11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * hugetlbpage-backed filesystem. Based on ramfs. 31da177e4SLinus Torvalds * 41da177e4SLinus Torvalds * William Irwin, 2002 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Copyright (C) 2002 Linus Torvalds. 71da177e4SLinus Torvalds */ 81da177e4SLinus Torvalds 91da177e4SLinus Torvalds #include <linux/module.h> 101da177e4SLinus Torvalds #include <linux/thread_info.h> 111da177e4SLinus Torvalds #include <asm/current.h> 121da177e4SLinus Torvalds #include <linux/sched.h> /* remove ASAP */ 131da177e4SLinus Torvalds #include <linux/fs.h> 141da177e4SLinus Torvalds #include <linux/mount.h> 151da177e4SLinus Torvalds #include <linux/file.h> 16e73a75faSRandy Dunlap #include <linux/kernel.h> 171da177e4SLinus Torvalds #include <linux/writeback.h> 181da177e4SLinus Torvalds #include <linux/pagemap.h> 191da177e4SLinus Torvalds #include <linux/highmem.h> 201da177e4SLinus Torvalds #include <linux/init.h> 211da177e4SLinus Torvalds #include <linux/string.h> 2216f7e0feSRandy Dunlap #include <linux/capability.h> 23e73a75faSRandy Dunlap #include <linux/ctype.h> 241da177e4SLinus Torvalds #include <linux/backing-dev.h> 251da177e4SLinus Torvalds #include <linux/hugetlb.h> 261da177e4SLinus Torvalds #include <linux/pagevec.h> 27e73a75faSRandy Dunlap #include <linux/parser.h> 28036e0856SBenjamin Herrenschmidt #include <linux/mman.h> 291da177e4SLinus Torvalds #include <linux/slab.h> 301da177e4SLinus Torvalds #include <linux/dnotify.h> 311da177e4SLinus Torvalds #include <linux/statfs.h> 321da177e4SLinus Torvalds #include <linux/security.h> 331fd7317dSNick Black #include <linux/magic.h> 34290408d4SNaoya Horiguchi #include <linux/migrate.h> 351da177e4SLinus Torvalds 361da177e4SLinus Torvalds #include <asm/uaccess.h> 371da177e4SLinus Torvalds 38ee9b6d61SJosef 'Jeff' Sipek static const struct super_operations hugetlbfs_ops; 39f5e54d6eSChristoph Hellwig static const struct address_space_operations hugetlbfs_aops; 404b6f5d20SArjan van de Ven const struct file_operations hugetlbfs_file_operations; 4192e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_dir_inode_operations; 4292e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_inode_operations; 431da177e4SLinus Torvalds 44a1d776eeSDavid Gibson struct hugetlbfs_config { 45a1d776eeSDavid Gibson uid_t uid; 46a1d776eeSDavid Gibson gid_t gid; 47a1d776eeSDavid Gibson umode_t mode; 48a1d776eeSDavid Gibson long nr_blocks; 49a1d776eeSDavid Gibson long nr_inodes; 50a1d776eeSDavid Gibson struct hstate *hstate; 51a1d776eeSDavid Gibson }; 52a1d776eeSDavid Gibson 53a1d776eeSDavid Gibson struct hugetlbfs_inode_info { 54a1d776eeSDavid Gibson struct shared_policy policy; 55a1d776eeSDavid Gibson struct inode vfs_inode; 56a1d776eeSDavid Gibson }; 57a1d776eeSDavid Gibson 58a1d776eeSDavid Gibson static inline struct hugetlbfs_inode_info *HUGETLBFS_I(struct inode *inode) 59a1d776eeSDavid Gibson { 60a1d776eeSDavid Gibson return container_of(inode, struct hugetlbfs_inode_info, vfs_inode); 61a1d776eeSDavid Gibson } 62a1d776eeSDavid Gibson 631da177e4SLinus Torvalds static struct backing_dev_info hugetlbfs_backing_dev_info = { 64d993831fSJens Axboe .name = "hugetlbfs", 651da177e4SLinus Torvalds .ra_pages = 0, /* No readahead */ 66e4ad08feSMiklos Szeredi .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK, 671da177e4SLinus Torvalds }; 681da177e4SLinus Torvalds 691da177e4SLinus Torvalds int sysctl_hugetlb_shm_group; 701da177e4SLinus Torvalds 71e73a75faSRandy Dunlap enum { 72e73a75faSRandy Dunlap Opt_size, Opt_nr_inodes, 73e73a75faSRandy Dunlap Opt_mode, Opt_uid, Opt_gid, 74a137e1ccSAndi Kleen Opt_pagesize, 75e73a75faSRandy Dunlap Opt_err, 76e73a75faSRandy Dunlap }; 77e73a75faSRandy Dunlap 78a447c093SSteven Whitehouse static const match_table_t tokens = { 79e73a75faSRandy Dunlap {Opt_size, "size=%s"}, 80e73a75faSRandy Dunlap {Opt_nr_inodes, "nr_inodes=%s"}, 81e73a75faSRandy Dunlap {Opt_mode, "mode=%o"}, 82e73a75faSRandy Dunlap {Opt_uid, "uid=%u"}, 83e73a75faSRandy Dunlap {Opt_gid, "gid=%u"}, 84a137e1ccSAndi Kleen {Opt_pagesize, "pagesize=%s"}, 85e73a75faSRandy Dunlap {Opt_err, NULL}, 86e73a75faSRandy Dunlap }; 87e73a75faSRandy Dunlap 882e9b367cSAdam Litke static void huge_pagevec_release(struct pagevec *pvec) 892e9b367cSAdam Litke { 902e9b367cSAdam Litke int i; 912e9b367cSAdam Litke 922e9b367cSAdam Litke for (i = 0; i < pagevec_count(pvec); ++i) 932e9b367cSAdam Litke put_page(pvec->pages[i]); 942e9b367cSAdam Litke 952e9b367cSAdam Litke pagevec_reinit(pvec); 962e9b367cSAdam Litke } 972e9b367cSAdam Litke 981da177e4SLinus Torvalds static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma) 991da177e4SLinus Torvalds { 100b39424e2SJosef Sipek struct inode *inode = file->f_path.dentry->d_inode; 1011da177e4SLinus Torvalds loff_t len, vma_len; 1021da177e4SLinus Torvalds int ret; 103a5516438SAndi Kleen struct hstate *h = hstate_file(file); 1041da177e4SLinus Torvalds 10568589bc3SHugh Dickins /* 106dec4ad86SDavid Gibson * vma address alignment (but not the pgoff alignment) has 107dec4ad86SDavid Gibson * already been checked by prepare_hugepage_range. If you add 108dec4ad86SDavid Gibson * any error returns here, do so after setting VM_HUGETLB, so 109dec4ad86SDavid Gibson * is_vm_hugetlb_page tests below unmap_region go the right 110dec4ad86SDavid Gibson * way when do_mmap_pgoff unwinds (may be important on powerpc 111dec4ad86SDavid Gibson * and ia64). 11268589bc3SHugh Dickins */ 11368589bc3SHugh Dickins vma->vm_flags |= VM_HUGETLB | VM_RESERVED; 11468589bc3SHugh Dickins vma->vm_ops = &hugetlb_vm_ops; 1151da177e4SLinus Torvalds 1162b37c35eSBecky Bruce if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT)) 117dec4ad86SDavid Gibson return -EINVAL; 118dec4ad86SDavid Gibson 1191da177e4SLinus Torvalds vma_len = (loff_t)(vma->vm_end - vma->vm_start); 1201da177e4SLinus Torvalds 1211b1dcc1bSJes Sorensen mutex_lock(&inode->i_mutex); 1221da177e4SLinus Torvalds file_accessed(file); 1231da177e4SLinus Torvalds 1241da177e4SLinus Torvalds ret = -ENOMEM; 1251da177e4SLinus Torvalds len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); 1261da177e4SLinus Torvalds 127a1e78772SMel Gorman if (hugetlb_reserve_pages(inode, 128a5516438SAndi Kleen vma->vm_pgoff >> huge_page_order(h), 1295a6fe125SMel Gorman len >> huge_page_shift(h), vma, 1305a6fe125SMel Gorman vma->vm_flags)) 131b45b5bd6SDavid Gibson goto out; 132b45b5bd6SDavid Gibson 1334c887265SAdam Litke ret = 0; 1344c887265SAdam Litke hugetlb_prefault_arch_hook(vma->vm_mm); 135b6174df5SZhang, Yanmin if (vma->vm_flags & VM_WRITE && inode->i_size < len) 1361da177e4SLinus Torvalds inode->i_size = len; 1371da177e4SLinus Torvalds out: 1381b1dcc1bSJes Sorensen mutex_unlock(&inode->i_mutex); 1391da177e4SLinus Torvalds 1401da177e4SLinus Torvalds return ret; 1411da177e4SLinus Torvalds } 1421da177e4SLinus Torvalds 1431da177e4SLinus Torvalds /* 144508034a3SHugh Dickins * Called under down_write(mmap_sem). 