huge_page_size(h);
2109
2110    cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2111
2112    for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2113        src_pte = huge_pte_offset(src, addr);
2114        if (!src_pte)
2115            continue;
2116        dst_pte = huge_pte_alloc(dst, addr, sz);
2117        if (!dst_pte)
2118            goto nomem;
2119
2120        /* If the pagetables are shared don't copy or take references */
2121        if (dst_pte == src_pte)
2122            continue;
2123
2124        spin_lock(&dst->page_table_lock);
2125        spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2126        if (!huge_pte_none(huge_ptep_get(src_pte))) {
2127            if (cow)
2128                huge_ptep_set_wrprotect(src, addr, src_pte);
2129            entry = huge_ptep_get(src_pte);
2130            ptepage = pte_page(entry);
2131            get_page(ptepage);
2132            set_huge_pte_at(dst, addr, dst_pte, entry);
2133        }
2134        spin_unlock(&src->page_table_lock);
2135        spin_unlock(&dst->page_table_lock);
2136    }
2137    return 0;
2138
2139nomem:
2140    return -ENOMEM;
2141}
2142
2143void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2144                unsigned long end, struct page *ref_page)
2145{
2146    struct mm_struct *mm = vma->vm_mm;
2147    unsigned long address;
2148    pte_t *ptep;
2149    pte_t pte;
2150    struct page *page;
2151    struct page *tmp;
2152    struct hstate *h = hstate_vma(vma);
2153    unsigned long sz = huge_page_size(h);
2154
2155    /*
2156     * A page gathering list, protected by per file i_mmap_lock. The
2157     * lock is used to avoid list corruption from multiple unmapping
2158     * of the same page since we are using page->lru.
2159     */
2160    LIST_HEAD(page_list);
2161
2162    WARN_ON(!is_vm_hugetlb_page(vma));
2163    BUG_ON(start & ~huge_page_mask(h));
2164    BUG_ON(end & ~huge_page_mask(h));
2165
2166    mmu_notifier_invalidate_range_start(mm, start, end);
2167    spin_lock(&mm->page_table_lock);
2168    for (address = start; address < end; address += sz) {
2169        ptep = huge_pte_offset(mm, address);
2170        if (!ptep)
2171            continue;
2172
2173        if (huge_pmd_unshare(mm, &address, ptep))
2174            continue;
2175
2176        /*
2177         * If a reference page is supplied, it is because a specific
2178         * page is being unmapped, not a range. Ensure the page we
2179         * are about to unmap is the actual page of interest.
2180         */
2181        if (ref_page) {
2182            pte = huge_ptep_get(ptep);
2183            if (huge_pte_none(pte))
2184                continue;
2185            page = pte_page(pte);
2186            if (page != ref_page)
2187                continue;
2188
2189            /*
2190             * Mark the VMA as having unmapped its page so that
2191             * future faults in this VMA will fail rather than
2192             * looking like data was lost
2193             */
2194            set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2195        }
2196
2197        pte = huge_ptep_get_and_clear(mm, address, ptep);
2198        if (huge_pte_none(pte))
2199            continue;
2200
2201        page = pte_page(pte);
2202        if (pte_dirty(pte))
2203            set_page_dirty(page);
2204        list_add(&page->lru, &page_list);
2205    }
2206    spin_unlock(&mm->page_table_lock);
2207    flush_tlb_range(vma, start, end);
2208    mmu_notifier_invalidate_range_end(mm, start, end);
2209    list_for_each_entry_safe(page, tmp, &page_list, lru) {
2210        list_del(&page->lru);
2211        put_page(page);
2212    }
2213}
2214
2215void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2216              unsigned long end, struct page *ref_page)
2217{
2218    spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2219    __unmap_hugepage_range(vma, start, end, ref_page);
2220    spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2221}
2222
2223/*
2224 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2225 * mappping it owns the reserve page for. The intention is to unmap the page
2226 * from other VMAs and let the children be SIGKILLed if they are faulting the
2227 * same region.
2228 */
2229static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2230                struct page *page, unsigned long address)
2231{
2232    struct hstate *h = hstate_vma(vma);
2233    struct vm_area_struct *iter_vma;
2234    struct address_space *mapping;
2235    struct prio_tree_iter iter;
2236    pgoff_t pgoff;
2237
2238    /*
2239     * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2240     * from page cache lookup which is in HPAGE_SIZE units.
