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> 341da177e4SLinus Torvalds 351da177e4SLinus Torvalds #include <asm/uaccess.h> 361da177e4SLinus Torvalds 37ee9b6d61SJosef 'Jeff' Sipek static const struct super_operations hugetlbfs_ops; 38f5e54d6eSChristoph Hellwig static const struct address_space_operations hugetlbfs_aops; 394b6f5d20SArjan van de Ven const struct file_operations hugetlbfs_file_operations; 4092e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_dir_inode_operations; 4192e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_inode_operations; 421da177e4SLinus Torvalds 431da177e4SLinus Torvalds static struct backing_dev_info hugetlbfs_backing_dev_info = { 44d993831fSJens Axboe .name = "hugetlbfs", 451da177e4SLinus Torvalds .ra_pages = 0, /* No readahead */ 46e4ad08feSMiklos Szeredi .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK, 471da177e4SLinus Torvalds }; 481da177e4SLinus Torvalds 491da177e4SLinus Torvalds int sysctl_hugetlb_shm_group; 501da177e4SLinus Torvalds 51e73a75faSRandy Dunlap enum { 52e73a75faSRandy Dunlap Opt_size, Opt_nr_inodes, 53e73a75faSRandy Dunlap Opt_mode, Opt_uid, Opt_gid, 54a137e1ccSAndi Kleen Opt_pagesize, 55e73a75faSRandy Dunlap Opt_err, 56e73a75faSRandy Dunlap }; 57e73a75faSRandy Dunlap 58a447c093SSteven Whitehouse static const match_table_t tokens = { 59e73a75faSRandy Dunlap {Opt_size, "size=%s"}, 60e73a75faSRandy Dunlap {Opt_nr_inodes, "nr_inodes=%s"}, 61e73a75faSRandy Dunlap {Opt_mode, "mode=%o"}, 62e73a75faSRandy Dunlap {Opt_uid, "uid=%u"}, 63e73a75faSRandy Dunlap {Opt_gid, "gid=%u"}, 64a137e1ccSAndi Kleen {Opt_pagesize, "pagesize=%s"}, 65e73a75faSRandy Dunlap {Opt_err, NULL}, 66e73a75faSRandy Dunlap }; 67e73a75faSRandy Dunlap 682e9b367cSAdam Litke static void huge_pagevec_release(struct pagevec *pvec) 692e9b367cSAdam Litke { 702e9b367cSAdam Litke int i; 712e9b367cSAdam Litke 722e9b367cSAdam Litke for (i = 0; i < pagevec_count(pvec); ++i) 732e9b367cSAdam Litke put_page(pvec->pages[i]); 742e9b367cSAdam Litke 752e9b367cSAdam Litke pagevec_reinit(pvec); 762e9b367cSAdam Litke } 772e9b367cSAdam Litke 781da177e4SLinus Torvalds static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma) 791da177e4SLinus Torvalds { 80b39424e2SJosef Sipek struct inode *inode = file->f_path.dentry->d_inode; 811da177e4SLinus Torvalds loff_t len, vma_len; 821da177e4SLinus Torvalds int ret; 83a5516438SAndi Kleen struct hstate *h = hstate_file(file); 841da177e4SLinus Torvalds 8568589bc3SHugh Dickins /* 86dec4ad86SDavid Gibson * vma address alignment (but not the pgoff alignment) has 87dec4ad86SDavid Gibson * already been checked by prepare_hugepage_range. If you add 88dec4ad86SDavid Gibson * any error returns here, do so after setting VM_HUGETLB, so 89dec4ad86SDavid Gibson * is_vm_hugetlb_page tests below unmap_region go the right 90dec4ad86SDavid Gibson * way when do_mmap_pgoff unwinds (may be important on powerpc 91dec4ad86SDavid Gibson * and ia64). 9268589bc3SHugh Dickins */ 9368589bc3SHugh Dickins vma->vm_flags |= VM_HUGETLB | VM_RESERVED; 9468589bc3SHugh Dickins vma->vm_ops = &hugetlb_vm_ops; 951da177e4SLinus Torvalds 96a5516438SAndi Kleen if (vma->vm_pgoff & ~(huge_page_mask(h) >> PAGE_SHIFT)) 97dec4ad86SDavid Gibson return -EINVAL; 98dec4ad86SDavid Gibson 991da177e4SLinus Torvalds vma_len = (loff_t)(vma->vm_end - vma->vm_start); 1001da177e4SLinus Torvalds 1011b1dcc1bSJes Sorensen mutex_lock(&inode->i_mutex); 1021da177e4SLinus Torvalds file_accessed(file); 1031da177e4SLinus Torvalds 1041da177e4SLinus Torvalds ret = -ENOMEM; 1051da177e4SLinus Torvalds len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); 1061da177e4SLinus Torvalds 107a1e78772SMel Gorman if (hugetlb_reserve_pages(inode, 108a5516438SAndi Kleen vma->vm_pgoff >> huge_page_order(h), 1095a6fe125SMel Gorman len >> huge_page_shift(h), vma, 1105a6fe125SMel Gorman vma->vm_flags)) 111b45b5bd6SDavid Gibson goto out; 112b45b5bd6SDavid Gibson 1134c887265SAdam Litke ret = 0; 1144c887265SAdam Litke hugetlb_prefault_arch_hook(vma->vm_mm); 115b6174df5SZhang, Yanmin if (vma->vm_flags & VM_WRITE && inode->i_size < len) 1161da177e4SLinus Torvalds inode->i_size = len; 1171da177e4SLinus Torvalds out: 1181b1dcc1bSJes Sorensen mutex_unlock(&inode->i_mutex); 1191da177e4SLinus Torvalds 1201da177e4SLinus Torvalds return ret; 1211da177e4SLinus Torvalds } 1221da177e4SLinus Torvalds 1231da177e4SLinus Torvalds /* 124508034a3SHugh Dickins * Called under down_write(mmap_sem). 1251da177e4SLinus Torvalds */ 1261da177e4SLinus Torvalds 127d2ba27e8SAdrian Bunk #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA 1281da177e4SLinus Torvalds static unsigned long 1291da177e4SLinus Torvalds hugetlb_get_unmapped_area(struct file *file, unsigned long addr, 1301da177e4SLinus Torvalds unsigned long len, unsigned long pgoff, unsigned long flags) 1311da177e4SLinus Torvalds { 1321da177e4SLinus Torvalds struct mm_struct *mm = current->mm; 1331da177e4SLinus Torvalds struct vm_area_struct *vma; 1341da177e4SLinus Torvalds unsigned long start_addr; 135a5516438SAndi Kleen struct hstate *h = hstate_file(file); 1361da177e4SLinus Torvalds 137a5516438SAndi Kleen if (len & ~huge_page_mask(h)) 1381da177e4SLinus Torvalds return -EINVAL; 1391da177e4SLinus Torvalds if (len > TASK_SIZE) 1401da177e4SLinus Torvalds return -ENOMEM; 1411da177e4SLinus Torvalds 142036e0856SBenjamin Herrenschmidt if (flags & MAP_FIXED) { 143a5516438SAndi Kleen if (prepare_hugepage_range(file, addr, len)) 144036e0856SBenjamin Herrenschmidt return -EINVAL; 145036e0856SBenjamin Herrenschmidt return addr; 146036e0856SBenjamin Herrenschmidt } 147036e0856SBenjamin Herrenschmidt 1481da177e4SLinus Torvalds if (addr) { 149a5516438SAndi Kleen addr = ALIGN(addr, huge_page_size(h)); 1501da177e4SLinus Torvalds vma = find_vma(mm, addr); 1511da177e4SLinus Torvalds if (TASK_SIZE - len >= addr && 1521da177e4SLinus Torvalds (!vma || addr + len <= vma->vm_start)) 1531da177e4SLinus Torvalds return addr; 1541da177e4SLinus Torvalds } 1551da177e4SLinus Torvalds 1561da177e4SLinus Torvalds start_addr = mm->free_area_cache; 1571da177e4SLinus Torvalds 1581363c3cdSWolfgang Wander if (len <= mm->cached_hole_size) 1591363c3cdSWolfgang Wander start_addr = TASK_UNMAPPED_BASE; 1601363c3cdSWolfgang Wander 1611da177e4SLinus Torvalds full_search: 162a5516438SAndi Kleen addr = ALIGN(start_addr, huge_page_size(h)); 1631da177e4SLinus Torvalds 1641da177e4SLinus Torvalds for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1651da177e4SLinus Torvalds /* At this point: (!vma || addr < vma->vm_end). */ 1661da177e4SLinus Torvalds if (TASK_SIZE - len < addr) { 1671da177e4SLinus Torvalds /* 1681da177e4SLinus Torvalds * Start a new search - just in case we missed 1691da177e4SLinus Torvalds * some holes. 