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> 161da177e4SLinus Torvalds #include <linux/writeback.h> 171da177e4SLinus Torvalds #include <linux/pagemap.h> 181da177e4SLinus Torvalds #include <linux/highmem.h> 191da177e4SLinus Torvalds #include <linux/init.h> 201da177e4SLinus Torvalds #include <linux/string.h> 2116f7e0feSRandy Dunlap #include <linux/capability.h> 221da177e4SLinus Torvalds #include <linux/backing-dev.h> 231da177e4SLinus Torvalds #include <linux/hugetlb.h> 241da177e4SLinus Torvalds #include <linux/pagevec.h> 251da177e4SLinus Torvalds #include <linux/quotaops.h> 261da177e4SLinus Torvalds #include <linux/slab.h> 271da177e4SLinus Torvalds #include <linux/dnotify.h> 281da177e4SLinus Torvalds #include <linux/statfs.h> 291da177e4SLinus Torvalds #include <linux/security.h> 301da177e4SLinus Torvalds 311da177e4SLinus Torvalds #include <asm/uaccess.h> 321da177e4SLinus Torvalds 331da177e4SLinus Torvalds /* some random number */ 341da177e4SLinus Torvalds #define HUGETLBFS_MAGIC 0x958458f6 351da177e4SLinus Torvalds 361da177e4SLinus Torvalds static struct super_operations hugetlbfs_ops; 37f5e54d6eSChristoph Hellwig static const struct address_space_operations hugetlbfs_aops; 384b6f5d20SArjan van de Ven const struct file_operations hugetlbfs_file_operations; 3992e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_dir_inode_operations; 4092e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_inode_operations; 411da177e4SLinus Torvalds 421da177e4SLinus Torvalds static struct backing_dev_info hugetlbfs_backing_dev_info = { 431da177e4SLinus Torvalds .ra_pages = 0, /* No readahead */ 441da177e4SLinus Torvalds .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK, 451da177e4SLinus Torvalds }; 461da177e4SLinus Torvalds 471da177e4SLinus Torvalds int sysctl_hugetlb_shm_group; 481da177e4SLinus Torvalds 492e9b367cSAdam Litke static void huge_pagevec_release(struct pagevec *pvec) 502e9b367cSAdam Litke { 512e9b367cSAdam Litke int i; 522e9b367cSAdam Litke 532e9b367cSAdam Litke for (i = 0; i < pagevec_count(pvec); ++i) 542e9b367cSAdam Litke put_page(pvec->pages[i]); 552e9b367cSAdam Litke 562e9b367cSAdam Litke pagevec_reinit(pvec); 572e9b367cSAdam Litke } 582e9b367cSAdam Litke 591da177e4SLinus Torvalds static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma) 601da177e4SLinus Torvalds { 61b39424e2SJosef Sipek struct inode *inode = file->f_path.dentry->d_inode; 621da177e4SLinus Torvalds loff_t len, vma_len; 631da177e4SLinus Torvalds int ret; 641da177e4SLinus Torvalds 6568589bc3SHugh Dickins /* 6668589bc3SHugh Dickins * vma alignment has already been checked by prepare_hugepage_range. 6768589bc3SHugh Dickins * If you add any error returns here, do so after setting VM_HUGETLB, 6868589bc3SHugh Dickins * so is_vm_hugetlb_page tests below unmap_region go the right way 6968589bc3SHugh Dickins * when do_mmap_pgoff unwinds (may be important on powerpc and ia64). 7068589bc3SHugh Dickins */ 7168589bc3SHugh Dickins vma->vm_flags |= VM_HUGETLB | VM_RESERVED; 7268589bc3SHugh Dickins vma->vm_ops = &hugetlb_vm_ops; 731da177e4SLinus Torvalds 741da177e4SLinus Torvalds vma_len = (loff_t)(vma->vm_end - vma->vm_start); 751da177e4SLinus Torvalds 761b1dcc1bSJes Sorensen mutex_lock(&inode->i_mutex); 771da177e4SLinus Torvalds file_accessed(file); 781da177e4SLinus Torvalds 791da177e4SLinus Torvalds ret = -ENOMEM; 801da177e4SLinus Torvalds len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); 811da177e4SLinus Torvalds 82a43a8c39SChen, Kenneth W if (vma->vm_flags & VM_MAYSHARE && 83a43a8c39SChen, Kenneth W hugetlb_reserve_pages(inode, vma->vm_pgoff >> (HPAGE_SHIFT-PAGE_SHIFT), 84a43a8c39SChen, Kenneth W len >> HPAGE_SHIFT)) 85b45b5bd6SDavid Gibson goto out; 86b45b5bd6SDavid Gibson 874c887265SAdam Litke ret = 0; 884c887265SAdam Litke hugetlb_prefault_arch_hook(vma->vm_mm); 89b6174df5SZhang, Yanmin if (vma->vm_flags & VM_WRITE && inode->i_size < len) 901da177e4SLinus Torvalds inode->i_size = len; 911da177e4SLinus Torvalds out: 921b1dcc1bSJes Sorensen mutex_unlock(&inode->i_mutex); 931da177e4SLinus Torvalds 941da177e4SLinus Torvalds return ret; 951da177e4SLinus Torvalds } 961da177e4SLinus Torvalds 971da177e4SLinus Torvalds /* 98508034a3SHugh Dickins * Called under down_write(mmap_sem). 