2241     */
2242    address = address & huge_page_mask(h);
2243    pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2244        + (vma->vm_pgoff >> PAGE_SHIFT);
2245    mapping = (struct address_space *)page_private(page);
2246
2247    /*
2248     * Take the mapping lock for the duration of the table walk. As
2249     * this mapping should be shared between all the VMAs,
2250     * __unmap_hugepage_range() is called as the lock is already held
2251     */
2252    spin_lock(&mapping->i_mmap_lock);
2253    vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2254        /* Do not unmap the current VMA */
2255        if (iter_vma == vma)
2256            continue;
2257
2258        /*
2259         * Unmap the page from other VMAs without their own reserves.
2260         * They get marked to be SIGKILLed if they fault in these
2261         * areas. This is because a future no-page fault on this VMA
2262         * could insert a zeroed page instead of the data existing
2263         * from the time of fork. This would look like data corruption
2264         */
2265        if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2266            __unmap_hugepage_range(iter_vma,
2267                address, address + huge_page_size(h),
2268                page);
2269    }
2270    spin_unlock(&mapping->i_mmap_lock);
2271
2272    return 1;
2273}
2274
2275static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2276            unsigned long address, pte_t *ptep, pte_t pte,
2277            struct page *pagecache_page)
2278{
2279    struct hstate *h = hstate_vma(vma);
2280    struct page *old_page, *new_page;
2281    int avoidcopy;
2282    int outside_reserve = 0;
2283
2284    old_page = pte_page(pte);
2285
2286retry_avoidcopy:
2287    /* If no-one else is actually using this page, avoid the copy
2288     * and just make the page writable */
2289    avoidcopy = (page_count(old_page) == 1);
2290    if (avoidcopy) {
2291        set_huge_ptep_writable(vma, address, ptep);
2292        return 0;
2293    }
2294
2295    /*
2296     * If the process that created a MAP_PRIVATE mapping is about to
2297     * perform a COW due to a shared page count, attempt to satisfy
2298     * the allocation without using the existing reserves. The pagecache
2299     * page is used to determine if the reserve at this address was
2300     * consumed or not. If reserves were used, a partial faulted mapping
2301     * at the time of fork() could consume its reserves on COW instead
2302     * of the full address range.
2303     */
2304    if (!(vma->vm_flags & VM_MAYSHARE) &&
2305            is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2306            old_page != pagecache_page)
2307        outside_reserve = 1;
2308
2309    page_cache_get(old_page);
2310
2311    /* Drop page_table_lock as buddy allocator may be called */
2312    spin_unlock(&mm->page_table_lock);
2313    new_page = alloc_huge_page(vma, address, outside_reserve);
2314
2315    if (IS_ERR(new_page)) {
2316        page_cache_release(old_page);
2317
2318        /*
2319         * If a process owning a MAP_PRIVATE mapping fails to COW,
2320         * it is due to references held by a child and an insufficient
2321         * huge page pool. To guarantee the original mappers
2322         * reliability, unmap the page from child processes. The child
2323         * may get SIGKILLed if it later faults.
2324         */
2325        if (outside_reserve) {
2326            BUG_ON(huge_pte_none(pte));
2327            if (unmap_ref_private(mm, vma, old_page, address)) {
2328                BUG_ON(page_count(old_page) != 1);
2329                BUG_ON(huge_pte_none(pte));
2330                spin_lock(&mm->page_table_lock);
2331                goto retry_avoidcopy;
2332            }
2333            WARN_ON_ONCE(1);
2334        }
2335
2336        /* Caller expects lock to be held */
2337        spin_lock(&mm->page_table_lock);
2338        return -PTR_ERR(new_page);
2339    }
2340
2341    copy_huge_page(new_page, old_page, address, vma);
2342    __SetPageUptodate(new_page);
2343
2344    /*
2345     * Retake the page_table_lock to check for racing updates
2346     * before the page tables are altered
2347     */
2348    spin_lock(&mm->page_table_lock);
2349    ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2350    if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2351        /* Break COW */
2352        huge_ptep_clear_flush(vma, address, ptep);
2353        set_huge_pte_at(mm, address, ptep,
2354                make_huge_pte(vma, new_page, 1));
2355        /* Make the old page be freed below */
2356        new_page = old_page;
2357    }
2358    page_cache_release(new_page);
2359    page_cache_release(old_page);
2360    return 0;
2361}
2362
2363/* Return the pagecache page at a given address within a VMA */
2364static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2365            struct vm_area_struct *vma, unsigned long address)
2366{
2367    struct address_space *mapping;
2368    pgoff_t idx;
2369
2370    mapping = vma->vm_file->f_mapping;
2371    idx = vma_hugecache_offset(h, vma, address);
2372
2373    return find_lock_page(mapping, idx);
2374}
2375
2376/*
2377 * Return whether there is a pagecache page to back given address within VMA.