1701da177e4SLinus Torvalds */ 1711da177e4SLinus Torvalds if (start_addr != TASK_UNMAPPED_BASE) { 1721da177e4SLinus Torvalds start_addr = TASK_UNMAPPED_BASE; 1731da177e4SLinus Torvalds goto full_search; 1741da177e4SLinus Torvalds } 1751da177e4SLinus Torvalds return -ENOMEM; 1761da177e4SLinus Torvalds } 1771da177e4SLinus Torvalds 1781da177e4SLinus Torvalds if (!vma || addr + len <= vma->vm_start) 1791da177e4SLinus Torvalds return addr; 180a5516438SAndi Kleen addr = ALIGN(vma->vm_end, huge_page_size(h)); 1811da177e4SLinus Torvalds } 1821da177e4SLinus Torvalds } 1831da177e4SLinus Torvalds #endif 1841da177e4SLinus Torvalds 185e63e1e5aSBadari Pulavarty static int 186e63e1e5aSBadari Pulavarty hugetlbfs_read_actor(struct page *page, unsigned long offset, 187e63e1e5aSBadari Pulavarty char __user *buf, unsigned long count, 188e63e1e5aSBadari Pulavarty unsigned long size) 189e63e1e5aSBadari Pulavarty { 190e63e1e5aSBadari Pulavarty char *kaddr; 191e63e1e5aSBadari Pulavarty unsigned long left, copied = 0; 192e63e1e5aSBadari Pulavarty int i, chunksize; 193e63e1e5aSBadari Pulavarty 194e63e1e5aSBadari Pulavarty if (size > count) 195e63e1e5aSBadari Pulavarty size = count; 196e63e1e5aSBadari Pulavarty 197e63e1e5aSBadari Pulavarty /* Find which 4k chunk and offset with in that chunk */ 198e63e1e5aSBadari Pulavarty i = offset >> PAGE_CACHE_SHIFT; 199e63e1e5aSBadari Pulavarty offset = offset & ~PAGE_CACHE_MASK; 200e63e1e5aSBadari Pulavarty 201e63e1e5aSBadari Pulavarty while (size) { 202e63e1e5aSBadari Pulavarty chunksize = PAGE_CACHE_SIZE; 203e63e1e5aSBadari Pulavarty if (offset) 204e63e1e5aSBadari Pulavarty chunksize -= offset; 205e63e1e5aSBadari Pulavarty if (chunksize > size) 206e63e1e5aSBadari Pulavarty chunksize = size; 207e63e1e5aSBadari Pulavarty kaddr = kmap(&page[i]); 208e63e1e5aSBadari Pulavarty left = __copy_to_user(buf, kaddr + offset, chunksize); 209e63e1e5aSBadari Pulavarty kunmap(&page[i]); 210e63e1e5aSBadari Pulavarty if (left) { 211e63e1e5aSBadari Pulavarty copied += (chunksize - left); 212e63e1e5aSBadari Pulavarty break; 213e63e1e5aSBadari Pulavarty } 214e63e1e5aSBadari Pulavarty offset = 0; 215e63e1e5aSBadari Pulavarty size -= chunksize; 216e63e1e5aSBadari Pulavarty buf += chunksize; 217e63e1e5aSBadari Pulavarty copied += chunksize; 218e63e1e5aSBadari Pulavarty i++; 219e63e1e5aSBadari Pulavarty } 220e63e1e5aSBadari Pulavarty return copied ? copied : -EFAULT; 221e63e1e5aSBadari Pulavarty } 222e63e1e5aSBadari Pulavarty 223e63e1e5aSBadari Pulavarty /* 224e63e1e5aSBadari Pulavarty * Support for read() - Find the page attached to f_mapping and copy out the 225e63e1e5aSBadari Pulavarty * data. Its *very* similar to do_generic_mapping_read(), we can't use that 226e63e1e5aSBadari Pulavarty * since it has PAGE_CACHE_SIZE assumptions. 227e63e1e5aSBadari Pulavarty */ 228e63e1e5aSBadari Pulavarty static ssize_t hugetlbfs_read(struct file *filp, char __user *buf, 229e63e1e5aSBadari Pulavarty size_t len, loff_t *ppos) 230e63e1e5aSBadari Pulavarty { 231a5516438SAndi Kleen struct hstate *h = hstate_file(filp); 232e63e1e5aSBadari Pulavarty struct address_space *mapping = filp->f_mapping; 233e63e1e5aSBadari Pulavarty struct inode *inode = mapping->host; 234a5516438SAndi Kleen unsigned long index = *ppos >> huge_page_shift(h); 235a5516438SAndi Kleen unsigned long offset = *ppos & ~huge_page_mask(h); 236e63e1e5aSBadari Pulavarty unsigned long end_index; 237e63e1e5aSBadari Pulavarty loff_t isize; 238e63e1e5aSBadari Pulavarty ssize_t retval = 0; 239e63e1e5aSBadari Pulavarty 240e63e1e5aSBadari Pulavarty mutex_lock(&inode->i_mutex); 241e63e1e5aSBadari Pulavarty 242e63e1e5aSBadari Pulavarty /* validate length */ 243e63e1e5aSBadari Pulavarty if (len == 0) 244e63e1e5aSBadari Pulavarty goto out; 245e63e1e5aSBadari Pulavarty 246e63e1e5aSBadari Pulavarty isize = i_size_read(inode); 247e63e1e5aSBadari Pulavarty if (!isize) 248e63e1e5aSBadari Pulavarty goto out; 249e63e1e5aSBadari Pulavarty 250a5516438SAndi Kleen end_index = (isize - 1) >> huge_page_shift(h); 251e63e1e5aSBadari Pulavarty for (;;) { 252e63e1e5aSBadari Pulavarty struct page *page; 253a5516438SAndi Kleen unsigned long nr, ret; 25491bf189cSRoel Kluin int ra; 255e63e1e5aSBadari Pulavarty 256e63e1e5aSBadari Pulavarty /* nr is the maximum number of bytes to copy from this page */ 257a5516438SAndi Kleen nr = huge_page_size(h); 258e63e1e5aSBadari Pulavarty if (index >= end_index) { 259e63e1e5aSBadari Pulavarty if (index > end_index) 260e63e1e5aSBadari Pulavarty goto out; 261a5516438SAndi Kleen nr = ((isize - 1) & ~huge_page_mask(h)) + 1; 262e63e1e5aSBadari Pulavarty if (nr <= offset) { 263e63e1e5aSBadari Pulavarty goto out; 264e63e1e5aSBadari Pulavarty } 265e63e1e5aSBadari Pulavarty } 266e63e1e5aSBadari Pulavarty nr = nr - offset; 267e63e1e5aSBadari Pulavarty 268e63e1e5aSBadari Pulavarty /* Find the page */ 269e63e1e5aSBadari Pulavarty page = find_get_page(mapping, index); 270e63e1e5aSBadari Pulavarty if (unlikely(page == NULL)) { 271e63e1e5aSBadari Pulavarty /* 272e63e1e5aSBadari Pulavarty * We have a HOLE, zero out the user-buffer for the 273e63e1e5aSBadari Pulavarty * length of the hole or request. 