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/quotaops.h> 301da177e4SLinus Torvalds #include <linux/slab.h> 311da177e4SLinus Torvalds #include <linux/dnotify.h> 321da177e4SLinus Torvalds #include <linux/statfs.h> 331da177e4SLinus Torvalds #include <linux/security.h> 341da177e4SLinus Torvalds 351da177e4SLinus Torvalds #include <asm/uaccess.h> 361da177e4SLinus Torvalds 371da177e4SLinus Torvalds /* some random number */ 381da177e4SLinus Torvalds #define HUGETLBFS_MAGIC 0x958458f6 391da177e4SLinus Torvalds 40ee9b6d61SJosef 'Jeff' Sipek static const struct super_operations hugetlbfs_ops; 41f5e54d6eSChristoph Hellwig static const struct address_space_operations hugetlbfs_aops; 424b6f5d20SArjan van de Ven const struct file_operations hugetlbfs_file_operations; 4392e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_dir_inode_operations; 4492e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_inode_operations; 451da177e4SLinus Torvalds 461da177e4SLinus Torvalds static struct backing_dev_info hugetlbfs_backing_dev_info = { 471da177e4SLinus Torvalds .ra_pages = 0, /* No readahead */ 48e4ad08feSMiklos Szeredi .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK, 491da177e4SLinus Torvalds }; 501da177e4SLinus Torvalds 511da177e4SLinus Torvalds int sysctl_hugetlb_shm_group; 521da177e4SLinus Torvalds 53e73a75faSRandy Dunlap enum { 54e73a75faSRandy Dunlap Opt_size, Opt_nr_inodes, 55e73a75faSRandy Dunlap Opt_mode, Opt_uid, Opt_gid, 56a137e1ccSAndi Kleen Opt_pagesize, 57e73a75faSRandy Dunlap Opt_err, 58e73a75faSRandy Dunlap }; 59e73a75faSRandy Dunlap 60e73a75faSRandy Dunlap static match_table_t tokens = { 61e73a75faSRandy Dunlap {Opt_size, "size=%s"}, 62e73a75faSRandy Dunlap {Opt_nr_inodes, "nr_inodes=%s"}, 63e73a75faSRandy Dunlap {Opt_mode, "mode=%o"}, 64e73a75faSRandy Dunlap {Opt_uid, "uid=%u"}, 65e73a75faSRandy Dunlap {Opt_gid, "gid=%u"}, 66a137e1ccSAndi Kleen {Opt_pagesize, "pagesize=%s"}, 67e73a75faSRandy Dunlap {Opt_err, NULL}, 68e73a75faSRandy Dunlap }; 69e73a75faSRandy Dunlap 702e9b367cSAdam Litke static void huge_pagevec_release(struct pagevec *pvec) 712e9b367cSAdam Litke { 722e9b367cSAdam Litke int i; 732e9b367cSAdam Litke 742e9b367cSAdam Litke for (i = 0; i < pagevec_count(pvec); ++i) 752e9b367cSAdam Litke put_page(pvec->pages[i]); 762e9b367cSAdam Litke 772e9b367cSAdam Litke pagevec_reinit(pvec); 782e9b367cSAdam Litke } 792e9b367cSAdam Litke 801da177e4SLinus Torvalds static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma) 811da177e4SLinus Torvalds { 82b39424e2SJosef Sipek struct inode *inode = file->f_path.dentry->d_inode; 831da177e4SLinus Torvalds loff_t len, vma_len; 841da177e4SLinus Torvalds int ret; 85a5516438SAndi Kleen struct hstate *h = hstate_file(file); 861da177e4SLinus Torvalds 8768589bc3SHugh Dickins /* 88dec4ad86SDavid Gibson * vma address alignment (but not the pgoff alignment) has 89dec4ad86SDavid Gibson * already been checked by prepare_hugepage_range. If you add 90dec4ad86SDavid Gibson * any error returns here, do so after setting VM_HUGETLB, so 91dec4ad86SDavid Gibson * is_vm_hugetlb_page tests below unmap_region go the right 92dec4ad86SDavid Gibson * way when do_mmap_pgoff unwinds (may be important on powerpc 93dec4ad86SDavid Gibson * and ia64). 9468589bc3SHugh Dickins */ 9568589bc3SHugh Dickins vma->vm_flags |= VM_HUGETLB | VM_RESERVED; 9668589bc3SHugh Dickins vma->vm_ops = &hugetlb_vm_ops; 971da177e4SLinus Torvalds 98a5516438SAndi Kleen if (vma->vm_pgoff & ~(huge_page_mask(h) >> PAGE_SHIFT)) 99dec4ad86SDavid Gibson return -EINVAL; 100dec4ad86SDavid Gibson 1011da177e4SLinus Torvalds vma_len = (loff_t)(vma->vm_end - vma->vm_start); 1021da177e4SLinus Torvalds 1031b1dcc1bSJes Sorensen mutex_lock(&inode->i_mutex); 1041da177e4SLinus Torvalds file_accessed(file); 1051da177e4SLinus Torvalds 1061da177e4SLinus Torvalds ret = -ENOMEM; 1071da177e4SLinus Torvalds len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); 1081da177e4SLinus Torvalds 109a1e78772SMel Gorman if (hugetlb_reserve_pages(inode, 110a5516438SAndi Kleen vma->vm_pgoff >> huge_page_order(h), 111a5516438SAndi Kleen len >> huge_page_shift(h), vma)) 112b45b5bd6SDavid Gibson goto out; 113b45b5bd6SDavid Gibson 1144c887265SAdam Litke ret = 0; 1154c887265SAdam Litke hugetlb_prefault_arch_hook(vma->vm_mm); 116b6174df5SZhang, Yanmin if (vma->vm_flags & VM_WRITE && inode->i_size < len) 1171da177e4SLinus Torvalds inode->i_size = len; 1181da177e4SLinus Torvalds out: 1191b1dcc1bSJes Sorensen mutex_unlock(&inode->i_mutex); 1201da177e4SLinus Torvalds 1211da177e4SLinus Torvalds return ret; 1221da177e4SLinus Torvalds } 1231da177e4SLinus Torvalds 1241da177e4SLinus Torvalds /* 125508034a3SHugh Dickins * Called under down_write(mmap_sem). 1261da177e4SLinus Torvalds */ 1271da177e4SLinus Torvalds 128d2ba27e8SAdrian Bunk #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA 1291da177e4SLinus Torvalds static unsigned long 1301da177e4SLinus Torvalds hugetlb_get_unmapped_area(struct file *file, unsigned long addr, 1311da177e4SLinus Torvalds unsigned long len, unsigned long pgoff, unsigned long flags) 1321da177e4SLinus Torvalds { 1331da177e4SLinus Torvalds struct mm_struct *mm = current->mm; 1341da177e4SLinus Torvalds struct vm_area_struct *vma; 1351da177e4SLinus Torvalds unsigned long start_addr; 136a5516438SAndi Kleen struct hstate *h = hstate_file(file); 1371da177e4SLinus Torvalds 138a5516438SAndi Kleen if (len & ~huge_page_mask(h)) 1391da177e4SLinus Torvalds return -EINVAL; 1401da177e4SLinus Torvalds if (len > TASK_SIZE) 1411da177e4SLinus Torvalds return -ENOMEM; 1421da177e4SLinus Torvalds 143036e0856SBenjamin Herrenschmidt if (flags & MAP_FIXED) { 144a5516438SAndi Kleen if (prepare_hugepage_range(file, addr, len)) 145036e0856SBenjamin Herrenschmidt return -EINVAL; 146036e0856SBenjamin Herrenschmidt return addr; 147036e0856SBenjamin Herrenschmidt } 148036e0856SBenjamin Herrenschmidt 1491da177e4SLinus Torvalds if (addr) { 