991da177e4SLinus Torvalds */ 1001da177e4SLinus Torvalds 1011da177e4SLinus Torvalds #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA 1021da177e4SLinus Torvalds unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, 1031da177e4SLinus Torvalds unsigned long len, unsigned long pgoff, unsigned long flags); 1041da177e4SLinus Torvalds #else 1051da177e4SLinus Torvalds static unsigned long 1061da177e4SLinus Torvalds hugetlb_get_unmapped_area(struct file *file, unsigned long addr, 1071da177e4SLinus Torvalds unsigned long len, unsigned long pgoff, unsigned long flags) 1081da177e4SLinus Torvalds { 1091da177e4SLinus Torvalds struct mm_struct *mm = current->mm; 1101da177e4SLinus Torvalds struct vm_area_struct *vma; 1111da177e4SLinus Torvalds unsigned long start_addr; 1121da177e4SLinus Torvalds 1131da177e4SLinus Torvalds if (len & ~HPAGE_MASK) 1141da177e4SLinus Torvalds return -EINVAL; 1151da177e4SLinus Torvalds if (len > TASK_SIZE) 1161da177e4SLinus Torvalds return -ENOMEM; 1171da177e4SLinus Torvalds 1181da177e4SLinus Torvalds if (addr) { 1191da177e4SLinus Torvalds addr = ALIGN(addr, HPAGE_SIZE); 1201da177e4SLinus Torvalds vma = find_vma(mm, addr); 1211da177e4SLinus Torvalds if (TASK_SIZE - len >= addr && 1221da177e4SLinus Torvalds (!vma || addr + len <= vma->vm_start)) 1231da177e4SLinus Torvalds return addr; 1241da177e4SLinus Torvalds } 1251da177e4SLinus Torvalds 1261da177e4SLinus Torvalds start_addr = mm->free_area_cache; 1271da177e4SLinus Torvalds 1281363c3cdSWolfgang Wander if (len <= mm->cached_hole_size) 1291363c3cdSWolfgang Wander start_addr = TASK_UNMAPPED_BASE; 1301363c3cdSWolfgang Wander 1311da177e4SLinus Torvalds full_search: 1321da177e4SLinus Torvalds addr = ALIGN(start_addr, HPAGE_SIZE); 1331da177e4SLinus Torvalds 1341da177e4SLinus Torvalds for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1351da177e4SLinus Torvalds /* At this point: (!vma || addr < vma->vm_end). */ 1361da177e4SLinus Torvalds if (TASK_SIZE - len < addr) { 1371da177e4SLinus Torvalds /* 1381da177e4SLinus Torvalds * Start a new search - just in case we missed 1391da177e4SLinus Torvalds * some holes. 1401da177e4SLinus Torvalds */ 1411da177e4SLinus Torvalds if (start_addr != TASK_UNMAPPED_BASE) { 1421da177e4SLinus Torvalds start_addr = TASK_UNMAPPED_BASE; 1431da177e4SLinus Torvalds goto full_search; 1441da177e4SLinus Torvalds } 1451da177e4SLinus Torvalds return -ENOMEM; 1461da177e4SLinus Torvalds } 1471da177e4SLinus Torvalds 1481da177e4SLinus Torvalds if (!vma || addr + len <= vma->vm_start) 1491da177e4SLinus Torvalds return addr; 1501da177e4SLinus Torvalds addr = ALIGN(vma->vm_end, HPAGE_SIZE); 1511da177e4SLinus Torvalds } 1521da177e4SLinus Torvalds } 1531da177e4SLinus Torvalds #endif 1541da177e4SLinus Torvalds 1551da177e4SLinus Torvalds /* 1561da177e4SLinus Torvalds * Read a page. Again trivial. If it didn't already exist 1571da177e4SLinus Torvalds * in the page cache, it is zero-filled. 1581da177e4SLinus Torvalds */ 1591da177e4SLinus Torvalds static int hugetlbfs_readpage(struct file *file, struct page * page) 1601da177e4SLinus Torvalds { 1611da177e4SLinus Torvalds unlock_page(page); 1621da177e4SLinus Torvalds return -EINVAL; 1631da177e4SLinus Torvalds } 1641da177e4SLinus Torvalds 1651da177e4SLinus Torvalds static int hugetlbfs_prepare_write(struct file *file, 1661da177e4SLinus Torvalds struct page *page, unsigned offset, unsigned to) 1671da177e4SLinus Torvalds { 1681da177e4SLinus Torvalds return -EINVAL; 1691da177e4SLinus Torvalds } 1701da177e4SLinus Torvalds 1711da177e4SLinus Torvalds static int hugetlbfs_commit_write(struct file *file, 1721da177e4SLinus Torvalds struct page *page, unsigned offset, unsigned to) 1731da177e4SLinus Torvalds { 1741da177e4SLinus Torvalds return -EINVAL; 1751da177e4SLinus Torvalds } 1761da177e4SLinus Torvalds 1771da177e4SLinus Torvalds static void truncate_huge_page(struct page *page) 1781da177e4SLinus Torvalds { 179fba2591bSLinus Torvalds cancel_dirty_page(page, /* No IO accounting for huge pages? */0); 1801da177e4SLinus Torvalds ClearPageUptodate(page); 1811da177e4SLinus Torvalds remove_from_page_cache(page); 1821da177e4SLinus Torvalds put_page(page); 1831da177e4SLinus Torvalds } 1841da177e4SLinus Torvalds 185b45b5bd6SDavid Gibson static void truncate_hugepages(struct inode *inode, loff_t lstart) 1861da177e4SLinus Torvalds { 187b45b5bd6SDavid Gibson struct address_space *mapping = &inode->i_data; 1881da177e4SLinus Torvalds const pgoff_t start = lstart >> HPAGE_SHIFT; 1891da177e4SLinus Torvalds struct pagevec pvec; 1901da177e4SLinus Torvalds pgoff_t next; 191a43a8c39SChen, Kenneth W int i, freed = 0; 1921da177e4SLinus Torvalds 1931da177e4SLinus Torvalds pagevec_init(&pvec, 0); 1941da177e4SLinus Torvalds next = start; 1951da177e4SLinus Torvalds while (1) { 1961da177e4SLinus Torvalds if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { 1971da177e4SLinus Torvalds if (next == start) 1981da177e4SLinus Torvalds break; 1991da177e4SLinus Torvalds next = start; 2001da177e4SLinus Torvalds continue; 2011da177e4SLinus Torvalds } 2021da177e4SLinus Torvalds 2031da177e4SLinus Torvalds for (i = 0; i < pagevec_count(&pvec); ++i) { 2041da177e4SLinus Torvalds struct page *page = pvec.pages[i]; 2051da177e4SLinus Torvalds 2061da177e4SLinus Torvalds lock_page(page); 2071da177e4SLinus Torvalds if (page->index > next) 2081da177e4SLinus Torvalds next = page->index; 2091da177e4SLinus Torvalds ++next; 2101da177e4SLinus Torvalds truncate_huge_page(page); 2111da177e4SLinus Torvalds unlock_page(page); 2121da177e4SLinus Torvalds hugetlb_put_quota(mapping); 213a43a8c39SChen, Kenneth W freed++; 2141da177e4SLinus Torvalds } 2151da177e4SLinus Torvalds huge_pagevec_release(&pvec); 2161da177e4SLinus Torvalds } 2171da177e4SLinus Torvalds BUG_ON(!lstart && mapping->nrpages); 218a43a8c39SChen, Kenneth W hugetlb_unreserve_pages(inode, start, freed); 2191da177e4SLinus Torvalds } 2201da177e4SLinus Torvalds 2211da177e4SLinus Torvalds static void hugetlbfs_delete_inode(struct inode *inode) 2221da177e4SLinus Torvalds { 223b45b5bd6SDavid Gibson truncate_hugepages(inode, 0); 224149f4211SChristoph Hellwig clear_inode(inode); 225149f4211SChristoph Hellwig } 226149f4211SChristoph Hellwig 227ddc0a51dSJosh Triplett static void hugetlbfs_forget_inode(struct inode *inode) __releases(inode_lock) 2281da177e4SLinus Torvalds { 2290b1533f6SChristoph Hellwig struct super_block *sb = inode->i_sb; 2301da177e4SLinus Torvalds 2310b1533f6SChristoph Hellwig if (!hlist_unhashed(&inode->i_hash)) { 2320b1533f6SChristoph Hellwig if (!(inode->i_state & (I_DIRTY|I_LOCK))) 2330b1533f6SChristoph Hellwig list_move(&inode->i_list, &inode_unused); 2341da177e4SLinus Torvalds inodes_stat.nr_unused++; 2350b1533f6SChristoph Hellwig if (!sb || (sb->s_flags & MS_ACTIVE)) { 2361da177e4SLinus Torvalds spin_unlock(&inode_lock); 2371da177e4SLinus Torvalds return; 2381da177e4SLinus Torvalds } 2390b1533f6SChristoph Hellwig inode->i_state |= I_WILL_FREE; 2400b1533f6SChristoph Hellwig spin_unlock(&inode_lock); 2410b1533f6SChristoph Hellwig /* 2420b1533f6SChristoph Hellwig * write_inode_now is a noop as we set BDI_CAP_NO_WRITEBACK 2430b1533f6SChristoph Hellwig * in our backing_dev_info. 2440b1533f6SChristoph Hellwig */ 2450b1533f6SChristoph Hellwig write_inode_now(inode, 1); 2460b1533f6SChristoph Hellwig spin_lock(&inode_lock); 2470b1533f6SChristoph Hellwig inode->i_state &= ~I_WILL_FREE; 2481da177e4SLinus Torvalds inodes_stat.nr_unused--; 2491da177e4SLinus Torvalds hlist_del_init(&inode->i_hash); 2500b1533f6SChristoph Hellwig } 2511da177e4SLinus Torvalds list_del_init(&inode->i_list); 2521da177e4SLinus Torvalds list_del_init(&inode->i_sb_list); 2531da177e4SLinus Torvalds inode->i_state |= I_FREEING; 2541da177e4SLinus Torvalds inodes_stat.nr_inodes--; 2551da177e4SLinus Torvalds spin_unlock(&inode_lock); 256b45b5bd6SDavid Gibson truncate_hugepages(inode, 0); 2571da177e4SLinus Torvalds clear_inode(inode); 2581da177e4SLinus Torvalds destroy_inode(inode); 2591da177e4SLinus Torvalds } 2601da177e4SLinus Torvalds 2611da177e4SLinus Torvalds static void hugetlbfs_drop_inode(struct inode *inode) 2621da177e4SLinus Torvalds { 2631da177e4SLinus Torvalds if (!inode->i_nlink) 2646b09b9dfSChristoph Hellwig generic_delete_inode(inode); 2651da177e4SLinus Torvalds else 2661da177e4SLinus Torvalds hugetlbfs_forget_inode(inode); 2671da177e4SLinus Torvalds } 2681da177e4SLinus Torvalds 2691da177e4SLinus Torvalds static inline void 270856fc295SHugh Dickins hugetlb_vmtruncate_list(struct prio_tree_root *root, pgoff_t pgoff) 2711da177e4SLinus Torvalds { 2721da177e4SLinus Torvalds struct vm_area_struct *vma; 2731da177e4SLinus Torvalds struct prio_tree_iter iter; 2741da177e4SLinus Torvalds 275856fc295SHugh Dickins vma_prio_tree_foreach(vma, &iter, root, pgoff, ULONG_MAX) { 2761da177e4SLinus Torvalds unsigned long v_offset; 2771da177e4SLinus Torvalds 2781da177e4SLinus Torvalds /* 279856fc295SHugh Dickins * Can the expression below overflow on 32-bit arches? 