1451da177e4SLinus Torvalds */ 1461da177e4SLinus Torvalds 147d2ba27e8SAdrian Bunk #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA 1481da177e4SLinus Torvalds static unsigned long 1491da177e4SLinus Torvalds hugetlb_get_unmapped_area(struct file *file, unsigned long addr, 1501da177e4SLinus Torvalds unsigned long len, unsigned long pgoff, unsigned long flags) 1511da177e4SLinus Torvalds { 1521da177e4SLinus Torvalds struct mm_struct *mm = current->mm; 1531da177e4SLinus Torvalds struct vm_area_struct *vma; 1541da177e4SLinus Torvalds unsigned long start_addr; 155a5516438SAndi Kleen struct hstate *h = hstate_file(file); 1561da177e4SLinus Torvalds 157a5516438SAndi Kleen if (len & ~huge_page_mask(h)) 1581da177e4SLinus Torvalds return -EINVAL; 1591da177e4SLinus Torvalds if (len > TASK_SIZE) 1601da177e4SLinus Torvalds return -ENOMEM; 1611da177e4SLinus Torvalds 162036e0856SBenjamin Herrenschmidt if (flags & MAP_FIXED) { 163a5516438SAndi Kleen if (prepare_hugepage_range(file, addr, len)) 164036e0856SBenjamin Herrenschmidt return -EINVAL; 165036e0856SBenjamin Herrenschmidt return addr; 166036e0856SBenjamin Herrenschmidt } 167036e0856SBenjamin Herrenschmidt 1681da177e4SLinus Torvalds if (addr) { 169a5516438SAndi Kleen addr = ALIGN(addr, huge_page_size(h)); 1701da177e4SLinus Torvalds vma = find_vma(mm, addr); 1711da177e4SLinus Torvalds if (TASK_SIZE - len >= addr && 1721da177e4SLinus Torvalds (!vma || addr + len <= vma->vm_start)) 1731da177e4SLinus Torvalds return addr; 1741da177e4SLinus Torvalds } 1751da177e4SLinus Torvalds 1764bfc130dSXiao Guangrong if (len > mm->cached_hole_size) 1771da177e4SLinus Torvalds start_addr = mm->free_area_cache; 1784bfc130dSXiao Guangrong else { 1791363c3cdSWolfgang Wander start_addr = TASK_UNMAPPED_BASE; 1804bfc130dSXiao Guangrong mm->cached_hole_size = 0; 1814bfc130dSXiao Guangrong } 1821363c3cdSWolfgang Wander 1831da177e4SLinus Torvalds full_search: 184a5516438SAndi Kleen addr = ALIGN(start_addr, huge_page_size(h)); 1851da177e4SLinus Torvalds 1861da177e4SLinus Torvalds for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1871da177e4SLinus Torvalds /* At this point: (!vma || addr < vma->vm_end). */ 1881da177e4SLinus Torvalds if (TASK_SIZE - len < addr) { 1891da177e4SLinus Torvalds /* 1901da177e4SLinus Torvalds * Start a new search - just in case we missed 1911da177e4SLinus Torvalds * some holes. 1921da177e4SLinus Torvalds */ 1931da177e4SLinus Torvalds if (start_addr != TASK_UNMAPPED_BASE) { 1941da177e4SLinus Torvalds start_addr = TASK_UNMAPPED_BASE; 1954bfc130dSXiao Guangrong mm->cached_hole_size = 0; 1961da177e4SLinus Torvalds goto full_search; 1971da177e4SLinus Torvalds } 1981da177e4SLinus Torvalds return -ENOMEM; 1991da177e4SLinus Torvalds } 2001da177e4SLinus Torvalds 2014bfc130dSXiao Guangrong if (!vma || addr + len <= vma->vm_start) { 2024bfc130dSXiao Guangrong mm->free_area_cache = addr + len; 2031da177e4SLinus Torvalds return addr; 2044bfc130dSXiao Guangrong } 2054bfc130dSXiao Guangrong if (addr + mm->cached_hole_size < vma->vm_start) 2064bfc130dSXiao Guangrong mm->cached_hole_size = vma->vm_start - addr; 207a5516438SAndi Kleen addr = ALIGN(vma->vm_end, huge_page_size(h)); 2081da177e4SLinus Torvalds } 2091da177e4SLinus Torvalds } 2101da177e4SLinus Torvalds #endif 2111da177e4SLinus Torvalds 212e63e1e5aSBadari Pulavarty static int 213e63e1e5aSBadari Pulavarty hugetlbfs_read_actor(struct page *page, unsigned long offset, 214e63e1e5aSBadari Pulavarty char __user *buf, unsigned long count, 215e63e1e5aSBadari Pulavarty unsigned long size) 216e63e1e5aSBadari Pulavarty { 217e63e1e5aSBadari Pulavarty char *kaddr; 218e63e1e5aSBadari Pulavarty unsigned long left, copied = 0; 219e63e1e5aSBadari Pulavarty int i, chunksize; 220e63e1e5aSBadari Pulavarty 221e63e1e5aSBadari Pulavarty if (size > count) 222e63e1e5aSBadari Pulavarty size = count; 223e63e1e5aSBadari Pulavarty 224e63e1e5aSBadari Pulavarty /* Find which 4k chunk and offset with in that chunk */ 225e63e1e5aSBadari Pulavarty i = offset >> PAGE_CACHE_SHIFT; 226e63e1e5aSBadari Pulavarty offset = offset & ~PAGE_CACHE_MASK; 227e63e1e5aSBadari Pulavarty 228e63e1e5aSBadari Pulavarty while (size) { 229e63e1e5aSBadari Pulavarty chunksize = PAGE_CACHE_SIZE; 230e63e1e5aSBadari Pulavarty if (offset) 231e63e1e5aSBadari Pulavarty chunksize -= offset; 232e63e1e5aSBadari Pulavarty if (chunksize > size) 233e63e1e5aSBadari Pulavarty chunksize = size; 234e63e1e5aSBadari Pulavarty kaddr = kmap(&page[i]); 235e63e1e5aSBadari Pulavarty left = __copy_to_user(buf, kaddr + offset, chunksize); 236e63e1e5aSBadari Pulavarty kunmap(&page[i]); 237e63e1e5aSBadari Pulavarty if (left) { 238e63e1e5aSBadari Pulavarty copied += (chunksize - left); 239e63e1e5aSBadari Pulavarty break; 240e63e1e5aSBadari Pulavarty } 241e63e1e5aSBadari Pulavarty offset = 0; 242e63e1e5aSBadari Pulavarty size -= chunksize; 243e63e1e5aSBadari Pulavarty buf += chunksize; 244e63e1e5aSBadari Pulavarty copied += chunksize; 245e63e1e5aSBadari Pulavarty i++; 246e63e1e5aSBadari Pulavarty } 247e63e1e5aSBadari Pulavarty return copied ? copied : -EFAULT; 248e63e1e5aSBadari Pulavarty } 249e63e1e5aSBadari Pulavarty 250e63e1e5aSBadari Pulavarty /* 251e63e1e5aSBadari Pulavarty * Support for read() - Find the page attached to f_mapping and copy out the 252e63e1e5aSBadari Pulavarty * data. Its *very* similar to do_generic_mapping_read(), we can't use that 253e63e1e5aSBadari Pulavarty * since it has PAGE_CACHE_SIZE assumptions. 