150a5516438SAndi Kleen addr = ALIGN(addr, huge_page_size(h)); 1511da177e4SLinus Torvalds vma = find_vma(mm, addr); 1521da177e4SLinus Torvalds if (TASK_SIZE - len >= addr && 1531da177e4SLinus Torvalds (!vma || addr + len <= vma->vm_start)) 1541da177e4SLinus Torvalds return addr; 1551da177e4SLinus Torvalds } 1561da177e4SLinus Torvalds 1571da177e4SLinus Torvalds start_addr = mm->free_area_cache; 1581da177e4SLinus Torvalds 1591363c3cdSWolfgang Wander if (len <= mm->cached_hole_size) 1601363c3cdSWolfgang Wander start_addr = TASK_UNMAPPED_BASE; 1611363c3cdSWolfgang Wander 1621da177e4SLinus Torvalds full_search: 163a5516438SAndi Kleen addr = ALIGN(start_addr, huge_page_size(h)); 1641da177e4SLinus Torvalds 1651da177e4SLinus Torvalds for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1661da177e4SLinus Torvalds /* At this point: (!vma || addr < vma->vm_end). */ 1671da177e4SLinus Torvalds if (TASK_SIZE - len < addr) { 1681da177e4SLinus Torvalds /* 1691da177e4SLinus Torvalds * Start a new search - just in case we missed 1701da177e4SLinus Torvalds * some holes. 1711da177e4SLinus Torvalds */ 1721da177e4SLinus Torvalds if (start_addr != TASK_UNMAPPED_BASE) { 1731da177e4SLinus Torvalds start_addr = TASK_UNMAPPED_BASE; 1741da177e4SLinus Torvalds goto full_search; 1751da177e4SLinus Torvalds } 1761da177e4SLinus Torvalds return -ENOMEM; 1771da177e4SLinus Torvalds } 1781da177e4SLinus Torvalds 1791da177e4SLinus Torvalds if (!vma || addr + len <= vma->vm_start) 1801da177e4SLinus Torvalds return addr; 181a5516438SAndi Kleen addr = ALIGN(vma->vm_end, huge_page_size(h)); 1821da177e4SLinus Torvalds } 1831da177e4SLinus Torvalds } 1841da177e4SLinus Torvalds #endif 1851da177e4SLinus Torvalds 186e63e1e5aSBadari Pulavarty static int 187e63e1e5aSBadari Pulavarty hugetlbfs_read_actor(struct page *page, unsigned long offset, 188e63e1e5aSBadari Pulavarty char __user *buf, unsigned long count, 189e63e1e5aSBadari Pulavarty unsigned long size) 190e63e1e5aSBadari Pulavarty { 191e63e1e5aSBadari Pulavarty char *kaddr; 192e63e1e5aSBadari Pulavarty unsigned long left, copied = 0; 193e63e1e5aSBadari Pulavarty int i, chunksize; 194e63e1e5aSBadari Pulavarty 195e63e1e5aSBadari Pulavarty if (size > count) 196e63e1e5aSBadari Pulavarty size = count; 197e63e1e5aSBadari Pulavarty 198e63e1e5aSBadari Pulavarty /* Find which 4k chunk and offset with in that chunk */ 199e63e1e5aSBadari Pulavarty i = offset >> PAGE_CACHE_SHIFT; 200e63e1e5aSBadari Pulavarty offset = offset & ~PAGE_CACHE_MASK; 201e63e1e5aSBadari Pulavarty 202e63e1e5aSBadari Pulavarty while (size) { 203e63e1e5aSBadari Pulavarty chunksize = PAGE_CACHE_SIZE; 204e63e1e5aSBadari Pulavarty if (offset) 205e63e1e5aSBadari Pulavarty chunksize -= offset; 206e63e1e5aSBadari Pulavarty if (chunksize > size) 207e63e1e5aSBadari Pulavarty chunksize = size; 208e63e1e5aSBadari Pulavarty kaddr = kmap(&page[i]); 209e63e1e5aSBadari Pulavarty left = __copy_to_user(buf, kaddr + offset, chunksize); 210e63e1e5aSBadari Pulavarty kunmap(&page[i]); 211e63e1e5aSBadari Pulavarty if (left) { 212e63e1e5aSBadari Pulavarty copied += (chunksize - left); 213e63e1e5aSBadari Pulavarty break; 214e63e1e5aSBadari Pulavarty } 215e63e1e5aSBadari Pulavarty offset = 0; 216e63e1e5aSBadari Pulavarty size -= chunksize; 217e63e1e5aSBadari Pulavarty buf += chunksize; 218e63e1e5aSBadari Pulavarty copied += chunksize; 219e63e1e5aSBadari Pulavarty i++; 220e63e1e5aSBadari Pulavarty } 221e63e1e5aSBadari Pulavarty return copied ? copied : -EFAULT; 222e63e1e5aSBadari Pulavarty } 223e63e1e5aSBadari Pulavarty 224e63e1e5aSBadari Pulavarty /* 225e63e1e5aSBadari Pulavarty * Support for read() - Find the page attached to f_mapping and copy out the 226e63e1e5aSBadari Pulavarty * data. Its *very* similar to do_generic_mapping_read(), we can't use that 227e63e1e5aSBadari Pulavarty * since it has PAGE_CACHE_SIZE assumptions. 228e63e1e5aSBadari Pulavarty */ 229e63e1e5aSBadari Pulavarty static ssize_t hugetlbfs_read(struct file *filp, char __user *buf, 230e63e1e5aSBadari Pulavarty size_t len, loff_t *ppos) 231e63e1e5aSBadari Pulavarty { 232a5516438SAndi Kleen struct hstate *h = hstate_file(filp); 233e63e1e5aSBadari Pulavarty struct address_space *mapping = filp->f_mapping; 234e63e1e5aSBadari Pulavarty struct inode *inode = mapping->host; 235a5516438SAndi Kleen unsigned long index = *ppos >> huge_page_shift(h); 236a5516438SAndi Kleen unsigned long offset = *ppos & ~huge_page_mask(h); 237e63e1e5aSBadari Pulavarty unsigned long end_index; 238e63e1e5aSBadari Pulavarty loff_t isize; 239e63e1e5aSBadari Pulavarty ssize_t retval = 0; 240e63e1e5aSBadari Pulavarty 241e63e1e5aSBadari Pulavarty mutex_lock(&inode->i_mutex); 242e63e1e5aSBadari Pulavarty 243e63e1e5aSBadari Pulavarty /* validate length */ 244e63e1e5aSBadari Pulavarty if (len == 0) 245e63e1e5aSBadari Pulavarty goto out; 246e63e1e5aSBadari Pulavarty 247e63e1e5aSBadari Pulavarty isize = i_size_read(inode); 248e63e1e5aSBadari Pulavarty if (!isize) 249e63e1e5aSBadari Pulavarty goto out; 250e63e1e5aSBadari Pulavarty 251a5516438SAndi Kleen end_index = (isize - 1) >> huge_page_shift(h); 252e63e1e5aSBadari Pulavarty for (;;) { 253e63e1e5aSBadari Pulavarty struct page *page; 254a5516438SAndi Kleen unsigned long nr, ret; 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)) 277e63e1e5aSBadari Pulavarty ret = -EFAULT; 278e63e1e5aSBadari Pulavarty } else { 279e63e1e5aSBadari Pulavarty /* 280e63e1e5aSBadari Pulavarty * We have the page, copy it to user space buffer. 