274e63e1e5aSBadari Pulavarty */ 275e63e1e5aSBadari Pulavarty ret = len < nr ? len : nr; 276e63e1e5aSBadari Pulavarty if (clear_user(buf, ret)) 27791bf189cSRoel Kluin ra = -EFAULT; 27891bf189cSRoel Kluin else 27991bf189cSRoel Kluin ra = 0; 280e63e1e5aSBadari Pulavarty } else { 281e63e1e5aSBadari Pulavarty /* 282e63e1e5aSBadari Pulavarty * We have the page, copy it to user space buffer. 283e63e1e5aSBadari Pulavarty */ 28491bf189cSRoel Kluin ra = hugetlbfs_read_actor(page, offset, buf, len, nr); 28591bf189cSRoel Kluin ret = ra; 286e63e1e5aSBadari Pulavarty } 28791bf189cSRoel Kluin if (ra < 0) { 288e63e1e5aSBadari Pulavarty if (retval == 0) 28991bf189cSRoel Kluin retval = ra; 290e63e1e5aSBadari Pulavarty if (page) 291e63e1e5aSBadari Pulavarty page_cache_release(page); 292e63e1e5aSBadari Pulavarty goto out; 293e63e1e5aSBadari Pulavarty } 294e63e1e5aSBadari Pulavarty 295e63e1e5aSBadari Pulavarty offset += ret; 296e63e1e5aSBadari Pulavarty retval += ret; 297e63e1e5aSBadari Pulavarty len -= ret; 298a5516438SAndi Kleen index += offset >> huge_page_shift(h); 299a5516438SAndi Kleen offset &= ~huge_page_mask(h); 300e63e1e5aSBadari Pulavarty 301e63e1e5aSBadari Pulavarty if (page) 302e63e1e5aSBadari Pulavarty page_cache_release(page); 303e63e1e5aSBadari Pulavarty 304e63e1e5aSBadari Pulavarty /* short read or no more work */ 305e63e1e5aSBadari Pulavarty if ((ret != nr) || (len == 0)) 306e63e1e5aSBadari Pulavarty break; 307e63e1e5aSBadari Pulavarty } 308e63e1e5aSBadari Pulavarty out: 309a5516438SAndi Kleen *ppos = ((loff_t)index << huge_page_shift(h)) + offset; 310e63e1e5aSBadari Pulavarty mutex_unlock(&inode->i_mutex); 311e63e1e5aSBadari Pulavarty return retval; 312e63e1e5aSBadari Pulavarty } 313e63e1e5aSBadari Pulavarty 314800d15a5SNick Piggin static int hugetlbfs_write_begin(struct file *file, 315800d15a5SNick Piggin struct address_space *mapping, 316800d15a5SNick Piggin loff_t pos, unsigned len, unsigned flags, 317800d15a5SNick Piggin struct page **pagep, void **fsdata) 3181da177e4SLinus Torvalds { 3191da177e4SLinus Torvalds return -EINVAL; 3201da177e4SLinus Torvalds } 3211da177e4SLinus Torvalds 322800d15a5SNick Piggin static int hugetlbfs_write_end(struct file *file, struct address_space *mapping, 323800d15a5SNick Piggin loff_t pos, unsigned len, unsigned copied, 324800d15a5SNick Piggin struct page *page, void *fsdata) 3251da177e4SLinus Torvalds { 326800d15a5SNick Piggin BUG(); 3271da177e4SLinus Torvalds return -EINVAL; 3281da177e4SLinus Torvalds } 3291da177e4SLinus Torvalds 3301da177e4SLinus Torvalds static void truncate_huge_page(struct page *page) 3311da177e4SLinus Torvalds { 332fba2591bSLinus Torvalds cancel_dirty_page(page, /* No IO accounting for huge pages? */0); 3331da177e4SLinus Torvalds ClearPageUptodate(page); 3341da177e4SLinus Torvalds remove_from_page_cache(page); 3351da177e4SLinus Torvalds put_page(page); 3361da177e4SLinus Torvalds } 3371da177e4SLinus Torvalds 338b45b5bd6SDavid Gibson static void truncate_hugepages(struct inode *inode, loff_t lstart) 3391da177e4SLinus Torvalds { 340a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 341b45b5bd6SDavid Gibson struct address_space *mapping = &inode->i_data; 342a5516438SAndi Kleen const pgoff_t start = lstart >> huge_page_shift(h); 3431da177e4SLinus Torvalds struct pagevec pvec; 3441da177e4SLinus Torvalds pgoff_t next; 345a43a8c39SChen, Kenneth W int i, freed = 0; 3461da177e4SLinus Torvalds 3471da177e4SLinus Torvalds pagevec_init(&pvec, 0); 3481da177e4SLinus Torvalds next = start; 3491da177e4SLinus Torvalds while (1) { 3501da177e4SLinus Torvalds if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { 3511da177e4SLinus Torvalds if (next == start) 3521da177e4SLinus Torvalds break; 3531da177e4SLinus Torvalds next = start; 3541da177e4SLinus Torvalds continue; 3551da177e4SLinus Torvalds } 3561da177e4SLinus Torvalds 3571da177e4SLinus Torvalds for (i = 0; i < pagevec_count(&pvec); ++i) { 3581da177e4SLinus Torvalds struct page *page = pvec.pages[i]; 3591da177e4SLinus Torvalds 3601da177e4SLinus Torvalds lock_page(page); 3611da177e4SLinus Torvalds if (page->index > next) 3621da177e4SLinus Torvalds next = page->index; 3631da177e4SLinus Torvalds ++next; 3641da177e4SLinus Torvalds truncate_huge_page(page); 3651da177e4SLinus Torvalds unlock_page(page); 366a43a8c39SChen, Kenneth W freed++; 3671da177e4SLinus Torvalds } 3681da177e4SLinus Torvalds huge_pagevec_release(&pvec); 3691da177e4SLinus Torvalds } 3701da177e4SLinus Torvalds BUG_ON(!lstart && mapping->nrpages); 371a43a8c39SChen, Kenneth W hugetlb_unreserve_pages(inode, start, freed); 3721da177e4SLinus Torvalds } 3731da177e4SLinus Torvalds 3742bbbda30SAl Viro static void hugetlbfs_evict_inode(struct inode *inode) 3751da177e4SLinus Torvalds { 376b45b5bd6SDavid Gibson truncate_hugepages(inode, 0); 377149f4211SChristoph Hellwig clear_inode(inode); 378149f4211SChristoph Hellwig } 379149f4211SChristoph Hellwig 3801da177e4SLinus Torvalds static inline void 381856fc295SHugh Dickins hugetlb_vmtruncate_list(struct prio_tree_root *root, pgoff_t pgoff) 3821da177e4SLinus Torvalds { 3831da177e4SLinus Torvalds struct vm_area_struct *vma; 3841da177e4SLinus Torvalds struct prio_tree_iter iter; 3851da177e4SLinus Torvalds 386856fc295SHugh Dickins vma_prio_tree_foreach(vma, &iter, root, pgoff, ULONG_MAX) { 3871da177e4SLinus Torvalds unsigned long v_offset; 3881da177e4SLinus Torvalds 3891da177e4SLinus Torvalds /* 390856fc295SHugh Dickins * Can the expression below overflow on 32-bit arches? 