254e63e1e5aSBadari Pulavarty */ 255e63e1e5aSBadari Pulavarty static ssize_t hugetlbfs_read(struct file *filp, char __user *buf, 256e63e1e5aSBadari Pulavarty size_t len, loff_t *ppos) 257e63e1e5aSBadari Pulavarty { 258a5516438SAndi Kleen struct hstate *h = hstate_file(filp); 259e63e1e5aSBadari Pulavarty struct address_space *mapping = filp->f_mapping; 260e63e1e5aSBadari Pulavarty struct inode *inode = mapping->host; 261a5516438SAndi Kleen unsigned long index = *ppos >> huge_page_shift(h); 262a5516438SAndi Kleen unsigned long offset = *ppos & ~huge_page_mask(h); 263e63e1e5aSBadari Pulavarty unsigned long end_index; 264e63e1e5aSBadari Pulavarty loff_t isize; 265e63e1e5aSBadari Pulavarty ssize_t retval = 0; 266e63e1e5aSBadari Pulavarty 267e63e1e5aSBadari Pulavarty /* validate length */ 268e63e1e5aSBadari Pulavarty if (len == 0) 269e63e1e5aSBadari Pulavarty goto out; 270e63e1e5aSBadari Pulavarty 271e63e1e5aSBadari Pulavarty for (;;) { 272e63e1e5aSBadari Pulavarty struct page *page; 273a5516438SAndi Kleen unsigned long nr, ret; 27491bf189cSRoel Kluin int ra; 275e63e1e5aSBadari Pulavarty 276e63e1e5aSBadari Pulavarty /* nr is the maximum number of bytes to copy from this page */ 277a5516438SAndi Kleen nr = huge_page_size(h); 278a05b0855SAneesh Kumar K.V isize = i_size_read(inode); 279a05b0855SAneesh Kumar K.V if (!isize) 280a05b0855SAneesh Kumar K.V goto out; 281a05b0855SAneesh Kumar K.V end_index = (isize - 1) >> huge_page_shift(h); 282e63e1e5aSBadari Pulavarty if (index >= end_index) { 283e63e1e5aSBadari Pulavarty if (index > end_index) 284e63e1e5aSBadari Pulavarty goto out; 285a5516438SAndi Kleen nr = ((isize - 1) & ~huge_page_mask(h)) + 1; 286a05b0855SAneesh Kumar K.V if (nr <= offset) 287e63e1e5aSBadari Pulavarty goto out; 288e63e1e5aSBadari Pulavarty } 289e63e1e5aSBadari Pulavarty nr = nr - offset; 290e63e1e5aSBadari Pulavarty 291e63e1e5aSBadari Pulavarty /* Find the page */ 292a05b0855SAneesh Kumar K.V page = find_lock_page(mapping, index); 293e63e1e5aSBadari Pulavarty if (unlikely(page == NULL)) { 294e63e1e5aSBadari Pulavarty /* 295e63e1e5aSBadari Pulavarty * We have a HOLE, zero out the user-buffer for the 296e63e1e5aSBadari Pulavarty * length of the hole or request. 297e63e1e5aSBadari Pulavarty */ 298e63e1e5aSBadari Pulavarty ret = len < nr ? len : nr; 299e63e1e5aSBadari Pulavarty if (clear_user(buf, ret)) 30091bf189cSRoel Kluin ra = -EFAULT; 30191bf189cSRoel Kluin else 30291bf189cSRoel Kluin ra = 0; 303e63e1e5aSBadari Pulavarty } else { 304a05b0855SAneesh Kumar K.V unlock_page(page); 305a05b0855SAneesh Kumar K.V 306e63e1e5aSBadari Pulavarty /* 307e63e1e5aSBadari Pulavarty * We have the page, copy it to user space buffer. 308e63e1e5aSBadari Pulavarty */ 30991bf189cSRoel Kluin ra = hugetlbfs_read_actor(page, offset, buf, len, nr); 31091bf189cSRoel Kluin ret = ra; 311a05b0855SAneesh Kumar K.V page_cache_release(page); 312e63e1e5aSBadari Pulavarty } 31391bf189cSRoel Kluin if (ra < 0) { 314e63e1e5aSBadari Pulavarty if (retval == 0) 31591bf189cSRoel Kluin retval = ra; 316e63e1e5aSBadari Pulavarty goto out; 317e63e1e5aSBadari Pulavarty } 318e63e1e5aSBadari Pulavarty 319e63e1e5aSBadari Pulavarty offset += ret; 320e63e1e5aSBadari Pulavarty retval += ret; 321e63e1e5aSBadari Pulavarty len -= ret; 322a5516438SAndi Kleen index += offset >> huge_page_shift(h); 323a5516438SAndi Kleen offset &= ~huge_page_mask(h); 324e63e1e5aSBadari Pulavarty 325e63e1e5aSBadari Pulavarty /* short read or no more work */ 326e63e1e5aSBadari Pulavarty if ((ret != nr) || (len == 0)) 327e63e1e5aSBadari Pulavarty break; 328e63e1e5aSBadari Pulavarty } 329e63e1e5aSBadari Pulavarty out: 330a5516438SAndi Kleen *ppos = ((loff_t)index << huge_page_shift(h)) + offset; 331e63e1e5aSBadari Pulavarty return retval; 332e63e1e5aSBadari Pulavarty } 333e63e1e5aSBadari Pulavarty 334800d15a5SNick Piggin static int hugetlbfs_write_begin(struct file *file, 335800d15a5SNick Piggin struct address_space *mapping, 336800d15a5SNick Piggin loff_t pos, unsigned len, unsigned flags, 337800d15a5SNick Piggin struct page **pagep, void **fsdata) 3381da177e4SLinus Torvalds { 3391da177e4SLinus Torvalds return -EINVAL; 3401da177e4SLinus Torvalds } 3411da177e4SLinus Torvalds 342800d15a5SNick Piggin static int hugetlbfs_write_end(struct file *file, struct address_space *mapping, 343800d15a5SNick Piggin loff_t pos, unsigned len, unsigned copied, 344800d15a5SNick Piggin struct page *page, void *fsdata) 3451da177e4SLinus Torvalds { 346800d15a5SNick Piggin BUG(); 3471da177e4SLinus Torvalds return -EINVAL; 3481da177e4SLinus Torvalds } 3491da177e4SLinus Torvalds 3501da177e4SLinus Torvalds static void truncate_huge_page(struct page *page) 3511da177e4SLinus Torvalds { 352fba2591bSLinus Torvalds cancel_dirty_page(page, /* No IO accounting for huge pages? */0); 3531da177e4SLinus Torvalds ClearPageUptodate(page); 354bd65cb86SMinchan Kim delete_from_page_cache(page); 3551da177e4SLinus Torvalds } 3561da177e4SLinus Torvalds 357b45b5bd6SDavid Gibson static void truncate_hugepages(struct inode *inode, loff_t lstart) 3581da177e4SLinus Torvalds { 359a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 360b45b5bd6SDavid Gibson struct address_space *mapping = &inode->i_data; 361a5516438SAndi Kleen const pgoff_t start = lstart >> huge_page_shift(h); 3621da177e4SLinus Torvalds struct pagevec pvec; 3631da177e4SLinus Torvalds pgoff_t next; 364a43a8c39SChen, Kenneth W int i, freed = 0; 3651da177e4SLinus Torvalds 3661da177e4SLinus Torvalds pagevec_init(&pvec, 0); 3671da177e4SLinus Torvalds next = start; 3681da177e4SLinus Torvalds while (1) { 3691da177e4SLinus Torvalds if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { 3701da177e4SLinus Torvalds if (next == start) 3711da177e4SLinus Torvalds break; 3721da177e4SLinus Torvalds next = start; 3731da177e4SLinus Torvalds continue; 3741da177e4SLinus Torvalds } 3751da177e4SLinus Torvalds 3761da177e4SLinus Torvalds for (i = 0; i < pagevec_count(&pvec); ++i) { 3771da177e4SLinus Torvalds struct page *page = pvec.pages[i]; 3781da177e4SLinus Torvalds 3791da177e4SLinus Torvalds lock_page(page); 3801da177e4SLinus Torvalds if (page->index > next) 3811da177e4SLinus Torvalds next = page->index; 3821da177e4SLinus Torvalds ++next; 3831da177e4SLinus Torvalds truncate_huge_page(page); 3841da177e4SLinus Torvalds unlock_page(page); 385a43a8c39SChen, Kenneth W freed++; 3861da177e4SLinus Torvalds } 3871da177e4SLinus Torvalds huge_pagevec_release(&pvec); 3881da177e4SLinus Torvalds } 3891da177e4SLinus Torvalds BUG_ON(!lstart && mapping->nrpages); 390a43a8c39SChen, Kenneth W hugetlb_unreserve_pages(inode, start, freed); 3911da177e4SLinus Torvalds } 3921da177e4SLinus Torvalds 3932bbbda30SAl Viro static void hugetlbfs_evict_inode(struct inode *inode) 3941da177e4SLinus Torvalds { 395b45b5bd6SDavid Gibson truncate_hugepages(inode, 0); 396dbd5768fSJan Kara clear_inode(inode); 397149f4211SChristoph Hellwig } 398149f4211SChristoph Hellwig 3991da177e4SLinus Torvalds static inline void 400856fc295SHugh Dickins hugetlb_vmtruncate_list(struct prio_tree_root *root, pgoff_t pgoff) 4011da177e4SLinus Torvalds { 4021da177e4SLinus Torvalds struct vm_area_struct *vma; 4031da177e4SLinus Torvalds struct prio_tree_iter iter; 4041da177e4SLinus Torvalds 405856fc295SHugh Dickins vma_prio_tree_foreach(vma, &iter, root, pgoff, ULONG_MAX) { 4061da177e4SLinus Torvalds unsigned long v_offset; 4071da177e4SLinus Torvalds 4081da177e4SLinus Torvalds /* 409856fc295SHugh Dickins * Can the expression below overflow on 32-bit arches? 