281e63e1e5aSBadari Pulavarty */ 282e63e1e5aSBadari Pulavarty ret = hugetlbfs_read_actor(page, offset, buf, len, nr); 283e63e1e5aSBadari Pulavarty } 284e63e1e5aSBadari Pulavarty if (ret < 0) { 285e63e1e5aSBadari Pulavarty if (retval == 0) 286e63e1e5aSBadari Pulavarty retval = ret; 287e63e1e5aSBadari Pulavarty if (page) 288e63e1e5aSBadari Pulavarty page_cache_release(page); 289e63e1e5aSBadari Pulavarty goto out; 290e63e1e5aSBadari Pulavarty } 291e63e1e5aSBadari Pulavarty 292e63e1e5aSBadari Pulavarty offset += ret; 293e63e1e5aSBadari Pulavarty retval += ret; 294e63e1e5aSBadari Pulavarty len -= ret; 295a5516438SAndi Kleen index += offset >> huge_page_shift(h); 296a5516438SAndi Kleen offset &= ~huge_page_mask(h); 297e63e1e5aSBadari Pulavarty 298e63e1e5aSBadari Pulavarty if (page) 299e63e1e5aSBadari Pulavarty page_cache_release(page); 300e63e1e5aSBadari Pulavarty 301e63e1e5aSBadari Pulavarty /* short read or no more work */ 302e63e1e5aSBadari Pulavarty if ((ret != nr) || (len == 0)) 303e63e1e5aSBadari Pulavarty break; 304e63e1e5aSBadari Pulavarty } 305e63e1e5aSBadari Pulavarty out: 306a5516438SAndi Kleen *ppos = ((loff_t)index << huge_page_shift(h)) + offset; 307e63e1e5aSBadari Pulavarty mutex_unlock(&inode->i_mutex); 308e63e1e5aSBadari Pulavarty return retval; 309e63e1e5aSBadari Pulavarty } 310e63e1e5aSBadari Pulavarty 3111da177e4SLinus Torvalds /* 3121da177e4SLinus Torvalds * Read a page. Again trivial. If it didn't already exist 3131da177e4SLinus Torvalds * in the page cache, it is zero-filled. 3141da177e4SLinus Torvalds */ 3151da177e4SLinus Torvalds static int hugetlbfs_readpage(struct file *file, struct page * page) 3161da177e4SLinus Torvalds { 3171da177e4SLinus Torvalds unlock_page(page); 3181da177e4SLinus Torvalds return -EINVAL; 3191da177e4SLinus Torvalds } 3201da177e4SLinus Torvalds 321800d15a5SNick Piggin static int hugetlbfs_write_begin(struct file *file, 322800d15a5SNick Piggin struct address_space *mapping, 323800d15a5SNick Piggin loff_t pos, unsigned len, unsigned flags, 324800d15a5SNick Piggin struct page **pagep, void **fsdata) 3251da177e4SLinus Torvalds { 3261da177e4SLinus Torvalds return -EINVAL; 3271da177e4SLinus Torvalds } 3281da177e4SLinus Torvalds 329800d15a5SNick Piggin static int hugetlbfs_write_end(struct file *file, struct address_space *mapping, 330800d15a5SNick Piggin loff_t pos, unsigned len, unsigned copied, 331800d15a5SNick Piggin struct page *page, void *fsdata) 3321da177e4SLinus Torvalds { 333800d15a5SNick Piggin BUG(); 3341da177e4SLinus Torvalds return -EINVAL; 3351da177e4SLinus Torvalds } 3361da177e4SLinus Torvalds 3371da177e4SLinus Torvalds static void truncate_huge_page(struct page *page) 3381da177e4SLinus Torvalds { 339fba2591bSLinus Torvalds cancel_dirty_page(page, /* No IO accounting for huge pages? */0); 3401da177e4SLinus Torvalds ClearPageUptodate(page); 3411da177e4SLinus Torvalds remove_from_page_cache(page); 3421da177e4SLinus Torvalds put_page(page); 3431da177e4SLinus Torvalds } 3441da177e4SLinus Torvalds 345b45b5bd6SDavid Gibson static void truncate_hugepages(struct inode *inode, loff_t lstart) 3461da177e4SLinus Torvalds { 347a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 348b45b5bd6SDavid Gibson struct address_space *mapping = &inode->i_data; 349a5516438SAndi Kleen const pgoff_t start = lstart >> huge_page_shift(h); 3501da177e4SLinus Torvalds struct pagevec pvec; 3511da177e4SLinus Torvalds pgoff_t next; 352a43a8c39SChen, Kenneth W int i, freed = 0; 3531da177e4SLinus Torvalds 3541da177e4SLinus Torvalds pagevec_init(&pvec, 0); 3551da177e4SLinus Torvalds next = start; 3561da177e4SLinus Torvalds while (1) { 3571da177e4SLinus Torvalds if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { 3581da177e4SLinus Torvalds if (next == start) 3591da177e4SLinus Torvalds break; 3601da177e4SLinus Torvalds next = start; 3611da177e4SLinus Torvalds continue; 3621da177e4SLinus Torvalds } 3631da177e4SLinus Torvalds 3641da177e4SLinus Torvalds for (i = 0; i < pagevec_count(&pvec); ++i) { 3651da177e4SLinus Torvalds struct page *page = pvec.pages[i]; 3661da177e4SLinus Torvalds 3671da177e4SLinus Torvalds lock_page(page); 3681da177e4SLinus Torvalds if (page->index > next) 3691da177e4SLinus Torvalds next = page->index; 3701da177e4SLinus Torvalds ++next; 3711da177e4SLinus Torvalds truncate_huge_page(page); 3721da177e4SLinus Torvalds unlock_page(page); 373a43a8c39SChen, Kenneth W freed++; 3741da177e4SLinus Torvalds } 3751da177e4SLinus Torvalds huge_pagevec_release(&pvec); 3761da177e4SLinus Torvalds } 3771da177e4SLinus Torvalds BUG_ON(!lstart && mapping->nrpages); 378a43a8c39SChen, Kenneth W hugetlb_unreserve_pages(inode, start, freed); 3791da177e4SLinus Torvalds } 3801da177e4SLinus Torvalds 3811da177e4SLinus Torvalds static void hugetlbfs_delete_inode(struct inode *inode) 3821da177e4SLinus Torvalds { 383b45b5bd6SDavid Gibson truncate_hugepages(inode, 0); 384149f4211SChristoph Hellwig clear_inode(inode); 385149f4211SChristoph Hellwig } 386149f4211SChristoph Hellwig 387ddc0a51dSJosh Triplett static void hugetlbfs_forget_inode(struct inode *inode) __releases(inode_lock) 3881da177e4SLinus Torvalds { 3890b1533f6SChristoph Hellwig struct super_block *sb = inode->i_sb; 3901da177e4SLinus Torvalds 3910b1533f6SChristoph Hellwig if (!hlist_unhashed(&inode->i_hash)) { 3921c0eeaf5SJoern Engel if (!(inode->i_state & (I_DIRTY|I_SYNC))) 3930b1533f6SChristoph Hellwig list_move(&inode->i_list, &inode_unused); 3941da177e4SLinus Torvalds inodes_stat.nr_unused++; 3950b1533f6SChristoph Hellwig if (!sb || (sb->s_flags & MS_ACTIVE)) { 3961da177e4SLinus Torvalds spin_unlock(&inode_lock); 3971da177e4SLinus Torvalds return; 3981da177e4SLinus Torvalds } 3990b1533f6SChristoph Hellwig inode->i_state |= I_WILL_FREE; 4000b1533f6SChristoph Hellwig spin_unlock(&inode_lock); 4010b1533f6SChristoph Hellwig /* 4020b1533f6SChristoph Hellwig * write_inode_now is a noop as we set BDI_CAP_NO_WRITEBACK 4030b1533f6SChristoph Hellwig * in our backing_dev_info. 4040b1533f6SChristoph Hellwig */ 4050b1533f6SChristoph Hellwig write_inode_now(inode, 1); 4060b1533f6SChristoph Hellwig spin_lock(&inode_lock); 4070b1533f6SChristoph Hellwig inode->i_state &= ~I_WILL_FREE; 4081da177e4SLinus Torvalds inodes_stat.nr_unused--; 4091da177e4SLinus Torvalds hlist_del_init(&inode->i_hash); 4100b1533f6SChristoph Hellwig } 4111da177e4SLinus Torvalds list_del_init(&inode->i_list); 4121da177e4SLinus