2378 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2379 */
2380static bool hugetlbfs_pagecache_present(struct hstate *h,
2381            struct vm_area_struct *vma, unsigned long address)
2382{
2383    struct address_space *mapping;
2384    pgoff_t idx;
2385    struct page *page;
2386
2387    mapping = vma->vm_file->f_mapping;
2388    idx = vma_hugecache_offset(h, vma, address);
2389
2390    page = find_get_page(mapping, idx);
2391    if (page)
2392        put_page(page);
2393    return page != NULL;
2394}
2395
2396static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2397            unsigned long address, pte_t *ptep, unsigned int flags)
2398{
2399    struct hstate *h = hstate_vma(vma);
2400    int ret = VM_FAULT_SIGBUS;
2401    pgoff_t idx;
2402    unsigned long size;
2403    struct page *page;
2404    struct address_space *mapping;
2405    pte_t new_pte;
2406
2407    /*
2408     * Currently, we are forced to kill the process in the event the
2409     * original mapper has unmapped pages from the child due to a failed
2410     * COW. Warn that such a situation has occured as it may not be obvious
2411     */
2412    if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2413        printk(KERN_WARNING
2414            "PID %d killed due to inadequate hugepage pool\n",
2415            current->pid);
2416        return ret;
2417    }
2418
2419    mapping = vma->vm_file->f_mapping;
2420    idx = vma_hugecache_offset(h, vma, address);
2421
2422    /*
2423     * Use page lock to guard against racing truncation
2424     * before we get page_table_lock.
2425     */
2426retry:
2427    page = find_lock_page(mapping, idx);
2428    if (!page) {
2429        size = i_size_read(mapping->host) >> huge_page_shift(h);
2430        if (idx >= size)
2431            goto out;
2432        page = alloc_huge_page(vma, address, 0);
2433        if (IS_ERR(page)) {
2434            ret = -PTR_ERR(page);
2435            goto out;
2436        }
2437        clear_huge_page(page, address, huge_page_size(h));
2438        __SetPageUptodate(page);
2439
2440        if (vma->vm_flags & VM_MAYSHARE) {
2441            int err;
2442            struct inode *inode = mapping->host;
2443
2444            err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2445            if (err) {
2446                put_page(page);
2447                if (err == -EEXIST)
2448                    goto retry;
2449                goto out;
2450            }
2451
2452            spin_lock(&inode->i_lock);
2453            inode->i_blocks += blocks_per_huge_page(h);
2454            spin_unlock(&inode->i_lock);
2455        } else {
2456            lock_page(page);
2457            page->mapping = HUGETLB_POISON;
2458        }
2459    }
2460
2461    /*
2462     * If we are going to COW a private mapping later, we examine the
2463     * pending reservations for this page now. This will ensure that
2464     * any allocations necessary to record that reservation occur outside
2465     * the spinlock.
2466     */
2467    if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2468        if (vma_needs_reservation(h, vma, address) < 0) {
2469            ret = VM_FAULT_OOM;
2470            goto backout_unlocked;
2471        }
2472
2473    spin_lock(&mm->page_table_lock);
2474    size = i_size_read(mapping->host) >> huge_page_shift(h);
2475    if (idx >= size)
2476        goto backout;
2477
2478    ret = 0;
2479    if (!huge_pte_none(huge_ptep_get(ptep)))
2480        goto backout;
2481
2482    new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2483                && (vma->vm_flags & VM_SHARED)));
2484    set_huge_pte_at(mm, address, ptep, new_pte);
2485
2486    if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2487        /* Optimization, do the COW without a second fault */
2488        ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2489    }
2490
2491    spin_unlock(&mm->page_table_lock);
2492    unlock_page(page);
2493out:
2494    return ret;
2495
2496backout:
2497    spin_unlock(&mm->page_table_lock);
2498backout_unlocked:
2499    unlock_page(page);
2500    put_page(page);
2501    goto out;
2502}
2503
2504int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2505            unsigned long address, unsigned int flags)
2506{
2507    pte_t *ptep;
2508    pte_t entry;
2509    int ret;
2510    struct page *pagecache_page = NULL;
2511    static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2512    struct hstate *h = hstate_vma(vma);
2513
2514    ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2515    if (!ptep)
2516        return VM_FAULT_OOM;
2517
2518    /*
2519     * Serialize hugepage allocation and instantiation, so that we don't
2520     * get spurious allocation failures if two CPUs race to instantiate
2521     * the same page in the page cache.