410856fc295SHugh Dickins * No, because the prio_tree returns us only those vmas 411856fc295SHugh Dickins * which overlap the truncated area starting at pgoff, 412856fc295SHugh Dickins * and no vma on a 32-bit arch can span beyond the 4GB. 4131da177e4SLinus Torvalds */ 414856fc295SHugh Dickins if (vma->vm_pgoff < pgoff) 415856fc295SHugh Dickins v_offset = (pgoff - vma->vm_pgoff) << PAGE_SHIFT; 416856fc295SHugh Dickins else 4171da177e4SLinus Torvalds v_offset = 0; 4181da177e4SLinus Torvalds 419502717f4SChen, Kenneth W __unmap_hugepage_range(vma, 42004f2cbe3SMel Gorman vma->vm_start + v_offset, vma->vm_end, NULL); 4211da177e4SLinus Torvalds } 4221da177e4SLinus Torvalds } 4231da177e4SLinus Torvalds 4241da177e4SLinus Torvalds static int hugetlb_vmtruncate(struct inode *inode, loff_t offset) 4251da177e4SLinus Torvalds { 426856fc295SHugh Dickins pgoff_t pgoff; 4271da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 428a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 4291da177e4SLinus Torvalds 430a5516438SAndi Kleen BUG_ON(offset & ~huge_page_mask(h)); 431856fc295SHugh Dickins pgoff = offset >> PAGE_SHIFT; 4321da177e4SLinus Torvalds 4337aa91e10SKen Chen i_size_write(inode, offset); 4343d48ae45SPeter Zijlstra mutex_lock(&mapping->i_mmap_mutex); 4351da177e4SLinus Torvalds if (!prio_tree_empty(&mapping->i_mmap)) 4361da177e4SLinus Torvalds hugetlb_vmtruncate_list(&mapping->i_mmap, pgoff); 4373d48ae45SPeter Zijlstra mutex_unlock(&mapping->i_mmap_mutex); 438b45b5bd6SDavid Gibson truncate_hugepages(inode, offset); 4391da177e4SLinus Torvalds return 0; 4401da177e4SLinus Torvalds } 4411da177e4SLinus Torvalds 4421da177e4SLinus Torvalds static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr) 4431da177e4SLinus Torvalds { 4441da177e4SLinus Torvalds struct inode *inode = dentry->d_inode; 445a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 4461da177e4SLinus Torvalds int error; 4471da177e4SLinus Torvalds unsigned int ia_valid = attr->ia_valid; 4481da177e4SLinus Torvalds 4491da177e4SLinus Torvalds BUG_ON(!inode); 4501da177e4SLinus Torvalds 4511da177e4SLinus Torvalds error = inode_change_ok(inode, attr); 4521da177e4SLinus Torvalds if (error) 4531025774cSChristoph Hellwig return error; 4541da177e4SLinus Torvalds 4551da177e4SLinus Torvalds if (ia_valid & ATTR_SIZE) { 4561da177e4SLinus Torvalds error = -EINVAL; 4571025774cSChristoph Hellwig if (attr->ia_size & ~huge_page_mask(h)) 4581025774cSChristoph Hellwig return -EINVAL; 4591da177e4SLinus Torvalds error = hugetlb_vmtruncate(inode, attr->ia_size); 4601da177e4SLinus Torvalds if (error) 4611da177e4SLinus Torvalds return error; 4621da177e4SLinus Torvalds } 4631da177e4SLinus Torvalds 4641025774cSChristoph Hellwig setattr_copy(inode, attr); 4651025774cSChristoph Hellwig mark_inode_dirty(inode); 4661025774cSChristoph Hellwig return 0; 4671025774cSChristoph Hellwig } 4681025774cSChristoph Hellwig 4697d54fa64SAl Viro static struct inode *hugetlbfs_get_root(struct super_block *sb, 4707d54fa64SAl Viro struct hugetlbfs_config *config) 4711da177e4SLinus Torvalds { 4721da177e4SLinus Torvalds struct inode *inode; 4731da177e4SLinus Torvalds 4741da177e4SLinus Torvalds inode = new_inode(sb); 4751da177e4SLinus Torvalds if (inode) { 4761da177e4SLinus Torvalds struct hugetlbfs_inode_info *info; 47785fe4025SChristoph Hellwig inode->i_ino = get_next_ino(); 4787d54fa64SAl Viro inode->i_mode = S_IFDIR | config->mode; 4797d54fa64SAl Viro inode->i_uid = config->uid; 4807d54fa64SAl Viro inode->i_gid = config->gid; 4817d54fa64SAl Viro inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME; 4827d54fa64SAl Viro info = HUGETLBFS_I(inode); 4837d54fa64SAl Viro mpol_shared_policy_init(&info->policy, NULL); 4847d54fa64SAl Viro inode->i_op = &hugetlbfs_dir_inode_operations; 4857d54fa64SAl Viro inode->i_fop = &simple_dir_operations; 4867d54fa64SAl Viro /* directory inodes start off with i_nlink == 2 (for "." entry) */ 4877d54fa64SAl Viro inc_nlink(inode); 4887d54fa64SAl Viro } 4897d54fa64SAl Viro return inode; 4907d54fa64SAl Viro } 4917d54fa64SAl Viro 4927d54fa64SAl Viro static struct inode *hugetlbfs_get_inode(struct super_block *sb, 4937d54fa64SAl Viro struct inode *dir, 49418df2252SAl Viro umode_t mode, dev_t dev) 4957d54fa64SAl Viro { 4967d54fa64SAl Viro struct inode *inode; 4977d54fa64SAl Viro 4987d54fa64SAl Viro inode = new_inode(sb); 4997d54fa64SAl Viro if (inode) { 5007d54fa64SAl Viro struct hugetlbfs_inode_info *info; 5017d54fa64SAl Viro inode->i_ino = get_next_ino(); 5027d54fa64SAl Viro inode_init_owner(inode, dir, mode); 5031da177e4SLinus Torvalds inode->i_mapping->a_ops = &hugetlbfs_aops; 5041da177e4SLinus Torvalds inode->i_mapping->backing_dev_info =&hugetlbfs_backing_dev_info; 5051da177e4SLinus Torvalds inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME; 506a43a8c39SChen, Kenneth W INIT_LIST_HEAD(&inode->i_mapping->private_list); 5071da177e4SLinus Torvalds info = HUGETLBFS_I(inode); 5086bfde05bSEric B Munson /* 5096bfde05bSEric B Munson * The policy is initialized here even if we are creating a 5106bfde05bSEric B Munson * private inode because initialization simply creates an 5116bfde05bSEric B Munson * an empty rb tree and calls spin_lock_init(), later when we 5126bfde05bSEric B Munson * call mpol_free_shared_policy() it will just return because 5136bfde05bSEric B Munson * the rb tree will still be empty. 