Torvalds list_del_init(&inode->i_sb_list); 4131da177e4SLinus Torvalds inode->i_state |= I_FREEING; 4141da177e4SLinus Torvalds inodes_stat.nr_inodes--; 4151da177e4SLinus Torvalds spin_unlock(&inode_lock); 416b45b5bd6SDavid Gibson truncate_hugepages(inode, 0); 4171da177e4SLinus Torvalds clear_inode(inode); 4181da177e4SLinus Torvalds destroy_inode(inode); 4191da177e4SLinus Torvalds } 4201da177e4SLinus Torvalds 4211da177e4SLinus Torvalds static void hugetlbfs_drop_inode(struct inode *inode) 4221da177e4SLinus Torvalds { 4231da177e4SLinus Torvalds if (!inode->i_nlink) 4246b09b9dfSChristoph Hellwig generic_delete_inode(inode); 4251da177e4SLinus Torvalds else 4261da177e4SLinus Torvalds hugetlbfs_forget_inode(inode); 4271da177e4SLinus Torvalds } 4281da177e4SLinus Torvalds 4291da177e4SLinus Torvalds static inline void 430856fc295SHugh Dickins hugetlb_vmtruncate_list(struct prio_tree_root *root, pgoff_t pgoff) 4311da177e4SLinus Torvalds { 4321da177e4SLinus Torvalds struct vm_area_struct *vma; 4331da177e4SLinus Torvalds struct prio_tree_iter iter; 4341da177e4SLinus Torvalds 435856fc295SHugh Dickins vma_prio_tree_foreach(vma, &iter, root, pgoff, ULONG_MAX) { 4361da177e4SLinus Torvalds unsigned long v_offset; 4371da177e4SLinus Torvalds 4381da177e4SLinus Torvalds /* 439856fc295SHugh Dickins * Can the expression below overflow on 32-bit arches? 440856fc295SHugh Dickins * No, because the prio_tree returns us only those vmas 441856fc295SHugh Dickins * which overlap the truncated area starting at pgoff, 442856fc295SHugh Dickins * and no vma on a 32-bit arch can span beyond the 4GB. 4431da177e4SLinus Torvalds */ 444856fc295SHugh Dickins if (vma->vm_pgoff < pgoff) 445856fc295SHugh Dickins v_offset = (pgoff - vma->vm_pgoff) << PAGE_SHIFT; 446856fc295SHugh Dickins else 4471da177e4SLinus Torvalds v_offset = 0; 4481da177e4SLinus Torvalds 449502717f4SChen, Kenneth W __unmap_hugepage_range(vma, 45004f2cbe3SMel Gorman vma->vm_start + v_offset, vma->vm_end, NULL); 4511da177e4SLinus Torvalds } 4521da177e4SLinus Torvalds } 4531da177e4SLinus Torvalds 4541da177e4SLinus Torvalds static int hugetlb_vmtruncate(struct inode *inode, loff_t offset) 4551da177e4SLinus Torvalds { 456856fc295SHugh Dickins pgoff_t pgoff; 4571da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 458a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 4591da177e4SLinus Torvalds 460a5516438SAndi Kleen BUG_ON(offset & ~huge_page_mask(h)); 461856fc295SHugh Dickins pgoff = offset >> PAGE_SHIFT; 4621da177e4SLinus Torvalds 4637aa91e10SKen Chen i_size_write(inode, offset); 4641da177e4SLinus Torvalds spin_lock(&mapping->i_mmap_lock); 4651da177e4SLinus Torvalds if (!prio_tree_empty(&mapping->i_mmap)) 4661da177e4SLinus Torvalds hugetlb_vmtruncate_list(&mapping->i_mmap, pgoff); 4671da177e4SLinus Torvalds spin_unlock(&mapping->i_mmap_lock); 468b45b5bd6SDavid Gibson truncate_hugepages(inode, offset); 4691da177e4SLinus Torvalds return 0; 4701da177e4SLinus Torvalds } 4711da177e4SLinus Torvalds 4721da177e4SLinus Torvalds static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr) 4731da177e4SLinus Torvalds { 4741da177e4SLinus Torvalds struct inode *inode = dentry->d_inode; 475a5516438SAndi Kleen struct hstate *h = hstate_inode(inode); 4761da177e4SLinus Torvalds int error; 4771da177e4SLinus Torvalds unsigned int ia_valid = attr->ia_valid; 4781da177e4SLinus Torvalds 4791da177e4SLinus Torvalds BUG_ON(!inode); 4801da177e4SLinus Torvalds 4811da177e4SLinus Torvalds error = inode_change_ok(inode, attr); 4821da177e4SLinus Torvalds if (error) 4831da177e4SLinus Torvalds goto out; 4841da177e4SLinus Torvalds 4851da177e4SLinus Torvalds if (ia_valid & ATTR_SIZE) { 4861da177e4SLinus Torvalds error = -EINVAL; 487a5516438SAndi Kleen if (!(attr->ia_size & ~huge_page_mask(h))) 4881da177e4SLinus Torvalds error = hugetlb_vmtruncate(inode, attr->ia_size); 4891da177e4SLinus Torvalds if (error) 4901da177e4SLinus Torvalds goto out; 4911da177e4SLinus Torvalds attr->ia_valid &= ~ATTR_SIZE; 4921da177e4SLinus Torvalds } 4931da177e4SLinus Torvalds error = inode_setattr(inode, attr); 4941da177e4SLinus Torvalds out: 4951da177e4SLinus Torvalds return error; 4961da177e4SLinus Torvalds } 4971da177e4SLinus Torvalds 4981da177e4SLinus Torvalds static struct inode *hugetlbfs_get_inode(struct super_block *sb, uid_t uid, 4991da177e4SLinus Torvalds gid_t gid, int mode, dev_t dev) 5001da177e4SLinus Torvalds { 5011da177e4SLinus Torvalds struct inode *inode; 5021da177e4SLinus Torvalds 5031da177e4SLinus Torvalds inode = new_inode(sb); 5041da177e4SLinus Torvalds if (inode) { 5051da177e4SLinus Torvalds struct hugetlbfs_inode_info *info; 5061da177e4SLinus Torvalds inode->i_mode = mode; 5071da177e4SLinus Torvalds inode->i_uid = uid; 5081da177e4SLinus Torvalds inode->i_gid = gid; 5091da177e4SLinus Torvalds inode->i_blocks = 0; 5101da177e4SLinus Torvalds inode->i_mapping->a_ops = &hugetlbfs_aops; 5111da177e4SLinus Torvalds inode->i_mapping->backing_dev_info =&hugetlbfs_backing_dev_info; 5121da177e4SLinus Torvalds inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME; 513a43a8c39SChen, Kenneth W INIT_LIST_HEAD(&inode->i_mapping->private_list); 5141da177e4SLinus Torvalds info = HUGETLBFS_I(inode); 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 } 5351da177e4SLinus Torvalds } 5361da177e4SLinus Torvalds return inode; 5371da177e4SLinus Torvalds } 5381da177e4SLinus Torvalds 5391da177e4SLinus Torvalds /* 5401da177e4SLinus Torvalds * File creation. Allocate an inode, and we're done.. 5411da177e4SLinus Torvalds */ 5421da177e4SLinus Torvalds static int hugetlbfs_mknod(struct inode *dir, 5431da177e4SLinus Torvalds struct dentry *dentry, int mode, dev_t dev) 5441da177e4SLinus Torvalds { 5451da177e4SLinus Torvalds struct inode *inode; 5461da177e4SLinus Torvalds int error = -ENOSPC; 5471da177e4SLinus Torvalds gid_t gid; 5481da177e4SLinus Torvalds 5491da177e4SLinus Torvalds if (dir->i_mode & S_ISGID) { 5501da177e4SLinus