280856fc295SHugh Dickins * No, because the prio_tree returns us only those vmas 281856fc295SHugh Dickins * which overlap the truncated area starting at pgoff, 282856fc295SHugh Dickins * and no vma on a 32-bit arch can span beyond the 4GB. 2831da177e4SLinus Torvalds */ 284856fc295SHugh Dickins if (vma->vm_pgoff < pgoff) 285856fc295SHugh Dickins v_offset = (pgoff - vma->vm_pgoff) << PAGE_SHIFT; 286856fc295SHugh Dickins else 2871da177e4SLinus Torvalds v_offset = 0; 2881da177e4SLinus Torvalds 289502717f4SChen, Kenneth W __unmap_hugepage_range(vma, 290508034a3SHugh Dickins vma->vm_start + v_offset, vma->vm_end); 2911da177e4SLinus Torvalds } 2921da177e4SLinus Torvalds } 2931da177e4SLinus Torvalds 2941da177e4SLinus Torvalds /* 2951da177e4SLinus Torvalds * Expanding truncates are not allowed. 2961da177e4SLinus Torvalds */ 2971da177e4SLinus Torvalds static int hugetlb_vmtruncate(struct inode *inode, loff_t offset) 2981da177e4SLinus Torvalds { 299856fc295SHugh Dickins pgoff_t pgoff; 3001da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 3011da177e4SLinus Torvalds 3021da177e4SLinus Torvalds if (offset > inode->i_size) 3031da177e4SLinus Torvalds return -EINVAL; 3041da177e4SLinus Torvalds 3051da177e4SLinus Torvalds BUG_ON(offset & ~HPAGE_MASK); 306856fc295SHugh Dickins pgoff = offset >> PAGE_SHIFT; 3071da177e4SLinus Torvalds 3081da177e4SLinus Torvalds inode->i_size = offset; 3091da177e4SLinus Torvalds spin_lock(&mapping->i_mmap_lock); 3101da177e4SLinus Torvalds if (!prio_tree_empty(&mapping->i_mmap)) 3111da177e4SLinus Torvalds hugetlb_vmtruncate_list(&mapping->i_mmap, pgoff); 3121da177e4SLinus Torvalds spin_unlock(&mapping->i_mmap_lock); 313b45b5bd6SDavid Gibson truncate_hugepages(inode, offset); 3141da177e4SLinus Torvalds return 0; 3151da177e4SLinus Torvalds } 3161da177e4SLinus Torvalds 3171da177e4SLinus Torvalds static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr) 3181da177e4SLinus Torvalds { 3191da177e4SLinus Torvalds struct inode *inode = dentry->d_inode; 3201da177e4SLinus Torvalds int error; 3211da177e4SLinus Torvalds unsigned int ia_valid = attr->ia_valid; 3221da177e4SLinus Torvalds 3231da177e4SLinus Torvalds BUG_ON(!inode); 3241da177e4SLinus Torvalds 3251da177e4SLinus Torvalds error = inode_change_ok(inode, attr); 3261da177e4SLinus Torvalds if (error) 3271da177e4SLinus Torvalds goto out; 3281da177e4SLinus Torvalds 3291da177e4SLinus Torvalds if (ia_valid & ATTR_SIZE) { 3301da177e4SLinus Torvalds error = -EINVAL; 3311da177e4SLinus Torvalds if (!(attr->ia_size & ~HPAGE_MASK)) 3321da177e4SLinus Torvalds error = hugetlb_vmtruncate(inode, attr->ia_size); 3331da177e4SLinus Torvalds if (error) 3341da177e4SLinus Torvalds goto out; 3351da177e4SLinus Torvalds attr->ia_valid &= ~ATTR_SIZE; 3361da177e4SLinus Torvalds } 3371da177e4SLinus Torvalds error = inode_setattr(inode, attr); 3381da177e4SLinus Torvalds out: 3391da177e4SLinus Torvalds return error; 3401da177e4SLinus Torvalds } 3411da177e4SLinus Torvalds 3421da177e4SLinus Torvalds static struct inode *hugetlbfs_get_inode(struct super_block *sb, uid_t uid, 3431da177e4SLinus Torvalds gid_t gid, int mode, dev_t dev) 3441da177e4SLinus Torvalds { 3451da177e4SLinus Torvalds struct inode *inode; 3461da177e4SLinus Torvalds 3471da177e4SLinus Torvalds inode = new_inode(sb); 3481da177e4SLinus Torvalds if (inode) { 3491da177e4SLinus Torvalds struct hugetlbfs_inode_info *info; 3501da177e4SLinus Torvalds inode->i_mode = mode; 3511da177e4SLinus Torvalds inode->i_uid = uid; 3521da177e4SLinus Torvalds inode->i_gid = gid; 3531da177e4SLinus Torvalds inode->i_blocks = 0; 3541da177e4SLinus Torvalds inode->i_mapping->a_ops = &hugetlbfs_aops; 3551da177e4SLinus Torvalds inode->i_mapping->backing_dev_info =&hugetlbfs_backing_dev_info; 3561da177e4SLinus Torvalds inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME; 357a43a8c39SChen, Kenneth W INIT_LIST_HEAD(&inode->i_mapping->private_list); 3581da177e4SLinus Torvalds info = HUGETLBFS_I(inode); 