2522     */
2523    mutex_lock(&hugetlb_instantiation_mutex);
2524    entry = huge_ptep_get(ptep);
2525    if (huge_pte_none(entry)) {
2526        ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2527        goto out_mutex;
2528    }
2529
2530    ret = 0;
2531
2532    /*
2533     * If we are going to COW the mapping later, we examine the pending
2534     * reservations for this page now. This will ensure that any
2535     * allocations necessary to record that reservation occur outside the
2536     * spinlock. For private mappings, we also lookup the pagecache
2537     * page now as it is used to determine if a reservation has been
2538     * consumed.
2539     */
2540    if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2541        if (vma_needs_reservation(h, vma, address) < 0) {
2542            ret = VM_FAULT_OOM;
2543            goto out_mutex;
2544        }
2545
2546        if (!(vma->vm_flags & VM_MAYSHARE))
2547            pagecache_page = hugetlbfs_pagecache_page(h,
2548                                vma, address);
2549    }
2550
2551    spin_lock(&mm->page_table_lock);
2552    /* Check for a racing update before calling hugetlb_cow */
2553    if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2554        goto out_page_table_lock;
2555
2556
2557    if (flags & FAULT_FLAG_WRITE) {
2558        if (!pte_write(entry)) {
2559            ret = hugetlb_cow(mm, vma, address, ptep, entry,
2560                            pagecache_page);
2561            goto out_page_table_lock;
2562        }
2563        entry = pte_mkdirty(entry);
2564    }
2565    entry = pte_mkyoung(entry);
2566    if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2567                        flags & FAULT_FLAG_WRITE))
2568        update_mmu_cache(vma, address, ptep);
2569
2570out_page_table_lock:
2571    spin_unlock(&mm->page_table_lock);
2572
2573    if (pagecache_page) {
2574        unlock_page(pagecache_page);
2575        put_page(pagecache_page);
2576    }
2577
2578out_mutex:
2579    mutex_unlock(&hugetlb_instantiation_mutex);
2580
2581    return ret;
2582}
2583
2584/* Can be overriden by architectures */
2585__attribute__((weak)) struct page *
2586follow_huge_pud(struct mm_struct *mm, unsigned long address,
2587           pud_t *pud, int write)
2588{
2589    BUG();
2590    return NULL;
2591}
2592
2593int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2594            struct page **pages, struct vm_area_struct **vmas,
2595            unsigned long *position, int *length, int i,
2596            unsigned int flags)
2597{
2598    unsigned long pfn_offset;
2599    unsigned long vaddr = *position;
2600    int remainder = *length;
2601    struct hstate *h = hstate_vma(vma);
2602
2603    spin_lock(&mm->page_table_lock);
2604    while (vaddr < vma->vm_end && remainder) {
2605        pte_t *pte;
2606        int absent;
2607        struct page *page;
2608
2609        /*
2610         * Some archs (sparc64, sh*) have multiple pte_ts to
2611         * each hugepage. We have to make sure we get the
2612         * first, for the page indexing below to work.
2613         */
2614        pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2615        absent = !pte || huge_pte_none(huge_ptep_get(pte));
2616
2617        /*
2618         * When coredumping, it suits get_dump_page if we just return
2619         * an error where there's an empty slot with no huge pagecache
2620         * to back it. This way, we avoid allocating a hugepage, and
2621         * the sparse dumpfile avoids allocating disk blocks, but its
2622         * huge holes still show up with zeroes where they need to be.
2623         */
2624        if (absent && (flags & FOLL_DUMP) &&
2625            !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2626            remainder = 0;
2627            break;
2628        }
2629
2630        if (absent ||
2631            ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2632            int ret;
2633
2634            spin_unlock(&mm->page_table_lock);
2635            ret = hugetlb_fault(mm, vma, vaddr,
2636                (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2637            spin_lock(&mm->page_table_lock);
2638            if (!(ret & VM_FAULT_ERROR))
2639                continue;
2640
2641            remainder = 0;
2642            break;
2643        }
2644
2645        pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2646        page = pte_page(huge_ptep_get(pte));
2647same_page:
2648        if (pages) {
2649            pages[i] = mem_map_offset(page, pfn_offset);
2650            get_page(pages[i]);
2651        }
2652
2653        if (vmas)
2654            vmas[i] = vma;
2655
2656        vaddr += PAGE_SIZE;
2657        ++pfn_offset;
2658        --remainder;
2659        ++i;
2660        if (vaddr < vma->vm_end && remainder &&
2661                pfn_offset < pages_per_huge_page(h)) {
2662            /*
2663             * We use pfn_offset to avoid touching the pageframes
2664             * of this compound page.