5146bfde05bSEric B Munson */ 51571fe804bSLee Schermerhorn mpol_shared_policy_init(&info->policy, NULL); 5161da177e4SLinus Torvalds switch (mode & S_IFMT) { 5171da177e4SLinus Torvalds default: 5181da177e4SLinus Torvalds init_special_inode(inode, mode, dev); 5191da177e4SLinus Torvalds break; 5201da177e4SLinus Torvalds case S_IFREG: 5211da177e4SLinus Torvalds inode->i_op = &hugetlbfs_inode_operations; 5221da177e4SLinus Torvalds inode->i_fop = &hugetlbfs_file_operations; 5231da177e4SLinus Torvalds break; 5241da177e4SLinus Torvalds case S_IFDIR: 5251da177e4SLinus Torvalds inode->i_op = &hugetlbfs_dir_inode_operations; 5261da177e4SLinus Torvalds inode->i_fop = &simple_dir_operations; 5271da177e4SLinus Torvalds 5281da177e4SLinus Torvalds /* directory inodes start off with i_nlink == 2 (for "." entry) */ 529d8c76e6fSDave Hansen inc_nlink(inode); 5301da177e4SLinus Torvalds break; 5311da177e4SLinus Torvalds case S_IFLNK: 5321da177e4SLinus Torvalds inode->i_op = &page_symlink_inode_operations; 5331da177e4SLinus Torvalds break; 5341da177e4SLinus Torvalds } 535e096d0c7SJosh Boyer lockdep_annotate_inode_mutex_key(inode); 5361da177e4SLinus Torvalds } 5371da177e4SLinus Torvalds return inode; 5381da177e4SLinus Torvalds } 5391da177e4SLinus Torvalds 5401da177e4SLinus Torvalds /* 5411da177e4SLinus Torvalds * File creation. Allocate an inode, and we're done.. 5421da177e4SLinus Torvalds */ 5431da177e4SLinus Torvalds static int hugetlbfs_mknod(struct inode *dir, 5441a67aafbSAl Viro struct dentry *dentry, umode_t mode, dev_t dev) 5451da177e4SLinus Torvalds { 5461da177e4SLinus Torvalds struct inode *inode; 5471da177e4SLinus Torvalds int error = -ENOSPC; 5481da177e4SLinus Torvalds 5497d54fa64SAl Viro inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev); 5501da177e4SLinus Torvalds if (inode) { 5511da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 5521da177e4SLinus Torvalds d_instantiate(dentry, inode); 5531da177e4SLinus Torvalds dget(dentry); /* Extra count - pin the dentry in core */ 5541da177e4SLinus Torvalds error = 0; 5551da177e4SLinus Torvalds } 5561da177e4SLinus Torvalds return error; 5571da177e4SLinus Torvalds } 5581da177e4SLinus Torvalds 55918bb1db3SAl Viro static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode) 5601da177e4SLinus Torvalds { 5611da177e4SLinus Torvalds int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0); 5621da177e4SLinus Torvalds if (!retval) 563d8c76e6fSDave Hansen inc_nlink(dir); 5641da177e4SLinus Torvalds return retval; 5651da177e4SLinus Torvalds } 5661da177e4SLinus Torvalds 5674acdaf27SAl Viro static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, struct nameidata *nd) 5681da177e4SLinus Torvalds { 5691da177e4SLinus Torvalds return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0); 5701da177e4SLinus Torvalds } 5711da177e4SLinus Torvalds 5721da177e4SLinus Torvalds static int hugetlbfs_symlink(struct inode *dir, 5731da177e4SLinus Torvalds struct dentry *dentry, const char *symname) 5741da177e4SLinus Torvalds { 5751da177e4SLinus Torvalds struct inode *inode; 5761da177e4SLinus Torvalds int error = -ENOSPC; 5771da177e4SLinus Torvalds 5787d54fa64SAl Viro inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0); 5791da177e4SLinus Torvalds if (inode) { 5801da177e4SLinus Torvalds int l = strlen(symname)+1; 5811da177e4SLinus Torvalds error = page_symlink(inode, symname, l); 5821da177e4SLinus Torvalds if (!error) { 5831da177e4SLinus Torvalds d_instantiate(dentry, inode); 5841da177e4SLinus Torvalds dget(dentry); 5851da177e4SLinus Torvalds } else 5861da177e4SLinus Torvalds iput(inode); 5871da177e4SLinus Torvalds } 5881da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 5891da177e4SLinus Torvalds 5901da177e4SLinus Torvalds return error; 5911da177e4SLinus Torvalds } 5921da177e4SLinus Torvalds 5931da177e4SLinus Torvalds /* 5946649a386SKen Chen * mark the head page dirty 5951da177e4SLinus Torvalds */ 5961da177e4SLinus Torvalds static int hugetlbfs_set_page_dirty(struct page *page) 5971da177e4SLinus Torvalds { 598d85f3385SChristoph Lameter struct page *head = compound_head(page); 5996649a386SKen Chen 6006649a386SKen Chen SetPageDirty(head); 6011da177e4SLinus Torvalds return 0; 6021da177e4SLinus Torvalds } 6031da177e4SLinus Torvalds 604290408d4SNaoya Horiguchi static int hugetlbfs_migrate_page(struct address_space *mapping, 605b969c4abSMel Gorman struct page *newpage, struct page *page, 606a6bc32b8SMel Gorman enum migrate_mode mode) 607290408d4SNaoya Horiguchi { 608290408d4SNaoya Horiguchi int rc; 609290408d4SNaoya Horiguchi 610290408d4SNaoya Horiguchi rc = migrate_huge_page_move_mapping(mapping, newpage, page); 611290408d4SNaoya Horiguchi if (rc) 612290408d4SNaoya Horiguchi return rc; 613290408d4SNaoya Horiguchi migrate_page_copy(newpage, page); 614290408d4SNaoya Horiguchi 615290408d4SNaoya Horiguchi return 0; 616290408d4SNaoya Horiguchi } 617290408d4SNaoya Horiguchi 618726c3342SDavid Howells static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf) 6191da177e4SLinus Torvalds { 620726c3342SDavid Howells struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb); 621a5516438SAndi Kleen struct hstate *h = hstate_inode(dentry->d_inode); 6221da177e4SLinus Torvalds 6231da177e4SLinus Torvalds buf->f_type = HUGETLBFS_MAGIC; 624a5516438SAndi Kleen buf->f_bsize = huge_page_size(h); 6251da177e4SLinus Torvalds if (sbinfo) { 6261da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 62774a8a65cSDavid Gibson /* If no limits set, just report 0 for max/free/used 62874a8a65cSDavid Gibson * blocks, like simple_statfs() */ 62990481622SDavid Gibson if (sbinfo->spool) { 63090481622SDavid Gibson long free_pages; 63190481622SDavid Gibson 63290481622SDavid Gibson spin_lock(&sbinfo->spool->lock); 63390481622SDavid Gibson buf->f_blocks = sbinfo->spool->max_hpages; 63490481622SDavid Gibson free_pages = sbinfo->spool->max_hpages 63590481622SDavid Gibson - sbinfo->spool->used_hpages; 63690481622SDavid Gibson buf->f_bavail = buf->f_bfree = free_pages; 63790481622SDavid Gibson spin_unlock(&sbinfo->spool->lock); 6381da177e4SLinus Torvalds buf->f_files = sbinfo->max_inodes; 6391da177e4SLinus Torvalds buf->f_ffree = sbinfo->free_inodes; 64074a8a65cSDavid Gibson } 6411da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 6421da177e4SLinus Torvalds } 6431da177e4SLinus Torvalds buf->f_namelen = NAME_MAX; 6441da177e4SLinus Torvalds return 0; 6451da177e4SLinus Torvalds } 6461da177e4SLinus Torvalds 6471da177e4SLinus Torvalds static void hugetlbfs_put_super(struct super_block *sb) 6481da177e4SLinus Torvalds { 6491da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb); 6501da177e4SLinus Torvalds 6511da177e4SLinus Torvalds if (sbi) { 6521da177e4SLinus Torvalds sb->s_fs_info = NULL; 65390481622SDavid Gibson 65490481622SDavid Gibson if (sbi->spool) 65590481622SDavid Gibson hugepage_put_subpool(sbi->spool); 65690481622SDavid Gibson 6571da177e4SLinus Torvalds kfree(sbi); 