3597339ff83SRobin Holt mpol_shared_policy_init(&info->policy, MPOL_DEFAULT, NULL); 3601da177e4SLinus Torvalds switch (mode & S_IFMT) { 3611da177e4SLinus Torvalds default: 3621da177e4SLinus Torvalds init_special_inode(inode, mode, dev); 3631da177e4SLinus Torvalds break; 3641da177e4SLinus Torvalds case S_IFREG: 3651da177e4SLinus Torvalds inode->i_op = &hugetlbfs_inode_operations; 3661da177e4SLinus Torvalds inode->i_fop = &hugetlbfs_file_operations; 3671da177e4SLinus Torvalds break; 3681da177e4SLinus Torvalds case S_IFDIR: 3691da177e4SLinus Torvalds inode->i_op = &hugetlbfs_dir_inode_operations; 3701da177e4SLinus Torvalds inode->i_fop = &simple_dir_operations; 3711da177e4SLinus Torvalds 3721da177e4SLinus Torvalds /* directory inodes start off with i_nlink == 2 (for "." entry) */ 373d8c76e6fSDave Hansen inc_nlink(inode); 3741da177e4SLinus Torvalds break; 3751da177e4SLinus Torvalds case S_IFLNK: 3761da177e4SLinus Torvalds inode->i_op = &page_symlink_inode_operations; 3771da177e4SLinus Torvalds break; 3781da177e4SLinus Torvalds } 3791da177e4SLinus Torvalds } 3801da177e4SLinus Torvalds return inode; 3811da177e4SLinus Torvalds } 3821da177e4SLinus Torvalds 3831da177e4SLinus Torvalds /* 3841da177e4SLinus Torvalds * File creation. Allocate an inode, and we're done.. 3851da177e4SLinus Torvalds */ 3861da177e4SLinus Torvalds static int hugetlbfs_mknod(struct inode *dir, 3871da177e4SLinus Torvalds struct dentry *dentry, int mode, dev_t dev) 3881da177e4SLinus Torvalds { 3891da177e4SLinus Torvalds struct inode *inode; 3901da177e4SLinus Torvalds int error = -ENOSPC; 3911da177e4SLinus Torvalds gid_t gid; 3921da177e4SLinus Torvalds 3931da177e4SLinus Torvalds if (dir->i_mode & S_ISGID) { 3941da177e4SLinus Torvalds gid = dir->i_gid; 3951da177e4SLinus Torvalds if (S_ISDIR(mode)) 3961da177e4SLinus Torvalds mode |= S_ISGID; 3971da177e4SLinus Torvalds } else { 3981da177e4SLinus Torvalds gid = current->fsgid; 3991da177e4SLinus Torvalds } 4001da177e4SLinus Torvalds inode = hugetlbfs_get_inode(dir->i_sb, current->fsuid, gid, mode, dev); 4011da177e4SLinus Torvalds if (inode) { 4021da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 4031da177e4SLinus Torvalds d_instantiate(dentry, inode); 4041da177e4SLinus Torvalds dget(dentry); /* Extra count - pin the dentry in core */ 4051da177e4SLinus Torvalds error = 0; 4061da177e4SLinus Torvalds } 4071da177e4SLinus Torvalds return error; 4081da177e4SLinus Torvalds } 4091da177e4SLinus Torvalds 4101da177e4SLinus Torvalds static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, int mode) 4111da177e4SLinus Torvalds { 4121da177e4SLinus Torvalds int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0); 4131da177e4SLinus Torvalds if (!retval) 414d8c76e6fSDave Hansen inc_nlink(dir); 4151da177e4SLinus Torvalds return retval; 4161da177e4SLinus Torvalds } 4171da177e4SLinus Torvalds 4181da177e4SLinus Torvalds static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, int mode, struct nameidata *nd) 4191da177e4SLinus Torvalds { 4201da177e4SLinus Torvalds return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0); 4211da177e4SLinus Torvalds } 4221da177e4SLinus Torvalds 4231da177e4SLinus Torvalds static int hugetlbfs_symlink(struct inode *dir, 4241da177e4SLinus Torvalds struct dentry *dentry, const char *symname) 4251da177e4SLinus Torvalds { 4261da177e4SLinus Torvalds struct inode *inode; 4271da177e4SLinus Torvalds int error = -ENOSPC; 4281da177e4SLinus Torvalds gid_t gid; 4291da177e4SLinus Torvalds 4301da177e4SLinus Torvalds if (dir->i_mode & S_ISGID) 4311da177e4SLinus Torvalds gid = dir->i_gid; 4321da177e4SLinus Torvalds else 4331da177e4SLinus Torvalds gid = current->fsgid; 4341da177e4SLinus Torvalds 4351da177e4SLinus Torvalds inode = hugetlbfs_get_inode(dir->i_sb, current->fsuid, 4361da177e4SLinus Torvalds gid, S_IFLNK|S_IRWXUGO, 0); 4371da177e4SLinus Torvalds if (inode) { 4381da177e4SLinus Torvalds int l = strlen(symname)+1; 4391da177e4SLinus Torvalds error = page_symlink(inode, symname, l); 4401da177e4SLinus Torvalds if (!error) { 4411da177e4SLinus Torvalds d_instantiate(dentry, inode); 4421da177e4SLinus Torvalds dget(dentry); 4431da177e4SLinus Torvalds } else 4441da177e4SLinus Torvalds iput(inode); 4451da177e4SLinus Torvalds } 4461da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 