391856fc295SHugh Dickins * No, because the prio_tree returns us only those vmas 392856fc295SHugh Dickins * which overlap the truncated area starting at pgoff, 393856fc295SHugh Dickins * and no vma on a 32-bit arch can span beyond the 4GB. 3941da177e4SLinus Torvalds */ 395856fc295SHugh Dickins if (vma->vm_pgoff < pgoff) 396856fc295SHugh Dickins v_offset = (pgoff - vma->vm_pgoff) << PAGE_SHIFT; 397856fc295SHugh Dickins else 3981da177e4SLinus Torvalds v_offset = 0; 3991da177e4SLinus Torvalds 400502717f4SChen, Kenneth W __unmap_hugepage_range(vma, 40104f2cbe3SMel Gorman vma->vm_start + v_offset, vma->vm_end, NULL); 4021da177e4SLinus Torvalds } 4031da177e4SLinus Torvalds } 4041da177e4SLinus Torvalds 4051da177e4SLinus Torvalds static int hugetlb_vmtruncate(struct inode *inode, loff_t offset) 4061da177e4SLinus Torvalds { 407856fc295SHugh Dickins pgoff_t pgoff; 4081da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 409a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 4101da177e4SLinus Torvalds 411a5516438SAndi Kleen BUG_ON(offset & ~huge_page_mask(h)); 412856fc295SHugh Dickins pgoff = offset >> PAGE_SHIFT; 4131da177e4SLinus Torvalds 4147aa91e10SKen Chen i_size_write(inode, offset); 4151da177e4SLinus Torvalds spin_lock(&mapping->i_mmap_lock); 4161da177e4SLinus Torvalds if (!prio_tree_empty(&mapping->i_mmap)) 4171da177e4SLinus Torvalds hugetlb_vmtruncate_list(&mapping->i_mmap, pgoff); 4181da177e4SLinus Torvalds spin_unlock(&mapping->i_mmap_lock); 419b45b5bd6SDavid Gibson truncate_hugepages(inode, offset); 4201da177e4SLinus Torvalds return 0; 4211da177e4SLinus Torvalds } 4221da177e4SLinus Torvalds 4231da177e4SLinus Torvalds static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr) 4241da177e4SLinus Torvalds { 4251da177e4SLinus Torvalds struct inode *inode = dentry->d_inode; 426a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 4271da177e4SLinus Torvalds int error; 4281da177e4SLinus Torvalds unsigned int ia_valid = attr->ia_valid; 4291da177e4SLinus Torvalds 4301da177e4SLinus Torvalds BUG_ON(!inode); 4311da177e4SLinus Torvalds 4321da177e4SLinus Torvalds error = inode_change_ok(inode, attr); 4331da177e4SLinus Torvalds if (error) 4341025774cSChristoph Hellwig return error; 4351da177e4SLinus Torvalds 4361da177e4SLinus Torvalds if (ia_valid & ATTR_SIZE) { 4371da177e4SLinus Torvalds error = -EINVAL; 4381025774cSChristoph Hellwig if (attr->ia_size & ~huge_page_mask(h)) 4391025774cSChristoph Hellwig return -EINVAL; 4401da177e4SLinus Torvalds error = hugetlb_vmtruncate(inode, attr->ia_size); 4411da177e4SLinus Torvalds if (error) 4421da177e4SLinus Torvalds return error; 4431da177e4SLinus Torvalds } 4441da177e4SLinus Torvalds 4451025774cSChristoph Hellwig setattr_copy(inode, attr); 4461025774cSChristoph Hellwig mark_inode_dirty(inode); 4471025774cSChristoph Hellwig return 0; 4481025774cSChristoph Hellwig } 4491025774cSChristoph Hellwig 4501da177e4SLinus Torvalds static struct inode *hugetlbfs_get_inode(struct super_block *sb, uid_t uid, 4511da177e4SLinus Torvalds gid_t gid, int mode, dev_t dev) 4521da177e4SLinus Torvalds { 4531da177e4SLinus Torvalds struct inode *inode; 4541da177e4SLinus Torvalds 4551da177e4SLinus Torvalds inode = new_inode(sb); 4561da177e4SLinus Torvalds if (inode) { 4571da177e4SLinus Torvalds struct hugetlbfs_inode_info *info; 4581da177e4SLinus Torvalds inode->i_mode = mode; 4591da177e4SLinus Torvalds inode->i_uid = uid; 4601da177e4SLinus Torvalds inode->i_gid = gid; 4611da177e4SLinus Torvalds inode->i_mapping->a_ops = &hugetlbfs_aops; 4621da177e4SLinus Torvalds inode->i_mapping->backing_dev_info =&hugetlbfs_backing_dev_info; 4631da177e4SLinus Torvalds inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME; 464a43a8c39SChen, Kenneth W INIT_LIST_HEAD(&inode->i_mapping->private_list); 4651da177e4SLinus Torvalds info = HUGETLBFS_I(inode); 4666bfde05bSEric B Munson /* 4676bfde05bSEric B Munson * The policy is initialized here even if we are creating a 4686bfde05bSEric B Munson * private inode because initialization simply creates an 4696bfde05bSEric B Munson * an empty rb tree and calls spin_lock_init(), later when we 4706bfde05bSEric B Munson * call mpol_free_shared_policy() it will just return because 4716bfde05bSEric B Munson * the rb tree will still be empty. 4726bfde05bSEric B Munson */ 47371fe804bSLee Schermerhorn mpol_shared_policy_init(&info->policy, NULL); 4741da177e4SLinus Torvalds switch (mode & S_IFMT) { 4751da177e4SLinus Torvalds default: 4761da177e4SLinus Torvalds init_special_inode(inode, mode, dev); 4771da177e4SLinus Torvalds break; 4781da177e4SLinus Torvalds case S_IFREG: 4791da177e4SLinus Torvalds inode->i_op = &hugetlbfs_inode_operations; 4801da177e4SLinus Torvalds inode->i_fop = &hugetlbfs_file_operations; 4811da177e4SLinus Torvalds break; 4821da177e4SLinus Torvalds case S_IFDIR: 4831da177e4SLinus Torvalds inode->i_op = &hugetlbfs_dir_inode_operations; 4841da177e4SLinus Torvalds inode->i_fop = &simple_dir_operations; 4851da177e4SLinus Torvalds 4861da177e4SLinus Torvalds /* directory inodes start off with i_nlink == 2 (for "." entry) */ 487d8c76e6fSDave Hansen inc_nlink(inode); 4881da177e4SLinus Torvalds break; 4891da177e4SLinus Torvalds case S_IFLNK: 4901da177e4SLinus Torvalds inode->i_op = &page_symlink_inode_operations; 4911da177e4SLinus Torvalds break; 4921da177e4SLinus Torvalds } 4931da177e4SLinus Torvalds } 4941da177e4SLinus Torvalds return inode; 4951da177e4SLinus Torvalds } 4961da177e4SLinus Torvalds 4971da177e4SLinus Torvalds /* 4981da177e4SLinus Torvalds * File creation. Allocate an inode, and we're done.. 4991da177e4SLinus Torvalds */ 5001da177e4SLinus Torvalds static int hugetlbfs_mknod(struct inode *dir, 5011da177e4SLinus Torvalds struct dentry *dentry, int mode, dev_t dev) 5021da177e4SLinus Torvalds { 5031da177e4SLinus Torvalds struct inode *inode; 5041da177e4SLinus Torvalds int error = -ENOSPC; 5051da177e4SLinus Torvalds gid_t gid; 5061da177e4SLinus Torvalds 5071da177e4SLinus Torvalds if (dir->i_mode & S_ISGID) { 5081da177e4SLinus Torvalds gid = dir->i_gid; 5091da177e4SLinus Torvalds if (S_ISDIR(mode)) 5101da177e4SLinus Torvalds mode |= S_ISGID; 5111da177e4SLinus Torvalds } else { 51277c70de1SDavid Howells gid = current_fsgid(); 5131da177e4SLinus Torvalds } 51477c70de1SDavid Howells inode = hugetlbfs_get_inode(dir->i_sb, current_fsuid(), gid, mode, dev); 5151da177e4SLinus Torvalds if (inode) { 5161da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 5171da177e4SLinus Torvalds d_instantiate(dentry, inode); 5181da177e4SLinus Torvalds dget(dentry); /* Extra count - pin the dentry in core */ 5191da177e4SLinus Torvalds error = 0; 5201da177e4SLinus