Torvalds gid = dir->i_gid; 5511da177e4SLinus Torvalds if (S_ISDIR(mode)) 5521da177e4SLinus Torvalds mode |= S_ISGID; 5531da177e4SLinus Torvalds } else { 5541da177e4SLinus Torvalds gid = current->fsgid; 5551da177e4SLinus Torvalds } 5561da177e4SLinus Torvalds inode = hugetlbfs_get_inode(dir->i_sb, current->fsuid, gid, mode, dev); 5571da177e4SLinus Torvalds if (inode) { 5581da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 5591da177e4SLinus Torvalds d_instantiate(dentry, inode); 5601da177e4SLinus Torvalds dget(dentry); /* Extra count - pin the dentry in core */ 5611da177e4SLinus Torvalds error = 0; 5621da177e4SLinus Torvalds } 5631da177e4SLinus Torvalds return error; 5641da177e4SLinus Torvalds } 5651da177e4SLinus Torvalds 5661da177e4SLinus Torvalds static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, int mode) 5671da177e4SLinus Torvalds { 5681da177e4SLinus Torvalds int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0); 5691da177e4SLinus Torvalds if (!retval) 570d8c76e6fSDave Hansen inc_nlink(dir); 5711da177e4SLinus Torvalds return retval; 5721da177e4SLinus Torvalds } 5731da177e4SLinus Torvalds 5741da177e4SLinus Torvalds static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, int mode, struct nameidata *nd) 5751da177e4SLinus Torvalds { 5761da177e4SLinus Torvalds return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0); 5771da177e4SLinus Torvalds } 5781da177e4SLinus Torvalds 5791da177e4SLinus Torvalds static int hugetlbfs_symlink(struct inode *dir, 5801da177e4SLinus Torvalds struct dentry *dentry, const char *symname) 5811da177e4SLinus Torvalds { 5821da177e4SLinus Torvalds struct inode *inode; 5831da177e4SLinus Torvalds int error = -ENOSPC; 5841da177e4SLinus Torvalds gid_t gid; 5851da177e4SLinus Torvalds 5861da177e4SLinus Torvalds if (dir->i_mode & S_ISGID) 5871da177e4SLinus Torvalds gid = dir->i_gid; 5881da177e4SLinus Torvalds else 5891da177e4SLinus Torvalds gid = current->fsgid; 5901da177e4SLinus Torvalds 5911da177e4SLinus Torvalds inode = hugetlbfs_get_inode(dir->i_sb, current->fsuid, 5921da177e4SLinus Torvalds gid, S_IFLNK|S_IRWXUGO, 0); 5931da177e4SLinus Torvalds if (inode) { 5941da177e4SLinus Torvalds int l = strlen(symname)+1; 5951da177e4SLinus Torvalds error = page_symlink(inode, symname, l); 5961da177e4SLinus Torvalds if (!error) { 5971da177e4SLinus Torvalds d_instantiate(dentry, inode); 5981da177e4SLinus Torvalds dget(dentry); 5991da177e4SLinus Torvalds } else 6001da177e4SLinus Torvalds iput(inode); 6011da177e4SLinus Torvalds } 6021da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 6031da177e4SLinus Torvalds 6041da177e4SLinus Torvalds return error; 6051da177e4SLinus Torvalds } 6061da177e4SLinus Torvalds 6071da177e4SLinus Torvalds /* 6086649a386SKen Chen * mark the head page dirty 6091da177e4SLinus Torvalds */ 6101da177e4SLinus Torvalds static int hugetlbfs_set_page_dirty(struct page *page) 6111da177e4SLinus Torvalds { 612d85f3385SChristoph Lameter struct page *head = compound_head(page); 6136649a386SKen Chen 6146649a386SKen Chen SetPageDirty(head); 6151da177e4SLinus Torvalds return 0; 6161da177e4SLinus Torvalds } 6171da177e4SLinus Torvalds 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() */ 62974a8a65cSDavid Gibson if (sbinfo->max_blocks >= 0) { 6301da177e4SLinus Torvalds buf->f_blocks = sbinfo->max_blocks; 6311da177e4SLinus Torvalds buf->f_bavail = buf->f_bfree = sbinfo->free_blocks; 6321da177e4SLinus Torvalds buf->f_files = sbinfo->max_inodes; 6331da177e4SLinus Torvalds buf->f_ffree = sbinfo->free_inodes; 63474a8a65cSDavid Gibson } 6351da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 6361da177e4SLinus Torvalds } 6371da177e4SLinus Torvalds buf->f_namelen = NAME_MAX; 6381da177e4SLinus Torvalds return 0; 6391da177e4SLinus Torvalds } 6401da177e4SLinus Torvalds 6411da177e4SLinus Torvalds static void hugetlbfs_put_super(struct super_block *sb) 6421da177e4SLinus Torvalds { 6431da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb); 6441da177e4SLinus Torvalds 6451da177e4SLinus Torvalds if (sbi) { 6461da177e4SLinus Torvalds sb->s_fs_info = NULL; 6471da177e4SLinus Torvalds kfree(sbi); 6481da177e4SLinus Torvalds } 6491da177e4SLinus Torvalds } 6501da177e4SLinus Torvalds 65196527980SChristoph Hellwig static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo) 65296527980SChristoph Hellwig { 65396527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 65496527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 65596527980SChristoph Hellwig if (unlikely(!sbinfo->free_inodes)) { 65696527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 65796527980SChristoph Hellwig return 0; 65896527980SChristoph Hellwig } 65996527980SChristoph Hellwig sbinfo->free_inodes--; 66096527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 66196527980SChristoph Hellwig } 66296527980SChristoph Hellwig 66396527980SChristoph Hellwig return 1; 66496527980SChristoph Hellwig } 66596527980SChristoph Hellwig 66696527980SChristoph Hellwig static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo) 66796527980SChristoph Hellwig { 66896527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 66996527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 67096527980SChristoph Hellwig sbinfo->free_inodes++; 67196527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 67296527980SChristoph Hellwig } 67396527980SChristoph Hellwig } 67496527980SChristoph Hellwig 67596527980SChristoph Hellwig 676e18b890bSChristoph Lameter static struct kmem_cache *hugetlbfs_inode_cachep; 6771da177e4SLinus Torvalds 6781da177e4SLinus Torvalds static struct inode *hugetlbfs_alloc_inode(struct super_block *sb) 6791da177e4SLinus Torvalds { 68096527980SChristoph Hellwig struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb); 6811da177e4SLinus Torvalds struct hugetlbfs_inode_info *p; 6821da177e4SLinus Torvalds 68396527980SChristoph Hellwig if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo))) 