4471da177e4SLinus Torvalds 4481da177e4SLinus Torvalds return error; 4491da177e4SLinus Torvalds } 4501da177e4SLinus Torvalds 4511da177e4SLinus Torvalds /* 4526649a386SKen Chen * mark the head page dirty 4531da177e4SLinus Torvalds */ 4541da177e4SLinus Torvalds static int hugetlbfs_set_page_dirty(struct page *page) 4551da177e4SLinus Torvalds { 4566649a386SKen Chen struct page *head = (struct page *)page_private(page); 4576649a386SKen Chen 4586649a386SKen Chen SetPageDirty(head); 4591da177e4SLinus Torvalds return 0; 4601da177e4SLinus Torvalds } 4611da177e4SLinus Torvalds 462726c3342SDavid Howells static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf) 4631da177e4SLinus Torvalds { 464726c3342SDavid Howells struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb); 4651da177e4SLinus Torvalds 4661da177e4SLinus Torvalds buf->f_type = HUGETLBFS_MAGIC; 4671da177e4SLinus Torvalds buf->f_bsize = HPAGE_SIZE; 4681da177e4SLinus Torvalds if (sbinfo) { 4691da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 47074a8a65cSDavid Gibson /* If no limits set, just report 0 for max/free/used 47174a8a65cSDavid Gibson * blocks, like simple_statfs() */ 47274a8a65cSDavid Gibson if (sbinfo->max_blocks >= 0) { 4731da177e4SLinus Torvalds buf->f_blocks = sbinfo->max_blocks; 4741da177e4SLinus Torvalds buf->f_bavail = buf->f_bfree = sbinfo->free_blocks; 4751da177e4SLinus Torvalds buf->f_files = sbinfo->max_inodes; 4761da177e4SLinus Torvalds buf->f_ffree = sbinfo->free_inodes; 47774a8a65cSDavid Gibson } 4781da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 4791da177e4SLinus Torvalds } 4801da177e4SLinus Torvalds buf->f_namelen = NAME_MAX; 4811da177e4SLinus Torvalds return 0; 4821da177e4SLinus Torvalds } 4831da177e4SLinus Torvalds 4841da177e4SLinus Torvalds static void hugetlbfs_put_super(struct super_block *sb) 4851da177e4SLinus Torvalds { 4861da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb); 4871da177e4SLinus Torvalds 4881da177e4SLinus Torvalds if (sbi) { 4891da177e4SLinus Torvalds sb->s_fs_info = NULL; 4901da177e4SLinus Torvalds kfree(sbi); 4911da177e4SLinus Torvalds } 4921da177e4SLinus Torvalds } 4931da177e4SLinus Torvalds 49496527980SChristoph Hellwig static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo) 49596527980SChristoph Hellwig { 49696527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 49796527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 49896527980SChristoph Hellwig if (unlikely(!sbinfo->free_inodes)) { 49996527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 50096527980SChristoph Hellwig return 0; 50196527980SChristoph Hellwig } 50296527980SChristoph Hellwig sbinfo->free_inodes--; 50396527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 50496527980SChristoph Hellwig } 50596527980SChristoph Hellwig 50696527980SChristoph Hellwig return 1; 50796527980SChristoph Hellwig } 50896527980SChristoph Hellwig 50996527980SChristoph Hellwig static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo) 51096527980SChristoph Hellwig { 51196527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 51296527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 51396527980SChristoph Hellwig sbinfo->free_inodes++; 51496527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 51596527980SChristoph Hellwig } 51696527980SChristoph Hellwig } 51796527980SChristoph Hellwig 51896527980SChristoph Hellwig 519e18b890bSChristoph Lameter static struct kmem_cache *hugetlbfs_inode_cachep; 5201da177e4SLinus Torvalds 5211da177e4SLinus Torvalds static struct inode *hugetlbfs_alloc_inode(struct super_block *sb) 5221da177e4SLinus Torvalds { 52396527980SChristoph Hellwig struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb); 5241da177e4SLinus Torvalds struct hugetlbfs_inode_info *p; 5251da177e4SLinus Torvalds 52696527980SChristoph Hellwig if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo))) 5271da177e4SLinus Torvalds return NULL; 528e94b1766SChristoph Lameter p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL); 52996527980SChristoph Hellwig if (unlikely(!p)) { 53096527980SChristoph Hellwig hugetlbfs_inc_free_inodes(sbinfo); 53196527980SChristoph Hellwig return NULL; 5321da177e4SLinus Torvalds } 53396527980SChristoph Hellwig return &p->vfs_inode; 5341da177e4SLinus Torvalds } 5351da177e4SLinus