Torvalds } 5211da177e4SLinus Torvalds return error; 5221da177e4SLinus Torvalds } 5231da177e4SLinus Torvalds 5241da177e4SLinus Torvalds static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, int mode) 5251da177e4SLinus Torvalds { 5261da177e4SLinus Torvalds int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0); 5271da177e4SLinus Torvalds if (!retval) 528d8c76e6fSDave Hansen inc_nlink(dir); 5291da177e4SLinus Torvalds return retval; 5301da177e4SLinus Torvalds } 5311da177e4SLinus Torvalds 5321da177e4SLinus Torvalds static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, int mode, struct nameidata *nd) 5331da177e4SLinus Torvalds { 5341da177e4SLinus Torvalds return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0); 5351da177e4SLinus Torvalds } 5361da177e4SLinus Torvalds 5371da177e4SLinus Torvalds static int hugetlbfs_symlink(struct inode *dir, 5381da177e4SLinus Torvalds struct dentry *dentry, const char *symname) 5391da177e4SLinus Torvalds { 5401da177e4SLinus Torvalds struct inode *inode; 5411da177e4SLinus Torvalds int error = -ENOSPC; 5421da177e4SLinus Torvalds gid_t gid; 5431da177e4SLinus Torvalds 5441da177e4SLinus Torvalds if (dir->i_mode & S_ISGID) 5451da177e4SLinus Torvalds gid = dir->i_gid; 5461da177e4SLinus Torvalds else 54777c70de1SDavid Howells gid = current_fsgid(); 5481da177e4SLinus Torvalds 54977c70de1SDavid Howells inode = hugetlbfs_get_inode(dir->i_sb, current_fsuid(), 5501da177e4SLinus Torvalds gid, S_IFLNK|S_IRWXUGO, 0); 5511da177e4SLinus Torvalds if (inode) { 5521da177e4SLinus Torvalds int l = strlen(symname)+1; 5531da177e4SLinus Torvalds error = page_symlink(inode, symname, l); 5541da177e4SLinus Torvalds if (!error) { 5551da177e4SLinus Torvalds d_instantiate(dentry, inode); 5561da177e4SLinus Torvalds dget(dentry); 5571da177e4SLinus Torvalds } else 5581da177e4SLinus Torvalds iput(inode); 5591da177e4SLinus Torvalds } 5601da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 5611da177e4SLinus Torvalds 5621da177e4SLinus Torvalds return error; 5631da177e4SLinus Torvalds } 5641da177e4SLinus Torvalds 5651da177e4SLinus Torvalds /* 5666649a386SKen Chen * mark the head page dirty 5671da177e4SLinus Torvalds */ 5681da177e4SLinus Torvalds static int hugetlbfs_set_page_dirty(struct page *page) 5691da177e4SLinus Torvalds { 570d85f3385SChristoph Lameter struct page *head = compound_head(page); 5716649a386SKen Chen 5726649a386SKen Chen SetPageDirty(head); 5731da177e4SLinus Torvalds return 0; 5741da177e4SLinus Torvalds } 5751da177e4SLinus Torvalds 576726c3342SDavid Howells static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf) 5771da177e4SLinus Torvalds { 578726c3342SDavid Howells struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb); 579a5516438SAndi Kleen struct hstate *h = hstate_inode(dentry->d_inode); 5801da177e4SLinus Torvalds 5811da177e4SLinus Torvalds buf->f_type = HUGETLBFS_MAGIC; 582a5516438SAndi Kleen buf->f_bsize = huge_page_size(h); 5831da177e4SLinus Torvalds if (sbinfo) { 5841da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 58574a8a65cSDavid Gibson /* If no limits set, just report 0 for max/free/used 58674a8a65cSDavid Gibson * blocks, like simple_statfs() */ 58774a8a65cSDavid Gibson if (sbinfo->max_blocks >= 0) { 5881da177e4SLinus Torvalds buf->f_blocks = sbinfo->max_blocks; 5891da177e4SLinus Torvalds buf->f_bavail = buf->f_bfree = sbinfo->free_blocks; 5901da177e4SLinus Torvalds buf->f_files = sbinfo->max_inodes; 5911da177e4SLinus Torvalds buf->f_ffree = sbinfo->free_inodes; 59274a8a65cSDavid Gibson } 5931da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 5941da177e4SLinus Torvalds } 5951da177e4SLinus Torvalds buf->f_namelen = NAME_MAX; 5961da177e4SLinus Torvalds return 0; 5971da177e4SLinus Torvalds } 5981da177e4SLinus Torvalds 5991da177e4SLinus Torvalds static void hugetlbfs_put_super(struct super_block *sb) 6001da177e4SLinus Torvalds { 6011da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb); 6021da177e4SLinus Torvalds 6031da177e4SLinus Torvalds if (sbi) { 6041da177e4SLinus Torvalds sb->s_fs_info = NULL; 6051da177e4SLinus Torvalds kfree(sbi); 6061da177e4SLinus Torvalds } 6071da177e4SLinus Torvalds } 6081da177e4SLinus Torvalds 60996527980SChristoph Hellwig static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo) 61096527980SChristoph Hellwig { 61196527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 61296527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 61396527980SChristoph Hellwig if (unlikely(!sbinfo->free_inodes)) { 61496527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 61596527980SChristoph Hellwig return 0; 61696527980SChristoph Hellwig } 61796527980SChristoph Hellwig sbinfo->free_inodes--; 61896527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 61996527980SChristoph Hellwig } 62096527980SChristoph Hellwig 62196527980SChristoph Hellwig return 1; 62296527980SChristoph Hellwig } 62396527980SChristoph Hellwig 62496527980SChristoph Hellwig static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo) 62596527980SChristoph Hellwig { 62696527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 62796527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 62896527980SChristoph Hellwig sbinfo->free_inodes++; 62996527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 63096527980SChristoph Hellwig } 63196527980SChristoph Hellwig } 63296527980SChristoph Hellwig 63396527980SChristoph Hellwig 634e18b890bSChristoph Lameter static struct kmem_cache *hugetlbfs_inode_cachep; 6351da177e4SLinus Torvalds 6361da177e4SLinus Torvalds static struct inode *hugetlbfs_alloc_inode(struct super_block *sb) 6371da177e4SLinus Torvalds { 63896527980SChristoph Hellwig struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb); 6391da177e4SLinus Torvalds struct hugetlbfs_inode_info *p; 6401da177e4SLinus Torvalds 64196527980SChristoph Hellwig if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo))) 6421da177e4SLinus Torvalds return NULL; 643e94b1766SChristoph Lameter