6841da177e4SLinus Torvalds return NULL; 685e94b1766SChristoph Lameter p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL); 68696527980SChristoph Hellwig if (unlikely(!p)) { 68796527980SChristoph Hellwig hugetlbfs_inc_free_inodes(sbinfo); 68896527980SChristoph Hellwig return NULL; 6891da177e4SLinus Torvalds } 69096527980SChristoph Hellwig return &p->vfs_inode; 6911da177e4SLinus Torvalds } 6921da177e4SLinus Torvalds 6931da177e4SLinus Torvalds static void hugetlbfs_destroy_inode(struct inode *inode) 6941da177e4SLinus Torvalds { 69596527980SChristoph Hellwig hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb)); 6961da177e4SLinus Torvalds mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy); 6971da177e4SLinus Torvalds kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode)); 6981da177e4SLinus Torvalds } 6991da177e4SLinus Torvalds 700f5e54d6eSChristoph Hellwig static const struct address_space_operations hugetlbfs_aops = { 7011da177e4SLinus Torvalds .readpage = hugetlbfs_readpage, 702800d15a5SNick Piggin .write_begin = hugetlbfs_write_begin, 703800d15a5SNick Piggin .write_end = hugetlbfs_write_end, 7041da177e4SLinus Torvalds .set_page_dirty = hugetlbfs_set_page_dirty, 7051da177e4SLinus Torvalds }; 7061da177e4SLinus Torvalds 70796527980SChristoph Hellwig 70851cc5068SAlexey Dobriyan static void init_once(void *foo) 70996527980SChristoph Hellwig { 71096527980SChristoph Hellwig struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo; 71196527980SChristoph Hellwig 71296527980SChristoph Hellwig inode_init_once(&ei->vfs_inode); 71396527980SChristoph Hellwig } 71496527980SChristoph Hellwig 7154b6f5d20SArjan van de Ven const struct file_operations hugetlbfs_file_operations = { 716e63e1e5aSBadari Pulavarty .read = hugetlbfs_read, 7171da177e4SLinus Torvalds .mmap = hugetlbfs_file_mmap, 7181da177e4SLinus Torvalds .fsync = simple_sync_file, 7191da177e4SLinus Torvalds .get_unmapped_area = hugetlb_get_unmapped_area, 7201da177e4SLinus Torvalds }; 7211da177e4SLinus Torvalds 72292e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_dir_inode_operations = { 7231da177e4SLinus Torvalds .create = hugetlbfs_create, 7241da177e4SLinus Torvalds .lookup = simple_lookup, 7251da177e4SLinus Torvalds .link = simple_link, 7261da177e4SLinus Torvalds .unlink = simple_unlink, 7271da177e4SLinus Torvalds .symlink = hugetlbfs_symlink, 7281da177e4SLinus Torvalds .mkdir = hugetlbfs_mkdir, 7291da177e4SLinus Torvalds .rmdir = simple_rmdir, 7301da177e4SLinus Torvalds .mknod = hugetlbfs_mknod, 7311da177e4SLinus Torvalds .rename = simple_rename, 7321da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 7331da177e4SLinus Torvalds }; 7341da177e4SLinus Torvalds 73592e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_inode_operations = { 7361da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 7371da177e4SLinus Torvalds }; 7381da177e4SLinus Torvalds 739ee9b6d61SJosef 'Jeff' Sipek static const struct super_operations hugetlbfs_ops = { 7401da177e4SLinus Torvalds .alloc_inode = hugetlbfs_alloc_inode, 7411da177e4SLinus Torvalds .destroy_inode = hugetlbfs_destroy_inode, 7421da177e4SLinus Torvalds .statfs = hugetlbfs_statfs, 743149f4211SChristoph Hellwig .delete_inode = hugetlbfs_delete_inode, 7441da177e4SLinus Torvalds .drop_inode = hugetlbfs_drop_inode, 7451da177e4SLinus Torvalds .put_super = hugetlbfs_put_super, 74610f19a86SMiklos Szeredi .show_options = generic_show_options, 7471da177e4SLinus Torvalds }; 7481da177e4SLinus Torvalds 7491da177e4SLinus Torvalds static int 7501da177e4SLinus Torvalds hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig) 7511da177e4SLinus Torvalds { 752e73a75faSRandy Dunlap char *p, *rest; 753e73a75faSRandy Dunlap substring_t args[MAX_OPT_ARGS]; 754e73a75faSRandy Dunlap int option; 755a137e1ccSAndi Kleen unsigned long long size = 0; 756a137e1ccSAndi Kleen enum { NO_SIZE, SIZE_STD, SIZE_PERCENT } setsize = NO_SIZE; 7571da177e4SLinus Torvalds 7581da177e4SLinus Torvalds if (!options) 7591da177e4SLinus Torvalds return 0; 7601da177e4SLinus Torvalds 761e73a75faSRandy Dunlap while ((p = strsep(&options, ",")) != NULL) { 762e73a75faSRandy Dunlap int token; 763b4c07bceSLee Schermerhorn if (!*p) 764b4c07bceSLee Schermerhorn continue; 7651da177e4SLinus Torvalds 766e73a75faSRandy Dunlap token = match_token(p, tokens, args); 767e73a75faSRandy Dunlap switch (token) { 768e73a75faSRandy Dunlap case Opt_uid: 769e73a75faSRandy Dunlap if (match_int(&args[0], &option)) 770e73a75faSRandy Dunlap goto bad_val; 771e73a75faSRandy Dunlap pconfig->uid = option; 772e73a75faSRandy Dunlap break; 773e73a75faSRandy Dunlap 774e73a75faSRandy Dunlap case Opt_gid: 775e73a75faSRandy Dunlap if (match_int(&args[0], &option)) 776e73a75faSRandy Dunlap goto bad_val; 777e73a75faSRandy Dunlap pconfig->gid = option; 778e73a75faSRandy Dunlap break; 779e73a75faSRandy Dunlap 780e73a75faSRandy Dunlap case Opt_mode: 781e73a75faSRandy Dunlap if (match_octal(&args[0], &option)) 782e73a75faSRandy Dunlap goto bad_val; 78375897d60SKen Chen pconfig->mode = option & 01777U; 784e73a75faSRandy Dunlap break; 785e73a75faSRandy Dunlap 786e73a75faSRandy Dunlap case Opt_size: { 787e73a75faSRandy Dunlap /* memparse() will accept a K/M/G without a digit */ 788e73a75faSRandy Dunlap if (!isdigit(*args[0].from)) 789e73a75faSRandy Dunlap goto bad_val; 790e73a75faSRandy Dunlap size = memparse(args[0].from, &rest); 791a137e1ccSAndi Kleen setsize = SIZE_STD; 792a137e1ccSAndi Kleen if (*rest == '%') 793a137e1ccSAndi Kleen setsize = SIZE_PERCENT; 794e73a75faSRandy Dunlap break; 795e73a75faSRandy Dunlap } 7961da177e4SLinus Torvalds 797e73a75faSRandy Dunlap case Opt_nr_inodes: 798e73a75faSRandy Dunlap /* memparse() will accept a K/M/G without a digit */ 799e73a75faSRandy Dunlap if (!isdigit(*args[0].from)) 800e73a75faSRandy Dunlap goto bad_val; 801e73a75faSRandy Dunlap pconfig->nr_inodes = memparse(args[0].from, &rest); 802e73a75faSRandy Dunlap break; 803e73a75faSRandy