6581da177e4SLinus Torvalds } 6591da177e4SLinus Torvalds } 6601da177e4SLinus Torvalds 66196527980SChristoph Hellwig static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo) 66296527980SChristoph Hellwig { 66396527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 66496527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 66596527980SChristoph Hellwig if (unlikely(!sbinfo->free_inodes)) { 66696527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 66796527980SChristoph Hellwig return 0; 66896527980SChristoph Hellwig } 66996527980SChristoph Hellwig sbinfo->free_inodes--; 67096527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 67196527980SChristoph Hellwig } 67296527980SChristoph Hellwig 67396527980SChristoph Hellwig return 1; 67496527980SChristoph Hellwig } 67596527980SChristoph Hellwig 67696527980SChristoph Hellwig static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo) 67796527980SChristoph Hellwig { 67896527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 67996527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 68096527980SChristoph Hellwig sbinfo->free_inodes++; 68196527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 68296527980SChristoph Hellwig } 68396527980SChristoph Hellwig } 68496527980SChristoph Hellwig 68596527980SChristoph Hellwig 686e18b890bSChristoph Lameter static struct kmem_cache *hugetlbfs_inode_cachep; 6871da177e4SLinus Torvalds 6881da177e4SLinus Torvalds static struct inode *hugetlbfs_alloc_inode(struct super_block *sb) 6891da177e4SLinus Torvalds { 69096527980SChristoph Hellwig struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb); 6911da177e4SLinus Torvalds struct hugetlbfs_inode_info *p; 6921da177e4SLinus Torvalds 69396527980SChristoph Hellwig if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo))) 6941da177e4SLinus Torvalds return NULL; 695e94b1766SChristoph Lameter p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL); 69696527980SChristoph Hellwig if (unlikely(!p)) { 69796527980SChristoph Hellwig hugetlbfs_inc_free_inodes(sbinfo); 69896527980SChristoph Hellwig return NULL; 6991da177e4SLinus Torvalds } 70096527980SChristoph Hellwig return &p->vfs_inode; 7011da177e4SLinus Torvalds } 7021da177e4SLinus Torvalds 703fa0d7e3dSNick Piggin static void hugetlbfs_i_callback(struct rcu_head *head) 704fa0d7e3dSNick Piggin { 705fa0d7e3dSNick Piggin struct inode *inode = container_of(head, struct inode, i_rcu); 706fa0d7e3dSNick Piggin kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode)); 707fa0d7e3dSNick Piggin } 708fa0d7e3dSNick Piggin 7091da177e4SLinus Torvalds static void hugetlbfs_destroy_inode(struct inode *inode) 7101da177e4SLinus Torvalds { 71196527980SChristoph Hellwig hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb)); 7121da177e4SLinus Torvalds mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy); 713fa0d7e3dSNick Piggin call_rcu(&inode->i_rcu, hugetlbfs_i_callback); 7141da177e4SLinus Torvalds } 7151da177e4SLinus Torvalds 716f5e54d6eSChristoph Hellwig static const struct address_space_operations hugetlbfs_aops = { 717800d15a5SNick Piggin .write_begin = hugetlbfs_write_begin, 718800d15a5SNick Piggin .write_end = hugetlbfs_write_end, 7191da177e4SLinus Torvalds .set_page_dirty = hugetlbfs_set_page_dirty, 720290408d4SNaoya Horiguchi .migratepage = hugetlbfs_migrate_page, 7211da177e4SLinus Torvalds }; 7221da177e4SLinus Torvalds 72396527980SChristoph Hellwig 72451cc5068SAlexey Dobriyan static void init_once(void *foo) 72596527980SChristoph Hellwig { 72696527980SChristoph Hellwig struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo; 72796527980SChristoph Hellwig 72896527980SChristoph Hellwig inode_init_once(&ei->vfs_inode); 72996527980SChristoph Hellwig } 73096527980SChristoph Hellwig 7314b6f5d20SArjan van de Ven const struct file_operations hugetlbfs_file_operations = { 732e63e1e5aSBadari Pulavarty .read = hugetlbfs_read, 7331da177e4SLinus Torvalds .mmap = hugetlbfs_file_mmap, 7341b061d92SChristoph Hellwig .fsync = noop_fsync, 7351da177e4SLinus Torvalds .get_unmapped_area = hugetlb_get_unmapped_area, 7366038f373SArnd Bergmann .llseek = default_llseek, 7371da177e4SLinus Torvalds }; 7381da177e4SLinus Torvalds 73992e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_dir_inode_operations = { 7401da177e4SLinus Torvalds .create = hugetlbfs_create, 7411da177e4SLinus Torvalds .lookup = simple_lookup, 7421da177e4SLinus Torvalds .link = simple_link, 7431da177e4SLinus Torvalds .unlink = simple_unlink, 7441da177e4SLinus Torvalds .symlink = hugetlbfs_symlink, 7451da177e4SLinus Torvalds .mkdir = hugetlbfs_mkdir, 7461da177e4SLinus Torvalds .rmdir = simple_rmdir, 7471da177e4SLinus Torvalds .mknod = hugetlbfs_mknod, 7481da177e4SLinus Torvalds .rename = simple_rename, 7491da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 7501da177e4SLinus Torvalds }; 7511da177e4SLinus Torvalds 75292e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_inode_operations = { 7531da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 7541da177e4SLinus Torvalds }; 7551da177e4SLinus Torvalds 756ee9b6d61SJosef 'Jeff' Sipek static const struct super_operations hugetlbfs_ops = { 7571da177e4SLinus Torvalds .alloc_inode = hugetlbfs_alloc_inode, 7581da177e4SLinus Torvalds .destroy_inode = hugetlbfs_destroy_inode, 7592bbbda30SAl Viro .evict_inode = hugetlbfs_evict_inode, 7601da177e4SLinus Torvalds .statfs = hugetlbfs_statfs, 7611da177e4SLinus Torvalds .put_super = hugetlbfs_put_super, 76210f19a86SMiklos Szeredi .show_options = generic_show_options, 7631da177e4SLinus Torvalds }; 7641da177e4SLinus Torvalds 7651da177e4SLinus Torvalds static int 7661da177e4SLinus Torvalds hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig) 7671da177e4SLinus Torvalds { 768e73a75faSRandy Dunlap char *p, *rest; 769e73a75faSRandy Dunlap substring_t args[MAX_OPT_ARGS]; 770e73a75faSRandy Dunlap int option; 771a137e1ccSAndi Kleen unsigned long long size = 0; 772a137e1ccSAndi Kleen enum { NO_SIZE, SIZE_STD, SIZE_PERCENT } setsize = NO_SIZE; 7731da177e4SLinus Torvalds 7741da177e4SLinus Torvalds if (!options) 7751da177e4SLinus Torvalds return 0; 7761da177e4SLinus Torvalds 777e73a75faSRandy Dunlap while ((p = strsep(&options, ",")) != NULL) { 778e73a75faSRandy Dunlap int token; 779b4c07bceSLee Schermerhorn if (!