p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL); 64496527980SChristoph Hellwig if (unlikely(!p)) { 64596527980SChristoph Hellwig hugetlbfs_inc_free_inodes(sbinfo); 64696527980SChristoph Hellwig return NULL; 6471da177e4SLinus Torvalds } 64896527980SChristoph Hellwig return &p->vfs_inode; 6491da177e4SLinus Torvalds } 6501da177e4SLinus Torvalds 6511da177e4SLinus Torvalds static void hugetlbfs_destroy_inode(struct inode *inode) 6521da177e4SLinus Torvalds { 65396527980SChristoph Hellwig hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb)); 6541da177e4SLinus Torvalds mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy); 6551da177e4SLinus Torvalds kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode)); 6561da177e4SLinus Torvalds } 6571da177e4SLinus Torvalds 658f5e54d6eSChristoph Hellwig static const struct address_space_operations hugetlbfs_aops = { 659800d15a5SNick Piggin .write_begin = hugetlbfs_write_begin, 660800d15a5SNick Piggin .write_end = hugetlbfs_write_end, 6611da177e4SLinus Torvalds .set_page_dirty = hugetlbfs_set_page_dirty, 6621da177e4SLinus Torvalds }; 6631da177e4SLinus Torvalds 66496527980SChristoph Hellwig 66551cc5068SAlexey Dobriyan static void init_once(void *foo) 66696527980SChristoph Hellwig { 66796527980SChristoph Hellwig struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo; 66896527980SChristoph Hellwig 66996527980SChristoph Hellwig inode_init_once(&ei->vfs_inode); 67096527980SChristoph Hellwig } 67196527980SChristoph Hellwig 6724b6f5d20SArjan van de Ven const struct file_operations hugetlbfs_file_operations = { 673e63e1e5aSBadari Pulavarty .read = hugetlbfs_read, 6741da177e4SLinus Torvalds .mmap = hugetlbfs_file_mmap, 6751b061d92SChristoph Hellwig .fsync = noop_fsync, 6761da177e4SLinus Torvalds .get_unmapped_area = hugetlb_get_unmapped_area, 6771da177e4SLinus Torvalds }; 6781da177e4SLinus Torvalds 67992e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_dir_inode_operations = { 6801da177e4SLinus Torvalds .create = hugetlbfs_create, 6811da177e4SLinus Torvalds .lookup = simple_lookup, 6821da177e4SLinus Torvalds .link = simple_link, 6831da177e4SLinus Torvalds .unlink = simple_unlink, 6841da177e4SLinus Torvalds .symlink = hugetlbfs_symlink, 6851da177e4SLinus Torvalds .mkdir = hugetlbfs_mkdir, 6861da177e4SLinus Torvalds .rmdir = simple_rmdir, 6871da177e4SLinus Torvalds .mknod = hugetlbfs_mknod, 6881da177e4SLinus Torvalds .rename = simple_rename, 6891da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 6901da177e4SLinus Torvalds }; 6911da177e4SLinus Torvalds 69292e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_inode_operations = { 6931da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 6941da177e4SLinus Torvalds }; 6951da177e4SLinus Torvalds 696ee9b6d61SJosef 'Jeff' Sipek static const struct super_operations hugetlbfs_ops = { 6971da177e4SLinus Torvalds .alloc_inode = hugetlbfs_alloc_inode, 6981da177e4SLinus Torvalds .destroy_inode = hugetlbfs_destroy_inode, 6992bbbda30SAl Viro .evict_inode = hugetlbfs_evict_inode, 7001da177e4SLinus Torvalds .statfs = hugetlbfs_statfs, 7011da177e4SLinus Torvalds .put_super = hugetlbfs_put_super, 70210f19a86SMiklos Szeredi .show_options = generic_show_options, 7031da177e4SLinus Torvalds }; 7041da177e4SLinus Torvalds 7051da177e4SLinus Torvalds static int 7061da177e4SLinus Torvalds hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig) 7071da177e4SLinus Torvalds { 708e73a75faSRandy Dunlap char *p, *rest; 709e73a75faSRandy Dunlap substring_t args[MAX_OPT_ARGS]; 710e73a75faSRandy Dunlap int option; 711a137e1ccSAndi Kleen unsigned long long size = 0; 712a137e1ccSAndi Kleen enum { NO_SIZE, SIZE_STD, SIZE_PERCENT } setsize = NO_SIZE; 7131da177e4SLinus Torvalds 7141da177e4SLinus Torvalds if (!options) 7151da177e4SLinus Torvalds return 0; 7161da177e4SLinus Torvalds 717e73a75faSRandy Dunlap while ((p = strsep(&options, ",")) != NULL) { 718e73a75faSRandy Dunlap int token; 719b4c07bceSLee Schermerhorn if (!*p) 720b4c07bceSLee Schermerhorn continue; 7211da177e4SLinus Torvalds 722e73a75faSRandy Dunlap token = match_token(p, tokens, args); 723e73a75faSRandy Dunlap switch (token) { 724e73a75faSRandy Dunlap case Opt_uid: 725e73a75faSRandy Dunlap if (match_int(&args[0], &option)) 726e73a75faSRandy Dunlap goto bad_val; 727e73a75faSRandy Dunlap pconfig->uid = option; 728e73a75faSRandy Dunlap break; 729e73a75faSRandy Dunlap 730e73a75faSRandy Dunlap case Opt_gid: 731e73a75faSRandy Dunlap if (match_int(&args[0], &option)) 732e73a75faSRandy Dunlap goto bad_val; 733e73a75faSRandy Dunlap pconfig->gid = option; 734e73a75faSRandy Dunlap break; 735e73a75faSRandy Dunlap 736e73a75faSRandy Dunlap case Opt_mode: 737e73a75faSRandy Dunlap if (match_octal(&args[0], &option)) 738e73a75faSRandy Dunlap goto bad_val; 73975897d60SKen Chen pconfig->mode = option & 01777U; 740e73a75faSRandy Dunlap break; 741e73a75faSRandy Dunlap 742e73a75faSRandy Dunlap case Opt_size: { 743e73a75faSRandy Dunlap /* memparse() will accept a K/M/G without a digit */ 744e73a75faSRandy Dunlap if (!isdigit(*args[0].from)) 745e73a75faSRandy Dunlap goto bad_val; 746e73a75faSRandy Dunlap size = memparse(args[0].from, &rest); 747a137e1ccSAndi Kleen setsize = SIZE_STD; 748a137e1ccSAndi Kleen if (*rest == '%') 749a137e1ccSAndi Kleen setsize = SIZE_PERCENT; 750e73a75faSRandy Dunlap break; 751e73a75faSRandy Dunlap } 7521da177e4SLinus Torvalds 753e73a75faSRandy Dunlap case Opt_nr_inodes: 754e73a75faSRandy Dunlap /* memparse() will accept a K/M/G without a digit */ 755e73a75faSRandy Dunlap if (!isdigit(*args[0].from)) 756e73a75faSRandy Dunlap goto bad_val; 757e73a75faSRandy Dunlap pconfig->nr_inodes = memparse(args[0].from, &rest); 758e73a75faSRandy Dunlap break; 759e73a75faSRandy Dunlap 760a137e1ccSAndi Kleen case Opt_pagesize: { 761a137e1ccSAndi Kleen unsigned long ps; 762a137e1ccSAndi Kleen ps = memparse(args[0].from, &rest); 763a137e1ccSAndi Kleen pconfig->hstate = size_to_hstate(ps); 764a137e1ccSAndi Kleen if (!pconfig->hstate) { 765a137e1ccSAndi Kleen printk(KERN_ERR 