Dunlap 804a137e1ccSAndi Kleen case Opt_pagesize: { 805a137e1ccSAndi Kleen unsigned long ps; 806a137e1ccSAndi Kleen ps = memparse(args[0].from, &rest); 807a137e1ccSAndi Kleen pconfig->hstate = size_to_hstate(ps); 808a137e1ccSAndi Kleen if (!pconfig->hstate) { 809a137e1ccSAndi Kleen printk(KERN_ERR 810a137e1ccSAndi Kleen "hugetlbfs: Unsupported page size %lu MB\n", 811a137e1ccSAndi Kleen ps >> 20); 812a137e1ccSAndi Kleen return -EINVAL; 813a137e1ccSAndi Kleen } 814a137e1ccSAndi Kleen break; 815a137e1ccSAndi Kleen } 816a137e1ccSAndi Kleen 817e73a75faSRandy Dunlap default: 818b4c07bceSLee Schermerhorn printk(KERN_ERR "hugetlbfs: Bad mount option: \"%s\"\n", 819b4c07bceSLee Schermerhorn p); 820b4c07bceSLee Schermerhorn return -EINVAL; 821e73a75faSRandy Dunlap break; 822e73a75faSRandy Dunlap } 8231da177e4SLinus Torvalds } 824a137e1ccSAndi Kleen 825a137e1ccSAndi Kleen /* Do size after hstate is set up */ 826a137e1ccSAndi Kleen if (setsize > NO_SIZE) { 827a137e1ccSAndi Kleen struct hstate *h = pconfig->hstate; 828a137e1ccSAndi Kleen if (setsize == SIZE_PERCENT) { 829a137e1ccSAndi Kleen size <<= huge_page_shift(h); 830a137e1ccSAndi Kleen size *= h->max_huge_pages; 831a137e1ccSAndi Kleen do_div(size, 100); 832a137e1ccSAndi Kleen } 833a137e1ccSAndi Kleen pconfig->nr_blocks = (size >> huge_page_shift(h)); 834a137e1ccSAndi Kleen } 835a137e1ccSAndi Kleen 8361da177e4SLinus Torvalds return 0; 837e73a75faSRandy Dunlap 838e73a75faSRandy Dunlap bad_val: 839e73a75faSRandy Dunlap printk(KERN_ERR "hugetlbfs: Bad value '%s' for mount option '%s'\n", 840e73a75faSRandy Dunlap args[0].from, p); 841e73a75faSRandy Dunlap return 1; 8421da177e4SLinus Torvalds } 8431da177e4SLinus Torvalds 8441da177e4SLinus Torvalds static int 8451da177e4SLinus Torvalds hugetlbfs_fill_super(struct super_block *sb, void *data, int silent) 8461da177e4SLinus Torvalds { 8471da177e4SLinus Torvalds struct inode * inode; 8481da177e4SLinus Torvalds struct dentry * root; 8491da177e4SLinus Torvalds int ret; 8501da177e4SLinus Torvalds struct hugetlbfs_config config; 8511da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo; 8521da177e4SLinus Torvalds 85310f19a86SMiklos Szeredi save_mount_options(sb, data); 85410f19a86SMiklos Szeredi 8551da177e4SLinus Torvalds config.nr_blocks = -1; /* No limit on size by default */ 8561da177e4SLinus Torvalds config.nr_inodes = -1; /* No limit on number of inodes by default */ 8571da177e4SLinus Torvalds config.uid = current->fsuid; 8581da177e4SLinus Torvalds config.gid = current->fsgid; 8591da177e4SLinus Torvalds config.mode = 0755; 860a137e1ccSAndi Kleen config.hstate = &default_hstate; 8611da177e4SLinus Torvalds ret = hugetlbfs_parse_options(data, &config); 8621da177e4SLinus Torvalds if (ret) 8631da177e4SLinus Torvalds return ret; 8641da177e4SLinus Torvalds 8651da177e4SLinus Torvalds sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL); 8661da177e4SLinus Torvalds if (!sbinfo) 8671da177e4SLinus Torvalds return -ENOMEM; 8681da177e4SLinus Torvalds sb->s_fs_info = sbinfo; 869a137e1ccSAndi Kleen sbinfo->hstate = config.hstate; 8701da177e4SLinus Torvalds spin_lock_init(&sbinfo->stat_lock); 8711da177e4SLinus Torvalds sbinfo->max_blocks = config.nr_blocks; 8721da177e4SLinus Torvalds sbinfo->free_blocks = config.nr_blocks; 8731da177e4SLinus Torvalds sbinfo->max_inodes = config.nr_inodes; 8741da177e4SLinus Torvalds sbinfo->free_inodes = config.nr_inodes; 8751da177e4SLinus Torvalds sb->s_maxbytes = MAX_LFS_FILESIZE; 876a137e1ccSAndi Kleen sb->s_blocksize = huge_page_size(config.hstate); 877a137e1ccSAndi Kleen sb->s_blocksize_bits = huge_page_shift(config.hstate); 8781da177e4SLinus Torvalds sb->s_magic = HUGETLBFS_MAGIC; 8791da177e4SLinus Torvalds sb->s_op = &hugetlbfs_ops; 8801da177e4SLinus Torvalds sb->s_time_gran = 1; 8811da177e4SLinus Torvalds inode = hugetlbfs_get_inode(sb, config.uid, config.gid, 8821da177e4SLinus Torvalds S_IFDIR | config.mode, 0); 8831da177e4SLinus Torvalds if (!inode) 8841da177e4SLinus Torvalds goto out_free; 8851da177e4SLinus Torvalds 8861da177e4SLinus Torvalds root = d_alloc_root(inode); 8871da177e4SLinus Torvalds if (!root) { 8881da177e4SLinus Torvalds iput(inode); 8891da177e4SLinus Torvalds goto out_free; 8901da177e4SLinus Torvalds } 8911da177e4SLinus Torvalds sb->s_root = root; 8921da177e4SLinus Torvalds return 0; 8931da177e4SLinus Torvalds out_free: 8941da177e4SLinus Torvalds kfree(sbinfo); 8951da177e4SLinus Torvalds return -ENOMEM; 8961da177e4SLinus Torvalds } 8971da177e4SLinus Torvalds 8989a119c05SAdam Litke int hugetlb_get_quota(struct address_space *mapping, long delta) 8991da177e4SLinus Torvalds { 9001da177e4SLinus Torvalds int ret = 0; 9011da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(mapping->host->i_sb); 9021da177e4SLinus Torvalds 9031da177e4SLinus Torvalds if (sbinfo->free_blocks > -1) { 9041da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 9059a119c05SAdam Litke if (sbinfo->free_blocks - delta >= 0) 9069a119c05SAdam Litke sbinfo->free_blocks -= delta; 9071da177e4SLinus Torvalds else 9081da177e4SLinus Torvalds ret = -ENOMEM; 9091da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 9101da177e4SLinus Torvalds } 9111da177e4SLinus Torvalds 9121da177e4SLinus Torvalds return ret; 9131da177e4SLinus Torvalds } 9141da177e4SLinus Torvalds 9159a119c05SAdam Litke void hugetlb_put_quota(struct address_space *mapping, long delta) 9161da177e4SLinus Torvalds { 9171da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(mapping->host->i_sb); 9181da177e4SLinus Torvalds 9191da177e4SLinus Torvalds if (sbinfo->free_blocks > -1) { 9201da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 9219a119c05SAdam Litke sbinfo->free_blocks += delta; 9221da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 9231da177e4SLinus Torvalds } 9241da177e4SLinus Torvalds } 9251da177e4SLinus Torvalds 926454e2398SDavid Howells static int hugetlbfs_get_sb(struct file_system_type *fs_type, 927454e2398SDavid Howells