*p) 780b4c07bceSLee Schermerhorn continue; 7811da177e4SLinus Torvalds 782e73a75faSRandy Dunlap token = match_token(p, tokens, args); 783e73a75faSRandy Dunlap switch (token) { 784e73a75faSRandy Dunlap case Opt_uid: 785e73a75faSRandy Dunlap if (match_int(&args[0], &option)) 786e73a75faSRandy Dunlap goto bad_val; 787e73a75faSRandy Dunlap pconfig->uid = option; 788e73a75faSRandy Dunlap break; 789e73a75faSRandy Dunlap 790e73a75faSRandy Dunlap case Opt_gid: 791e73a75faSRandy Dunlap if (match_int(&args[0], &option)) 792e73a75faSRandy Dunlap goto bad_val; 793e73a75faSRandy Dunlap pconfig->gid = option; 794e73a75faSRandy Dunlap break; 795e73a75faSRandy Dunlap 796e73a75faSRandy Dunlap case Opt_mode: 797e73a75faSRandy Dunlap if (match_octal(&args[0], &option)) 798e73a75faSRandy Dunlap goto bad_val; 79975897d60SKen Chen pconfig->mode = option & 01777U; 800e73a75faSRandy Dunlap break; 801e73a75faSRandy Dunlap 802e73a75faSRandy Dunlap case Opt_size: { 803e73a75faSRandy Dunlap /* memparse() will accept a K/M/G without a digit */ 804e73a75faSRandy Dunlap if (!isdigit(*args[0].from)) 805e73a75faSRandy Dunlap goto bad_val; 806e73a75faSRandy Dunlap size = memparse(args[0].from, &rest); 807a137e1ccSAndi Kleen setsize = SIZE_STD; 808a137e1ccSAndi Kleen if (*rest == '%') 809a137e1ccSAndi Kleen setsize = SIZE_PERCENT; 810e73a75faSRandy Dunlap break; 811e73a75faSRandy Dunlap } 8121da177e4SLinus Torvalds 813e73a75faSRandy Dunlap case Opt_nr_inodes: 814e73a75faSRandy Dunlap /* memparse() will accept a K/M/G without a digit */ 815e73a75faSRandy Dunlap if (!isdigit(*args[0].from)) 816e73a75faSRandy Dunlap goto bad_val; 817e73a75faSRandy Dunlap pconfig->nr_inodes = memparse(args[0].from, &rest); 818e73a75faSRandy Dunlap break; 819e73a75faSRandy Dunlap 820a137e1ccSAndi Kleen case Opt_pagesize: { 821a137e1ccSAndi Kleen unsigned long ps; 822a137e1ccSAndi Kleen ps = memparse(args[0].from, &rest); 823a137e1ccSAndi Kleen pconfig->hstate = size_to_hstate(ps); 824a137e1ccSAndi Kleen if (!pconfig->hstate) { 825a137e1ccSAndi Kleen printk(KERN_ERR 826a137e1ccSAndi Kleen "hugetlbfs: Unsupported page size %lu MB\n", 827a137e1ccSAndi Kleen ps >> 20); 828a137e1ccSAndi Kleen return -EINVAL; 829a137e1ccSAndi Kleen } 830a137e1ccSAndi Kleen break; 831a137e1ccSAndi Kleen } 832a137e1ccSAndi Kleen 833e73a75faSRandy Dunlap default: 834b4c07bceSLee Schermerhorn printk(KERN_ERR "hugetlbfs: Bad mount option: \"%s\"\n", 835b4c07bceSLee Schermerhorn p); 836b4c07bceSLee Schermerhorn return -EINVAL; 837e73a75faSRandy Dunlap break; 838e73a75faSRandy Dunlap } 8391da177e4SLinus Torvalds } 840a137e1ccSAndi Kleen 841a137e1ccSAndi Kleen /* Do size after hstate is set up */ 842a137e1ccSAndi Kleen if (setsize > NO_SIZE) { 843a137e1ccSAndi Kleen struct hstate *h = pconfig->hstate; 844a137e1ccSAndi Kleen if (setsize == SIZE_PERCENT) { 845a137e1ccSAndi Kleen size <<= huge_page_shift(h); 846a137e1ccSAndi Kleen size *= h->max_huge_pages; 847a137e1ccSAndi Kleen do_div(size, 100); 848a137e1ccSAndi Kleen } 849a137e1ccSAndi Kleen pconfig->nr_blocks = (size >> huge_page_shift(h)); 850a137e1ccSAndi Kleen } 851a137e1ccSAndi Kleen 8521da177e4SLinus Torvalds return 0; 853e73a75faSRandy Dunlap 854e73a75faSRandy Dunlap bad_val: 855e73a75faSRandy Dunlap printk(KERN_ERR "hugetlbfs: Bad value '%s' for mount option '%s'\n", 856e73a75faSRandy Dunlap args[0].from, p); 857c12ddba0SAkinobu Mita return -EINVAL; 8581da177e4SLinus Torvalds } 8591da177e4SLinus Torvalds 8601da177e4SLinus Torvalds static int 8611da177e4SLinus Torvalds hugetlbfs_fill_super(struct super_block *sb, void *data, int silent) 8621da177e4SLinus Torvalds { 8631da177e4SLinus Torvalds int ret; 8641da177e4SLinus Torvalds struct hugetlbfs_config config; 8651da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo; 8661da177e4SLinus Torvalds 86710f19a86SMiklos Szeredi save_mount_options(sb, data); 86810f19a86SMiklos Szeredi 8691da177e4SLinus Torvalds config.nr_blocks = -1; /* No limit on size by default */ 8701da177e4SLinus Torvalds config.nr_inodes = -1; /* No limit on number of inodes by default */ 87177c70de1SDavid Howells config.uid = current_fsuid(); 87277c70de1SDavid Howells config.gid = current_fsgid(); 8731da177e4SLinus Torvalds config.mode = 0755; 874a137e1ccSAndi Kleen config.hstate = &default_hstate; 8751da177e4SLinus Torvalds ret = hugetlbfs_parse_options(data, &config); 8761da177e4SLinus Torvalds if (ret) 8771da177e4SLinus Torvalds return ret; 8781da177e4SLinus Torvalds 8791da177e4SLinus Torvalds sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL); 8801da177e4SLinus Torvalds if (!sbinfo) 8811da177e4SLinus Torvalds return -ENOMEM; 8821da177e4SLinus Torvalds sb->s_fs_info = sbinfo; 883a137e1ccSAndi Kleen sbinfo->hstate = config.hstate; 8841da177e4SLinus Torvalds spin_lock_init(&sbinfo->stat_lock); 8851da177e4SLinus Torvalds sbinfo->max_inodes = config.nr_inodes; 8861da177e4SLinus Torvalds sbinfo->free_inodes = config.nr_inodes; 88790481622SDavid Gibson sbinfo->spool = NULL; 88890481622SDavid Gibson if (config.nr_blocks != -1) { 88990481622SDavid Gibson sbinfo->spool = hugepage_new_subpool(config.nr_blocks); 89090481622SDavid Gibson if (!sbinfo->spool) 89190481622SDavid Gibson goto out_free; 89290481622SDavid Gibson } 8931da177e4SLinus Torvalds sb->s_maxbytes = MAX_LFS_FILESIZE; 894a137e1ccSAndi Kleen sb->s_blocksize = huge_page_size(config.hstate); 895a137e1ccSAndi Kleen sb->s_blocksize_bits = huge_page_shift(config.hstate); 8961da177e4SLinus Torvalds sb->s_magic = HUGETLBFS_MAGIC; 8971da177e4SLinus Torvalds sb->s_op = &hugetlbfs_ops; 8981da177e4SLinus Torvalds sb->s_time_gran = 1; 89948fde701SAl Viro sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config)); 90048fde701SAl Viro if (!sb->s_root) 9011da177e4SLinus Torvalds goto out_free; 9021da177e4SLinus Torvalds return 0; 9031da177e4SLinus Torvalds out_free: 90490481622SDavid Gibson if (sbinfo->spool) 90590481622SDavid Gibson kfree(sbinfo->spool); 9061da177e4SLinus Torvalds kfree(sbinfo); 9071da177e4SLinus Torvalds return -ENOMEM; 9081da177e4SLinus Torvalds } 9091da177e4SLinus Torvalds 9103c26ff6eSAl Viro static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type, 9113c26ff6eSAl Viro int flags, const char *dev_name, void *data) 9121da177e4SLinus Torvalds { 9133c26ff6eSAl Viro return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super); 9141da177e4SLinus Torvalds } 9151da177e4SLinus Torvalds 9161da177e4SLinus Torvalds static struct file_system_type hugetlbfs_fs_type = { 9171da177e4SLinus Torvalds .name = "hugetlbfs", 9183c26ff6eSAl Viro .mount = hugetlbfs_mount, 9191da177e4SLinus Torvalds .kill_sb = kill_litter_super, 9201da177e4SLinus Torvalds }; 9211da177e4SLinus Torvalds 9221da177e4SLinus Torvalds static struct vfsmount *hugetlbfs_vfsmount; 9231da177e4SLinus Torvalds 924ef1ff6b8SFrom: Mel Gorman static int can_do_hugetlb_shm(void) 9251da177e4SLinus Torvalds { 926ef1ff6b8SFrom: Mel Gorman return capable(CAP_IPC_LOCK) || in_group_p(sysctl_hugetlb_shm_group); 9271da177e4SLinus Torvalds } 9281da177e4SLinus Torvalds 92940716e29SSteven Truelove struct file *hugetlb_file_setup(const char *name, unsigned long addr, 93040716e29SSteven Truelove size_t size, vm_flags_t acctflag, 9316bfde05bSEric B Munson struct user_struct **user, int creat_flags) 9321da177e4SLinus Torvalds { 9331da177e4SLinus Torvalds int error = -ENOMEM; 9341da177e4SLinus Torvalds struct file *file; 9351da177e4SLinus Torvalds struct inode *inode; 9362c48b9c4SAl Viro struct path path; 9372c48b9c4SAl Viro struct dentry *root; 9381da177e4SLinus Torvalds struct qstr quick_string; 93940716e29SSteven Truelove struct hstate *hstate; 94040716e29SSteven Truelove unsigned long num_pages; 9411da177e4SLinus Torvalds 942353d5c30SHugh Dickins *user = NULL; 9435bc98594SAkinobu Mita if (!hugetlbfs_vfsmount) 9445bc98594SAkinobu Mita return ERR_PTR(-ENOENT); 9455bc98594SAkinobu Mita 946ef1ff6b8SFrom: Mel Gorman if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) { 947353d5c30SHugh Dickins *user = current_user(); 948353d5c30SHugh Dickins if (user_shm_lock(size, *user)) { 94921a3c273SDavid Rientjes task_lock(current); 95021a3c273SDavid Rientjes printk_once(KERN_WARNING 95121a3c273SDavid Rientjes "%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n", 95221a3c273SDavid Rientjes current->comm, current->pid); 95321a3c273SDavid Rientjes task_unlock(current); 954353d5c30SHugh Dickins } else { 955353d5c30SHugh Dickins *user = NULL; 9561da177e4SLinus Torvalds return ERR_PTR(-EPERM); 9572584e517SRavikiran G Thirumalai } 958353d5c30SHugh Dickins } 9591da177e4SLinus Torvalds 9601da177e4SLinus Torvalds root = hugetlbfs_vfsmount->mnt_root; 9619d66586fSEric W. Biederman quick_string.name = name; 9621da177e4SLinus Torvalds quick_string.len = strlen(quick_string.name); 9631da177e4SLinus Torvalds quick_string.hash = 0; 9642c48b9c4SAl Viro path.dentry = d_alloc(root, &quick_string); 9652c48b9c4SAl Viro if (!path.dentry) 9661da177e4SLinus Torvalds goto out_shm_unlock; 9671da177e4SLinus Torvalds 9682c48b9c4SAl Viro path.mnt = mntget(hugetlbfs_vfsmount); 9691da177e4SLinus Torvalds error = -ENOSPC; 9707d54fa64SAl Viro inode = hugetlbfs_get_inode(root->d_sb, NULL, S_IFREG | S_IRWXUGO, 0); 9711da177e4SLinus Torvalds if (!inode) 972ce8d2cdfSDave Hansen goto out_dentry; 9731da177e4SLinus Torvalds 97440716e29SSteven Truelove hstate = hstate_inode(inode); 97540716e29SSteven Truelove size += addr & ~huge_page_mask(hstate); 97640716e29SSteven Truelove num_pages = ALIGN(size, huge_page_size(hstate)) >> 97740716e29SSteven Truelove huge_page_shift(hstate); 978b45b5bd6SDavid Gibson error = -ENOMEM; 97940716e29SSteven Truelove if (hugetlb_reserve_pages(inode, 0, num_pages, NULL, acctflag)) 980b45b5bd6SDavid Gibson goto out_inode; 981b45b5bd6SDavid Gibson 9822c48b9c4SAl Viro d_instantiate(path.dentry, inode); 9831da177e4SLinus Torvalds inode->i_size = size; 9846d6b77f1SMiklos Szeredi clear_nlink(inode); 985ce8d2cdfSDave Hansen 986ce8d2cdfSDave Hansen error = -ENFILE; 9872c48b9c4SAl Viro file = alloc_file(&path, FMODE_WRITE | FMODE_READ, 988ce8d2cdfSDave Hansen &hugetlbfs_file_operations); 989ce8d2cdfSDave Hansen if (!file) 990b4d232e6SAl Viro goto out_dentry; /* inode is already attached */ 991ce8d2cdfSDave Hansen 9921da177e4SLinus Torvalds return file; 9931da177e4SLinus Torvalds 994b45b5bd6SDavid Gibson out_inode: 995b45b5bd6SDavid Gibson iput(inode); 9961da177e4SLinus Torvalds out_dentry: 9972c48b9c4SAl Viro path_put(&path); 9981da177e4SLinus Torvalds out_shm_unlock: 999353d5c30SHugh Dickins if (*user) { 1000353d5c30SHugh Dickins user_shm_unlock(size, *user); 1001353d5c30SHugh Dickins *user = NULL; 1002353d5c30SHugh Dickins } 10031da177e4SLinus Torvalds return ERR_PTR(error); 10041da177e4SLinus Torvalds } 10051da177e4SLinus Torvalds 10061da177e4SLinus Torvalds static int __init init_hugetlbfs_fs(void) 10071da177e4SLinus Torvalds { 10081da177e4SLinus Torvalds int error; 10091da177e4SLinus Torvalds struct vfsmount *vfsmount; 10101da177e4SLinus Torvalds 1011e0bf68ddSPeter Zijlstra error = bdi_init(&hugetlbfs_backing_dev_info); 1012e0bf68ddSPeter Zijlstra if (error) 1013e0bf68ddSPeter Zijlstra return error; 1014e0bf68ddSPeter Zijlstra 1015d1d5e05fSHillf Danton error = -ENOMEM; 10161da177e4SLinus Torvalds hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache", 10171da177e4SLinus Torvalds sizeof(struct hugetlbfs_inode_info), 101820c2df83SPaul Mundt 0, 0, init_once); 10191da177e4SLinus Torvalds if (hugetlbfs_inode_cachep == NULL) 1020e0bf68ddSPeter Zijlstra goto out2; 10211da177e4SLinus Torvalds 10221da177e4SLinus Torvalds error = register_filesystem(&hugetlbfs_fs_type); 10231da177e4SLinus Torvalds if (error) 10241da177e4SLinus Torvalds goto out; 10251da177e4SLinus Torvalds 10261da177e4SLinus Torvalds vfsmount = kern_mount(&hugetlbfs_fs_type); 10271da177e4SLinus Torvalds 10281da177e4SLinus Torvalds if (!IS_ERR(vfsmount)) { 10291da177e4SLinus Torvalds hugetlbfs_vfsmount = vfsmount; 10301da177e4SLinus Torvalds return 0; 10311da177e4SLinus Torvalds } 10321da177e4SLinus Torvalds 10331da177e4SLinus Torvalds error = PTR_ERR(vfsmount); 10341da177e4SLinus Torvalds 10351da177e4SLinus Torvalds out: 10361da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 1037e0bf68ddSPeter Zijlstra out2: 1038e0bf68ddSPeter Zijlstra bdi_destroy(&hugetlbfs_backing_dev_info); 10391da177e4SLinus Torvalds return error; 10401da177e4SLinus Torvalds } 10411da177e4SLinus Torvalds 10421da177e4SLinus Torvalds static void __exit exit_hugetlbfs_fs(void) 10431da177e4SLinus Torvalds { 10441da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 1045423e0ab0STim Chen kern_unmount(hugetlbfs_vfsmount); 10461da177e4SLinus Torvalds unregister_filesystem(&hugetlbfs_fs_type); 1047e0bf68ddSPeter Zijlstra bdi_destroy(&hugetlbfs_backing_dev_info); 10481da177e4SLinus Torvalds } 10491da177e4SLinus Torvalds 10501da177e4SLinus Torvalds module_init(init_hugetlbfs_fs) 10511da177e4SLinus Torvalds module_exit(exit_hugetlbfs_fs) 10521da177e4SLinus Torvalds 10531da177e4SLinus Torvalds MODULE_LICENSE("GPL"); 1054