Torvalds 5361da177e4SLinus Torvalds static void hugetlbfs_destroy_inode(struct inode *inode) 5371da177e4SLinus Torvalds { 53896527980SChristoph Hellwig hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb)); 5391da177e4SLinus Torvalds mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy); 5401da177e4SLinus Torvalds kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode)); 5411da177e4SLinus Torvalds } 5421da177e4SLinus Torvalds 543f5e54d6eSChristoph Hellwig static const struct address_space_operations hugetlbfs_aops = { 5441da177e4SLinus Torvalds .readpage = hugetlbfs_readpage, 5451da177e4SLinus Torvalds .prepare_write = hugetlbfs_prepare_write, 5461da177e4SLinus Torvalds .commit_write = hugetlbfs_commit_write, 5471da177e4SLinus Torvalds .set_page_dirty = hugetlbfs_set_page_dirty, 5481da177e4SLinus Torvalds }; 5491da177e4SLinus Torvalds 55096527980SChristoph Hellwig 551e18b890bSChristoph Lameter static void init_once(void *foo, struct kmem_cache *cachep, unsigned long flags) 55296527980SChristoph Hellwig { 55396527980SChristoph Hellwig struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo; 55496527980SChristoph Hellwig 55596527980SChristoph Hellwig if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) == 55696527980SChristoph Hellwig SLAB_CTOR_CONSTRUCTOR) 55796527980SChristoph Hellwig inode_init_once(&ei->vfs_inode); 55896527980SChristoph Hellwig } 55996527980SChristoph Hellwig 5604b6f5d20SArjan van de Ven const struct file_operations hugetlbfs_file_operations = { 5611da177e4SLinus Torvalds .mmap = hugetlbfs_file_mmap, 5621da177e4SLinus Torvalds .fsync = simple_sync_file, 5631da177e4SLinus Torvalds .get_unmapped_area = hugetlb_get_unmapped_area, 5641da177e4SLinus Torvalds }; 5651da177e4SLinus Torvalds 56692e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_dir_inode_operations = { 5671da177e4SLinus Torvalds .create = hugetlbfs_create, 5681da177e4SLinus Torvalds .lookup = simple_lookup, 5691da177e4SLinus Torvalds .link = simple_link, 5701da177e4SLinus Torvalds .unlink = simple_unlink, 5711da177e4SLinus Torvalds .symlink = hugetlbfs_symlink, 5721da177e4SLinus Torvalds .mkdir = hugetlbfs_mkdir, 5731da177e4SLinus Torvalds .rmdir = simple_rmdir, 5741da177e4SLinus Torvalds .mknod = hugetlbfs_mknod, 5751da177e4SLinus Torvalds .rename = simple_rename, 5761da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 5771da177e4SLinus Torvalds }; 5781da177e4SLinus Torvalds 57992e1d5beSArjan van de Ven static const struct inode_operations hugetlbfs_inode_operations = { 5801da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 5811da177e4SLinus Torvalds }; 5821da177e4SLinus Torvalds 5831da177e4SLinus Torvalds static struct super_operations hugetlbfs_ops = { 5841da177e4SLinus Torvalds .alloc_inode = hugetlbfs_alloc_inode, 5851da177e4SLinus Torvalds .destroy_inode = hugetlbfs_destroy_inode, 5861da177e4SLinus Torvalds .statfs = hugetlbfs_statfs, 587149f4211SChristoph Hellwig .delete_inode = hugetlbfs_delete_inode, 5881da177e4SLinus Torvalds .drop_inode = hugetlbfs_drop_inode, 5891da177e4SLinus Torvalds .put_super = hugetlbfs_put_super, 5901da177e4SLinus Torvalds }; 5911da177e4SLinus Torvalds 5921da177e4SLinus Torvalds static int 5931da177e4SLinus Torvalds hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig) 5941da177e4SLinus Torvalds { 5951da177e4SLinus Torvalds char *opt, *value, *rest; 5961da177e4SLinus Torvalds 5971da177e4SLinus Torvalds if (!options) 5981da177e4SLinus Torvalds return 0; 5991da177e4SLinus Torvalds while ((opt = strsep(&options, ",")) != NULL) { 6001da177e4SLinus Torvalds if (!*opt) 6011da177e4SLinus Torvalds continue; 6021da177e4SLinus Torvalds 6031da177e4SLinus Torvalds value = strchr(opt, '='); 6041da177e4SLinus Torvalds if (!value || !