int flags, const char *dev_name, void *data, struct vfsmount *mnt) 9281da177e4SLinus Torvalds { 929454e2398SDavid Howells return get_sb_nodev(fs_type, flags, data, hugetlbfs_fill_super, mnt); 9301da177e4SLinus Torvalds } 9311da177e4SLinus Torvalds 9321da177e4SLinus Torvalds static struct file_system_type hugetlbfs_fs_type = { 9331da177e4SLinus Torvalds .name = "hugetlbfs", 9341da177e4SLinus Torvalds .get_sb = hugetlbfs_get_sb, 9351da177e4SLinus Torvalds .kill_sb = kill_litter_super, 9361da177e4SLinus Torvalds }; 9371da177e4SLinus Torvalds 9381da177e4SLinus Torvalds static struct vfsmount *hugetlbfs_vfsmount; 9391da177e4SLinus Torvalds 9401da177e4SLinus Torvalds static int can_do_hugetlb_shm(void) 9411da177e4SLinus Torvalds { 9421da177e4SLinus Torvalds return likely(capable(CAP_IPC_LOCK) || 9431da177e4SLinus Torvalds in_group_p(sysctl_hugetlb_shm_group) || 9441da177e4SLinus Torvalds can_do_mlock()); 9451da177e4SLinus Torvalds } 9461da177e4SLinus Torvalds 9479d66586fSEric W. Biederman struct file *hugetlb_file_setup(const char *name, size_t size) 9481da177e4SLinus Torvalds { 9491da177e4SLinus Torvalds int error = -ENOMEM; 9501da177e4SLinus Torvalds struct file *file; 9511da177e4SLinus Torvalds struct inode *inode; 9521da177e4SLinus Torvalds struct dentry *dentry, *root; 9531da177e4SLinus Torvalds struct qstr quick_string; 9541da177e4SLinus Torvalds 9555bc98594SAkinobu Mita if (!hugetlbfs_vfsmount) 9565bc98594SAkinobu Mita return ERR_PTR(-ENOENT); 9575bc98594SAkinobu Mita 9581da177e4SLinus Torvalds if (!can_do_hugetlb_shm()) 9591da177e4SLinus Torvalds return ERR_PTR(-EPERM); 9601da177e4SLinus Torvalds 9611da177e4SLinus Torvalds if (!user_shm_lock(size, current->user)) 9621da177e4SLinus Torvalds return ERR_PTR(-ENOMEM); 9631da177e4SLinus Torvalds 9641da177e4SLinus Torvalds root = hugetlbfs_vfsmount->mnt_root; 9659d66586fSEric W. Biederman quick_string.name = name; 9661da177e4SLinus Torvalds quick_string.len = strlen(quick_string.name); 9671da177e4SLinus Torvalds quick_string.hash = 0; 9681da177e4SLinus Torvalds dentry = d_alloc(root, &quick_string); 9691da177e4SLinus Torvalds if (!dentry) 9701da177e4SLinus Torvalds goto out_shm_unlock; 9711da177e4SLinus Torvalds 9721da177e4SLinus Torvalds error = -ENOSPC; 9731da177e4SLinus Torvalds inode = hugetlbfs_get_inode(root->d_sb, current->fsuid, 9741da177e4SLinus Torvalds current->fsgid, S_IFREG | S_IRWXUGO, 0); 9751da177e4SLinus Torvalds if (!inode) 976ce8d2cdfSDave Hansen goto out_dentry; 9771da177e4SLinus Torvalds 978b45b5bd6SDavid Gibson error = -ENOMEM; 979a5516438SAndi Kleen if (hugetlb_reserve_pages(inode, 0, 980a5516438SAndi Kleen size >> huge_page_shift(hstate_inode(inode)), NULL)) 981b45b5bd6SDavid Gibson goto out_inode; 982b45b5bd6SDavid Gibson 9831da177e4SLinus Torvalds d_instantiate(dentry, inode); 9841da177e4SLinus Torvalds inode->i_size = size; 9851da177e4SLinus Torvalds inode->i_nlink = 0; 986ce8d2cdfSDave Hansen 987ce8d2cdfSDave Hansen error = -ENFILE; 988ce8d2cdfSDave Hansen file = alloc_file(hugetlbfs_vfsmount, dentry, 989ce8d2cdfSDave Hansen FMODE_WRITE | FMODE_READ, 990ce8d2cdfSDave Hansen &hugetlbfs_file_operations); 991ce8d2cdfSDave Hansen if (!file) 992b4d232e6SAl Viro goto out_dentry; /* inode is already attached */ 993ce8d2cdfSDave Hansen 9941da177e4SLinus Torvalds return file; 9951da177e4SLinus Torvalds 996b45b5bd6SDavid Gibson out_inode: 997b45b5bd6SDavid Gibson iput(inode); 9981da177e4SLinus Torvalds out_dentry: 9991da177e4SLinus Torvalds dput(dentry); 10001da177e4SLinus Torvalds out_shm_unlock: 10011da177e4SLinus Torvalds user_shm_unlock(size, current->user); 10021da177e4SLinus Torvalds return ERR_PTR(error); 10031da177e4SLinus Torvalds } 10041da177e4SLinus Torvalds 10051da177e4SLinus Torvalds static int __init init_hugetlbfs_fs(void) 10061da177e4SLinus Torvalds { 10071da177e4SLinus Torvalds int error; 10081da177e4SLinus Torvalds struct vfsmount *vfsmount; 10091da177e4SLinus Torvalds 1010e0bf68ddSPeter Zijlstra error = bdi_init(&hugetlbfs_backing_dev_info); 1011e0bf68ddSPeter Zijlstra if (error) 1012e0bf68ddSPeter Zijlstra return error; 1013e0bf68ddSPeter Zijlstra 10141da177e4SLinus Torvalds hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache", 10151da177e4SLinus Torvalds sizeof(struct hugetlbfs_inode_info), 101620c2df83SPaul Mundt 0, 0, init_once); 10171da177e4SLinus Torvalds if (hugetlbfs_inode_cachep == NULL) 1018e0bf68ddSPeter Zijlstra goto out2; 10191da177e4SLinus Torvalds 10201da177e4SLinus Torvalds error = register_filesystem(&hugetlbfs_fs_type); 10211da177e4SLinus Torvalds if (error) 10221da177e4SLinus Torvalds goto out; 10231da177e4SLinus Torvalds 10241da177e4SLinus Torvalds vfsmount = kern_mount(&hugetlbfs_fs_type); 10251da177e4SLinus Torvalds 10261da177e4SLinus Torvalds if (!IS_ERR(vfsmount)) { 10271da177e4SLinus Torvalds hugetlbfs_vfsmount = vfsmount; 10281da177e4SLinus Torvalds return 0; 10291da177e4SLinus Torvalds } 10301da177e4SLinus Torvalds 10311da177e4SLinus Torvalds error = PTR_ERR(vfsmount); 10321da177e4SLinus Torvalds 10331da177e4SLinus Torvalds out: 10341da177e4SLinus Torvalds if (error) 10351da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 1036e0bf68ddSPeter Zijlstra out2: 1037e0bf68ddSPeter Zijlstra bdi_destroy(&hugetlbfs_backing_dev_info); 10381da177e4SLinus Torvalds return error; 10391da177e4SLinus Torvalds } 10401da177e4SLinus Torvalds 10411da177e4SLinus Torvalds static void __exit exit_hugetlbfs_fs(void) 10421da177e4SLinus Torvalds { 10431da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 10441da177e4SLinus Torvalds unregister_filesystem(&hugetlbfs_fs_type); 1045e0bf68ddSPeter Zijlstra bdi_destroy(&hugetlbfs_backing_dev_info); 10461da177e4SLinus Torvalds } 10471da177e4SLinus Torvalds 10481da177e4SLinus Torvalds module_init(init_hugetlbfs_fs) 10491da177e4SLinus Torvalds module_exit(exit_hugetlbfs_fs) 10501da177e4SLinus Torvalds 10511da177e4SLinus Torvalds MODULE_LICENSE("GPL"); 1052