766a137e1ccSAndi Kleen "hugetlbfs: Unsupported page size %lu MB\n", 767a137e1ccSAndi Kleen ps >> 20); 768a137e1ccSAndi Kleen return -EINVAL; 769a137e1ccSAndi Kleen } 770a137e1ccSAndi Kleen break; 771a137e1ccSAndi Kleen } 772a137e1ccSAndi Kleen 773e73a75faSRandy Dunlap default: 774b4c07bceSLee Schermerhorn printk(KERN_ERR "hugetlbfs: Bad mount option: \"%s\"\n", 775b4c07bceSLee Schermerhorn p); 776b4c07bceSLee Schermerhorn return -EINVAL; 777e73a75faSRandy Dunlap break; 778e73a75faSRandy Dunlap } 7791da177e4SLinus Torvalds } 780a137e1ccSAndi Kleen 781a137e1ccSAndi Kleen /* Do size after hstate is set up */ 782a137e1ccSAndi Kleen if (setsize > NO_SIZE) { 783a137e1ccSAndi Kleen struct hstate *h = pconfig->hstate; 784a137e1ccSAndi Kleen if (setsize == SIZE_PERCENT) { 785a137e1ccSAndi Kleen size <<= huge_page_shift(h); 786a137e1ccSAndi Kleen size *= h->max_huge_pages; 787a137e1ccSAndi Kleen do_div(size, 100); 788a137e1ccSAndi Kleen } 789a137e1ccSAndi Kleen pconfig->nr_blocks = (size >> huge_page_shift(h)); 790a137e1ccSAndi Kleen } 791a137e1ccSAndi Kleen 7921da177e4SLinus Torvalds return 0; 793e73a75faSRandy Dunlap 794e73a75faSRandy Dunlap bad_val: 795e73a75faSRandy Dunlap printk(KERN_ERR "hugetlbfs: Bad value '%s' for mount option '%s'\n", 796e73a75faSRandy Dunlap args[0].from, p); 797c12ddba0SAkinobu Mita return -EINVAL; 7981da177e4SLinus Torvalds } 7991da177e4SLinus Torvalds 8001da177e4SLinus Torvalds static int 8011da177e4SLinus Torvalds hugetlbfs_fill_super(struct super_block *sb, void *data, int silent) 8021da177e4SLinus Torvalds { 8031da177e4SLinus Torvalds struct inode * inode; 8041da177e4SLinus Torvalds struct dentry * root; 8051da177e4SLinus Torvalds int ret; 8061da177e4SLinus Torvalds struct hugetlbfs_config config; 8071da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo; 8081da177e4SLinus Torvalds 80910f19a86SMiklos Szeredi save_mount_options(sb, data); 81010f19a86SMiklos Szeredi 8111da177e4SLinus Torvalds config.nr_blocks = -1; /* No limit on size by default */ 8121da177e4SLinus Torvalds config.nr_inodes = -1; /* No limit on number of inodes by default */ 81377c70de1SDavid Howells config.uid = current_fsuid(); 81477c70de1SDavid Howells config.gid = current_fsgid(); 8151da177e4SLinus Torvalds config.mode = 0755; 816a137e1ccSAndi Kleen config.hstate = &default_hstate; 8171da177e4SLinus Torvalds ret = hugetlbfs_parse_options(data, &config); 8181da177e4SLinus Torvalds if (ret) 8191da177e4SLinus Torvalds return ret; 8201da177e4SLinus Torvalds 8211da177e4SLinus Torvalds sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL); 8221da177e4SLinus Torvalds if (!sbinfo) 8231da177e4SLinus Torvalds return -ENOMEM; 8241da177e4SLinus Torvalds sb->s_fs_info = sbinfo; 825a137e1ccSAndi Kleen sbinfo->hstate = config.hstate; 8261da177e4SLinus Torvalds spin_lock_init(&sbinfo->stat_lock); 8271da177e4SLinus Torvalds sbinfo->max_blocks = config.nr_blocks; 8281da177e4SLinus Torvalds sbinfo->free_blocks = config.nr_blocks; 8291da177e4SLinus Torvalds sbinfo->max_inodes = config.nr_inodes; 8301da177e4SLinus Torvalds sbinfo->free_inodes = config.nr_inodes; 8311da177e4SLinus Torvalds sb->s_maxbytes = MAX_LFS_FILESIZE; 832a137e1ccSAndi Kleen sb->s_blocksize = huge_page_size(config.hstate); 833a137e1ccSAndi Kleen sb->s_blocksize_bits = huge_page_shift(config.hstate); 8341da177e4SLinus Torvalds sb->s_magic = HUGETLBFS_MAGIC; 8351da177e4SLinus Torvalds sb->s_op = &hugetlbfs_ops; 8361da177e4SLinus Torvalds sb->s_time_gran = 1; 8371da177e4SLinus Torvalds inode = hugetlbfs_get_inode(sb, config.uid, config.gid, 8381da177e4SLinus Torvalds S_IFDIR | config.mode, 0); 8391da177e4SLinus Torvalds if (!inode) 8401da177e4SLinus Torvalds goto out_free; 8411da177e4SLinus Torvalds 8421da177e4SLinus Torvalds root = d_alloc_root(inode); 8431da177e4SLinus Torvalds if (!root) { 8441da177e4SLinus Torvalds iput(inode); 8451da177e4SLinus Torvalds goto out_free; 8461da177e4SLinus Torvalds } 8471da177e4SLinus Torvalds sb->s_root = root; 8481da177e4SLinus Torvalds return 0; 8491da177e4SLinus Torvalds out_free: 8501da177e4SLinus Torvalds kfree(sbinfo); 8511da177e4SLinus Torvalds return -ENOMEM; 8521da177e4SLinus Torvalds } 8531da177e4SLinus Torvalds 8549a119c05SAdam Litke int hugetlb_get_quota(struct address_space *mapping, long delta) 8551da177e4SLinus Torvalds { 8561da177e4SLinus Torvalds int ret = 0; 8571da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(mapping->host->i_sb); 8581da177e4SLinus Torvalds 8591da177e4SLinus Torvalds if (sbinfo->free_blocks > -1) { 8601da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 8619a119c05SAdam Litke if (sbinfo->free_blocks - delta >= 0) 8629a119c05SAdam Litke sbinfo->free_blocks -= delta; 8631da177e4SLinus Torvalds else 8641da177e4SLinus Torvalds ret = -ENOMEM; 8651da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 8661da177e4SLinus Torvalds } 8671da177e4SLinus Torvalds 8681da177e4SLinus Torvalds return ret; 8691da177e4SLinus Torvalds } 8701da177e4SLinus Torvalds 8719a119c05SAdam Litke void hugetlb_put_quota(struct address_space *mapping, long delta) 8721da177e4SLinus Torvalds { 8731da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(mapping->host->i_sb); 8741da177e4SLinus Torvalds 8751da177e4SLinus Torvalds if (sbinfo->free_blocks > -1) { 8761da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 8779a119c05SAdam Litke sbinfo->free_blocks += delta; 8781da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 8791da177e4SLinus Torvalds } 8801da177e4SLinus Torvalds } 8811da177e4SLinus Torvalds 882454e2398SDavid Howells static int hugetlbfs_get_sb(struct file_system_type *fs_type, 883454e2398SDavid Howells int flags, const char *dev_name, void *data, struct vfsmount *mnt) 8841da177e4SLinus Torvalds { 885454e2398SDavid Howells return get_sb_nodev(fs_type, flags, data, hugetlbfs_fill_super, mnt); 8861da177e4SLinus Torvalds } 8871da177e4SLinus Torvalds 8881da177e4SLinus Torvalds static struct file_system_type hugetlbfs_fs_type = { 8891da177e4SLinus Torvalds .name = "hugetlbfs", 8901da177e4SLinus Torvalds .get_sb = hugetlbfs_get_sb, 8911da177e4SLinus Torvalds .kill_sb = kill_litter_super, 8921da177e4SLinus Torvalds }; 8931da177e4SLinus Torvalds 8941da177e4SLinus Torvalds static struct vfsmount *hugetlbfs_vfsmount; 8951da177e4SLinus Torvalds 896ef1ff6b8SFrom: Mel Gorman static int can_do_hugetlb_shm(void) 8971da177e4SLinus Torvalds { 898ef1ff6b8SFrom: Mel Gorman return capable(CAP_IPC_LOCK) || in_group_p(sysctl_hugetlb_shm_group); 8991da177e4SLinus Torvalds } 9001da177e4SLinus Torvalds 901353d5c30SHugh Dickins struct file *hugetlb_file_setup(const char *name, size_t size, int acctflag, 9026bfde05bSEric B Munson struct user_struct **user, int creat_flags) 9031da177e4SLinus Torvalds { 9041da177e4SLinus Torvalds int error = -ENOMEM; 9051da177e4SLinus Torvalds struct file *file; 9061da177e4SLinus Torvalds struct inode *inode; 9072c48b9c4SAl Viro struct path path; 9082c48b9c4SAl Viro struct dentry *root; 9091da177e4SLinus Torvalds struct qstr quick_string; 9101da177e4SLinus Torvalds 911353d5c30SHugh Dickins *user = NULL; 9125bc98594SAkinobu Mita if (!hugetlbfs_vfsmount) 9135bc98594SAkinobu Mita return ERR_PTR(-ENOENT); 9145bc98594SAkinobu Mita 915ef1ff6b8SFrom: Mel Gorman if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) { 916353d5c30SHugh Dickins *user = current_user(); 917353d5c30SHugh Dickins if (user_shm_lock(size, *user)) { 9182584e517SRavikiran G Thirumalai WARN_ONCE(1, 9192584e517SRavikiran G Thirumalai "Using mlock ulimits for SHM_HUGETLB deprecated\n"); 920353d5c30SHugh Dickins } else { 921353d5c30SHugh Dickins *user = NULL; 9221da177e4SLinus Torvalds return ERR_PTR(-EPERM); 9232584e517SRavikiran G Thirumalai } 924353d5c30SHugh Dickins } 9251da177e4SLinus Torvalds 9261da177e4SLinus Torvalds root = hugetlbfs_vfsmount->mnt_root; 9279d66586fSEric W. Biederman quick_string.name = name; 9281da177e4SLinus Torvalds quick_string.len = strlen(quick_string.name); 9291da177e4SLinus Torvalds quick_string.hash = 0; 9302c48b9c4SAl Viro path.dentry = d_alloc(root, &quick_string); 9312c48b9c4SAl Viro if (!path.dentry) 9321da177e4SLinus Torvalds goto out_shm_unlock; 9331da177e4SLinus Torvalds 9342c48b9c4SAl Viro path.mnt = mntget(hugetlbfs_vfsmount); 9351da177e4SLinus Torvalds error = -ENOSPC; 93677c70de1SDavid Howells inode = hugetlbfs_get_inode(root->d_sb, current_fsuid(), 93777c70de1SDavid Howells current_fsgid(), S_IFREG | S_IRWXUGO, 0); 9381da177e4SLinus Torvalds if (!inode) 939ce8d2cdfSDave Hansen goto out_dentry; 9401da177e4SLinus Torvalds 941b45b5bd6SDavid Gibson error = -ENOMEM; 942a5516438SAndi Kleen if (hugetlb_reserve_pages(inode, 0, 9435a6fe125SMel Gorman size >> huge_page_shift(hstate_inode(inode)), NULL, 9445a6fe125SMel Gorman acctflag)) 945b45b5bd6SDavid Gibson goto out_inode; 946b45b5bd6SDavid Gibson 9472c48b9c4SAl Viro d_instantiate(path.dentry, inode); 9481da177e4SLinus Torvalds inode->i_size = size; 9491da177e4SLinus Torvalds inode->i_nlink = 0; 950ce8d2cdfSDave Hansen 951ce8d2cdfSDave Hansen error = -ENFILE; 9522c48b9c4SAl Viro file = alloc_file(&path, FMODE_WRITE | FMODE_READ, 953ce8d2cdfSDave Hansen &hugetlbfs_file_operations); 954ce8d2cdfSDave Hansen if (!file) 955b4d232e6SAl Viro goto out_dentry; /* inode is already attached */ 956ce8d2cdfSDave Hansen 9571da177e4SLinus Torvalds return file; 9581da177e4SLinus Torvalds 959b45b5bd6SDavid Gibson out_inode: 960b45b5bd6SDavid Gibson iput(inode); 9611da177e4SLinus Torvalds out_dentry: 9622c48b9c4SAl Viro path_put(&path); 9631da177e4SLinus Torvalds out_shm_unlock: 964353d5c30SHugh Dickins if (*user) { 965353d5c30SHugh Dickins user_shm_unlock(size, *user); 966353d5c30SHugh Dickins *user = NULL; 967353d5c30SHugh Dickins } 9681da177e4SLinus Torvalds return ERR_PTR(error); 9691da177e4SLinus Torvalds } 9701da177e4SLinus Torvalds 9711da177e4SLinus Torvalds static int __init init_hugetlbfs_fs(void) 9721da177e4SLinus Torvalds { 9731da177e4SLinus Torvalds int error; 9741da177e4SLinus Torvalds struct vfsmount *vfsmount; 9751da177e4SLinus Torvalds 976e0bf68ddSPeter Zijlstra error = bdi_init(&hugetlbfs_backing_dev_info); 977e0bf68ddSPeter Zijlstra if (error) 978e0bf68ddSPeter Zijlstra return error; 979e0bf68ddSPeter Zijlstra 9801da177e4SLinus Torvalds hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache", 9811da177e4SLinus Torvalds sizeof(struct hugetlbfs_inode_info), 98220c2df83SPaul Mundt 0, 0, init_once); 9831da177e4SLinus Torvalds if (hugetlbfs_inode_cachep == NULL) 984e0bf68ddSPeter Zijlstra goto out2; 9851da177e4SLinus Torvalds 9861da177e4SLinus Torvalds error = register_filesystem(&hugetlbfs_fs_type); 9871da177e4SLinus Torvalds if (error) 9881da177e4SLinus Torvalds goto out; 9891da177e4SLinus Torvalds 9901da177e4SLinus Torvalds vfsmount = kern_mount(&hugetlbfs_fs_type); 9911da177e4SLinus Torvalds 9921da177e4SLinus Torvalds if (!IS_ERR(vfsmount)) { 9931da177e4SLinus Torvalds hugetlbfs_vfsmount = vfsmount; 9941da177e4SLinus Torvalds return 0; 9951da177e4SLinus Torvalds } 9961da177e4SLinus Torvalds 9971da177e4SLinus Torvalds error = PTR_ERR(vfsmount); 9981da177e4SLinus Torvalds 9991da177e4SLinus Torvalds out: 10001da177e4SLinus Torvalds if (error) 10011da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 1002e0bf68ddSPeter Zijlstra out2: 1003e0bf68ddSPeter Zijlstra bdi_destroy(&hugetlbfs_backing_dev_info); 10041da177e4SLinus Torvalds return error; 10051da177e4SLinus Torvalds } 10061da177e4SLinus Torvalds 10071da177e4SLinus Torvalds static void __exit exit_hugetlbfs_fs(void) 10081da177e4SLinus Torvalds { 10091da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 10101da177e4SLinus Torvalds unregister_filesystem(&hugetlbfs_fs_type); 1011e0bf68ddSPeter Zijlstra bdi_destroy(&hugetlbfs_backing_dev_info); 10121da177e4SLinus Torvalds } 10131da177e4SLinus Torvalds 10141da177e4SLinus Torvalds module_init(init_hugetlbfs_fs) 10151da177e4SLinus Torvalds module_exit(exit_hugetlbfs_fs) 10161da177e4SLinus Torvalds 10171da177e4SLinus Torvalds MODULE_LICENSE("GPL"); 1018