mm/hugetlb.c

   1 /*
   2  * Generic hugetlb support.
   3  * (C) William Irwin, April 2004
   4  */
   5 #include <linux/gfp.h>
   6 #include <linux/list.h>
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/sysctl.h>
  12 #include <linux/highmem.h>
  13
  14 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  15 static unsigned long nr_huge_pages, free_huge_pages;
  16 unsigned long max_huge_pages;
  17 static struct list_head hugepage_freelists[MAX_NUMNODES];
  18 static spinlock_t hugetlb_lock = SPIN_LOCK_UNLOCKED;
  19
  20 static void enqueue_huge_page(struct page *page)
  21 {
  22         list_add(&page->lru,
  23                  &hugepage_freelists[page_zone(page)->zone_pgdat->node_id]);
  24 }
  25
  26 static struct page *dequeue_huge_page(void)
  27 {
  28         int nid = numa_node_id();
  29         struct page *page = NULL;
  30
  31         if (list_empty(&hugepage_freelists[nid])) {
  32                 for (nid = 0; nid < MAX_NUMNODES; ++nid)
  33                         if (!list_empty(&hugepage_freelists[nid]))
  34                                 break;
  35         }
  36         if (nid >= 0 && nid < MAX_NUMNODES &&
  37             !list_empty(&hugepage_freelists[nid])) {
  38                 page = list_entry(hugepage_freelists[nid].next,
  39                                   struct page, lru);
  40                 list_del(&page->lru);
  41         }
  42         return page;
  43 }
  44
  45 static struct page *alloc_fresh_huge_page(void)
  46 {
  47         static int nid = 0;
  48         struct page *page;
  49         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP,
  50                                         HUGETLB_PAGE_ORDER);
  51         nid = (nid + 1) % numnodes;
  52         return page;
  53 }
  54
  55 void free_huge_page(struct page *page)
  56 {
  57         BUG_ON(page_count(page));
  58
  59         INIT_LIST_HEAD(&page->lru);
  60
  61         spin_lock(&hugetlb_lock);
  62         enqueue_huge_page(page);
  63         free_huge_pages++;
  64         spin_unlock(&hugetlb_lock);
  65 }
  66
  67 struct page *alloc_huge_page(void)
  68 {
  69         struct page *page;
  70         int i;
  71
  72         spin_lock(&hugetlb_lock);
  73         page = dequeue_huge_page();
  74         if (!page) {
  75                 spin_unlock(&hugetlb_lock);
  76                 return NULL;
  77         }
  78         free_huge_pages--;
  79         spin_unlock(&hugetlb_lock);
  80         set_page_count(page, 1);
  81         page[1].mapping = (void *)free_huge_page;
  82         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
  83                 clear_highpage(&page[i]);
  84         return page;
  85 }
  86
  87 static int __init hugetlb_init(void)
  88 {
  89         unsigned long i;
  90         struct page *page;
  91
  92         for (i = 0; i < MAX_NUMNODES; ++i)
  93                 INIT_LIST_HEAD(&hugepage_freelists[i]);
  94
  95         for (i = 0; i < max_huge_pages; ++i) {
  96                 page = alloc_fresh_huge_page();
  97                 if (!page)
  98                         break;
  99                 spin_lock(&hugetlb_lock);
 100                 enqueue_huge_page(page);
 101                 spin_unlock(&hugetlb_lock);
 102         }
 103         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 104         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 105         return 0;
 106 }
 107 module_init(hugetlb_init);
 108
 109 static int __init hugetlb_setup(char *s)
 110 {
 111         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 112                 max_huge_pages = 0;
 113         return 1;
 114 }
 115 __setup("hugepages=", hugetlb_setup);
 116
 117 static void update_and_free_page(struct page *page)
 118 {
 119         int i;
 120         nr_huge_pages--;
 121         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 122                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 123                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 124                                 1 << PG_private | 1<< PG_writeback);
 125                 set_page_count(&page[i], 0);
 126         }
 127         set_page_count(page, 1);
 128         __free_pages(page, HUGETLB_PAGE_ORDER);
 129 }
 130
 131 #ifdef CONFIG_HIGHMEM
 132 static int try_to_free_low(unsigned long count)
 133 {
 134         int i;
 135         for (i = 0; i < MAX_NUMNODES; ++i) {
 136                 struct page *page;
 137                 list_for_each_entry(page, &hugepage_freelists[i], lru) {
 138                         if (PageHighMem(page))
 139                                 continue;
 140                         list_del(&page->lru);
 141                         update_and_free_page(page);
 142                         --free_huge_pages;
 143                         if (!--count)
 144                                 return 0;
 145                 }
 146         }
 147         return count;
 148 }
 149 #else
 150 static inline int try_to_free_low(unsigned long count)
 151 {
 152         return count;
 153 }
 154 #endif
 155
 156 static unsigned long set_max_huge_pages(unsigned long count)
 157 {
 158         while (count > nr_huge_pages) {
 159                 struct page *page = alloc_fresh_huge_page();
 160                 if (!page)
 161                         return nr_huge_pages;
 162                 spin_lock(&hugetlb_lock);
 163                 enqueue_huge_page(page);
 164                 free_huge_pages++;
 165                 nr_huge_pages++;
 166                 spin_unlock(&hugetlb_lock);
 167         }
 168         if (count >= nr_huge_pages)
 169                 return nr_huge_pages;
 170
 171         spin_lock(&hugetlb_lock);
 172         for (count = try_to_free_low(count); count < nr_huge_pages; --free_huge_pages) {
 173                 struct page *page = dequeue_huge_page();
 174                 if (!page)
 175                         break;
 176                 update_and_free_page(page);
 177         }
 178         spin_unlock(&hugetlb_lock);
 179         return nr_huge_pages;
 180 }
 181
 182 #ifdef CONFIG_SYSCTL
 183 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 184                            struct file *file, void *buffer, size_t *length)
 185 {
 186         proc_doulongvec_minmax(table, write, file, buffer, length);
 187         max_huge_pages = set_max_huge_pages(max_huge_pages);
 188         return 0;
 189 }
 190 #endif /* CONFIG_SYSCTL */
 191
 192 int hugetlb_report_meminfo(char *buf)
 193 {
 194         return sprintf(buf,
 195                         "HugePages_Total: %5lu\n"
 196                         "HugePages_Free:  %5lu\n"
 197                         "Hugepagesize:    %5lu kB\n",
 198                         nr_huge_pages,
 199                         free_huge_pages,
 200                         HPAGE_SIZE/1024);
 201 }
 202
 203 int is_hugepage_mem_enough(size_t size)
 204 {
 205         return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
 206 }
 207
 208 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 209 unsigned long hugetlb_total_pages(void)
 210 {
 211         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 212 }
 213 EXPORT_SYMBOL(hugetlb_total_pages);
 214
 215 /*
 216  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 217  * handle_mm_fault() to try to instantiate regular-sized pages in the
 218  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 219  * this far.
 220  */
 221 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
 222                                 unsigned long address, int *unused)
 223 {
 224         BUG();
 225         return NULL;
 226 }
 227
 228 struct vm_operations_struct hugetlb_vm_ops = {
 229         .nopage = hugetlb_nopage,
 230 };
 231
 232 void zap_hugepage_range(struct vm_area_struct *vma,
 233                         unsigned long start, unsigned long length)
 234 {
 235         struct mm_struct *mm = vma->vm_mm;
 236
 237         spin_lock(&mm->page_table_lock);
 238         unmap_hugepage_range(vma, start, start + length);
 239         spin_unlock(&mm->page_table_lock);
 240 }