2665             */
2666            goto same_page;
2667        }
2668    }
2669    spin_unlock(&mm->page_table_lock);
2670    *length = remainder;
2671    *position = vaddr;
2672
2673    return i ? i : -EFAULT;
2674}
2675
2676void hugetlb_change_protection(struct vm_area_struct *vma,
2677        unsigned long address, unsigned long end, pgprot_t newprot)
2678{
2679    struct mm_struct *mm = vma->vm_mm;
2680    unsigned long start = address;
2681    pte_t *ptep;
2682    pte_t pte;
2683    struct hstate *h = hstate_vma(vma);
2684
2685    BUG_ON(address >= end);
2686    flush_cache_range(vma, address, end);
2687
2688    spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2689    spin_lock(&mm->page_table_lock);
2690    for (; address < end; address += huge_page_size(h)) {
2691        ptep = huge_pte_offset(mm, address);
2692        if (!ptep)
2693            continue;
2694        if (huge_pmd_unshare(mm, &address, ptep))
2695            continue;
2696        if (!huge_pte_none(huge_ptep_get(ptep))) {
2697            pte = huge_ptep_get_and_clear(mm, address, ptep);
2698            pte = pte_mkhuge(pte_modify(pte, newprot));
2699            set_huge_pte_at(mm, address, ptep, pte);
2700        }
2701    }
2702    spin_unlock(&mm->page_table_lock);
2703    spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2704
2705    flush_tlb_range(vma, start, end);
2706}
2707
2708int hugetlb_reserve_pages(struct inode *inode,
2709                    long from, long to,
2710                    struct vm_area_struct *vma,
2711                    int acctflag)
2712{
2713    long ret, chg;
2714    struct hstate *h = hstate_inode(inode);
2715
2716    /*
2717     * Only apply hugepage reservation if asked. At fault time, an
2718     * attempt will be made for VM_NORESERVE to allocate a page
2719     * and filesystem quota without using reserves
2720     */
2721    if (acctflag & VM_NORESERVE)
2722        return 0;
2723
2724    /*
2725     * Shared mappings base their reservation on the number of pages that
2726     * are already allocated on behalf of the file. Private mappings need
2727     * to reserve the full area even if read-only as mprotect() may be
2728     * called to make the mapping read-write. Assume !vma is a shm mapping
2729     */
2730    if (!vma || vma->vm_flags & VM_MAYSHARE)
2731        chg = region_chg(&inode->i_mapping->private_list, from, to);
2732    else {
2733        struct resv_map *resv_map = resv_map_alloc();
2734        if (!resv_map)
2735            return -ENOMEM;
2736
2737        chg = to - from;
2738
2739        set_vma_resv_map(vma, resv_map);
2740        set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2741    }
2742
2743    if (chg < 0)
2744        return chg;
2745
2746    /* There must be enough filesystem quota for the mapping */
2747    if (hugetlb_get_quota(inode->i_mapping, chg))
2748        return -ENOSPC;
2749
2750    /*
2751     * Check enough hugepages are available for the reservation.
2752     * Hand back the quota if there are not
2753     */
2754    ret = hugetlb_acct_memory(h, chg);
2755    if (ret < 0) {
2756        hugetlb_put_quota(inode->i_mapping, chg);
2757        return ret;
2758    }
2759
2760    /*
2761     * Account for the reservations made. Shared mappings record regions
2762     * that have reservations as they are shared by multiple VMAs.
2763     * When the last VMA disappears, the region map says how much
2764     * the reservation was and the page cache tells how much of
2765     * the reservation was consumed. Private mappings are per-VMA and
2766     * only the consumed reservations are tracked. When the VMA
2767     * disappears, the original reservation is the VMA size and the
2768     * consumed reservations are stored in the map. Hence, nothing
2769     * else has to be done for private mappings here
2770     */
2771    if (!vma || vma->vm_flags & VM_MAYSHARE)
2772        region_add(&inode->i_mapping->private_list, from, to);
2773    return 0;
2774}
2775
2776void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2777{
2778    struct hstate *h = hstate_inode(inode);
2779    long chg = region_truncate(&inode->i_mapping->private_list, offset);
2780
2781    spin_lock(&inode->i_lock);
2782    inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2783    spin_unlock(&inode->i_lock);
2784
2785    hugetlb_put_quota(inode->i_mapping, (chg - freed));
2786    hugetlb_acct_memory(h, -(chg - freed));
2787}
2788

Archive Download this file



interactive