*value) 6051da177e4SLinus Torvalds return -EINVAL; 6061da177e4SLinus Torvalds else 6071da177e4SLinus Torvalds *value++ = '\0'; 6081da177e4SLinus Torvalds 6091da177e4SLinus Torvalds if (!strcmp(opt, "uid")) 6101da177e4SLinus Torvalds pconfig->uid = simple_strtoul(value, &value, 0); 6111da177e4SLinus Torvalds else if (!strcmp(opt, "gid")) 6121da177e4SLinus Torvalds pconfig->gid = simple_strtoul(value, &value, 0); 6131da177e4SLinus Torvalds else if (!strcmp(opt, "mode")) 6141da177e4SLinus Torvalds pconfig->mode = simple_strtoul(value,&value,0) & 0777U; 6151da177e4SLinus Torvalds else if (!strcmp(opt, "size")) { 6161da177e4SLinus Torvalds unsigned long long size = memparse(value, &rest); 6171da177e4SLinus Torvalds if (*rest == '%') { 6181da177e4SLinus Torvalds size <<= HPAGE_SHIFT; 6191da177e4SLinus Torvalds size *= max_huge_pages; 6201da177e4SLinus Torvalds do_div(size, 100); 6211da177e4SLinus Torvalds rest++; 6221da177e4SLinus Torvalds } 6231da177e4SLinus Torvalds pconfig->nr_blocks = (size >> HPAGE_SHIFT); 6241da177e4SLinus Torvalds value = rest; 6251da177e4SLinus Torvalds } else if (!strcmp(opt,"nr_inodes")) { 6261da177e4SLinus Torvalds pconfig->nr_inodes = memparse(value, &rest); 6271da177e4SLinus Torvalds value = rest; 6281da177e4SLinus Torvalds } else 6291da177e4SLinus Torvalds return -EINVAL; 6301da177e4SLinus Torvalds 6311da177e4SLinus Torvalds if (*value) 6321da177e4SLinus Torvalds return -EINVAL; 6331da177e4SLinus Torvalds } 6341da177e4SLinus Torvalds return 0; 6351da177e4SLinus Torvalds } 6361da177e4SLinus Torvalds 6371da177e4SLinus Torvalds static int 6381da177e4SLinus Torvalds hugetlbfs_fill_super(struct super_block *sb, void *data, int silent) 6391da177e4SLinus Torvalds { 6401da177e4SLinus Torvalds struct inode * inode; 6411da177e4SLinus Torvalds struct dentry * root; 6421da177e4SLinus Torvalds int ret; 6431da177e4SLinus Torvalds struct hugetlbfs_config config; 6441da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo; 6451da177e4SLinus Torvalds 6461da177e4SLinus Torvalds config.nr_blocks = -1; /* No limit on size by default */ 6471da177e4SLinus Torvalds config.nr_inodes = -1; /* No limit on number of inodes by default */ 6481da177e4SLinus Torvalds config.uid = current->fsuid; 6491da177e4SLinus Torvalds config.gid = current->fsgid; 6501da177e4SLinus Torvalds config.mode = 0755; 6511da177e4SLinus Torvalds ret = hugetlbfs_parse_options(data, &config); 6521da177e4SLinus Torvalds 6531da177e4SLinus Torvalds if (ret) 6541da177e4SLinus Torvalds return ret; 6551da177e4SLinus Torvalds 6561da177e4SLinus Torvalds sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL); 6571da177e4SLinus Torvalds if (!sbinfo) 6581da177e4SLinus Torvalds return -ENOMEM; 6591da177e4SLinus Torvalds sb->s_fs_info = sbinfo; 6601da177e4SLinus Torvalds spin_lock_init(&sbinfo->stat_lock); 6611da177e4SLinus Torvalds sbinfo->max_blocks = config.nr_blocks; 6621da177e4SLinus Torvalds sbinfo->free_blocks = config.nr_blocks; 6631da177e4SLinus Torvalds sbinfo->max_inodes = config.nr_inodes; 6641da177e4SLinus Torvalds sbinfo->free_inodes = config.nr_inodes; 6651da177e4SLinus Torvalds sb->s_maxbytes = MAX_LFS_FILESIZE; 6661da177e4SLinus Torvalds sb->s_blocksize = HPAGE_SIZE; 6671da177e4SLinus Torvalds sb->s_blocksize_bits = HPAGE_SHIFT; 6681da177e4SLinus Torvalds sb->s_magic = HUGETLBFS_MAGIC; 6691da177e4SLinus Torvalds sb->s_op = &hugetlbfs_ops; 6701da177e4SLinus Torvalds sb->s_time_gran = 1; 6711da177e4SLinus Torvalds inode = hugetlbfs_get_inode(sb, config.uid, config.gid, 6721da177e4SLinus Torvalds S_IFDIR | config.mode, 0); 6731da177e4SLinus Torvalds if (!inode) 6741da177e4SLinus Torvalds goto out_free; 6751da177e4SLinus Torvalds 6761da177e4SLinus Torvalds root = d_alloc_root(inode); 6771da177e4SLinus Torvalds if (!root) { 6781da177e4SLinus Torvalds iput(inode); 6791da177e4SLinus Torvalds goto out_free; 6801da177e4SLinus Torvalds } 6811da177e4SLinus Torvalds sb->s_root = root; 6821da177e4SLinus Torvalds return 0; 6831da177e4SLinus Torvalds out_free: 6841da177e4SLinus Torvalds kfree(sbinfo); 6851da177e4SLinus Torvalds return -ENOMEM; 6861da177e4SLinus Torvalds } 6871da177e4SLinus Torvalds 6881da177e4SLinus Torvalds int hugetlb_get_quota(struct address_space *mapping) 6891da177e4SLinus Torvalds { 6901da177e4SLinus Torvalds int ret = 0; 6911da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(mapping->host->i_sb); 6921da177e4SLinus Torvalds 6931da177e4SLinus Torvalds if (sbinfo->free_blocks > -1) { 6941da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 6951da177e4SLinus Torvalds if (sbinfo->free_blocks > 0) 6961da177e4SLinus Torvalds sbinfo->free_blocks--; 6971da177e4SLinus Torvalds else 6981da177e4SLinus Torvalds ret = -ENOMEM; 6991da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 7001da177e4SLinus Torvalds } 7011da177e4SLinus Torvalds 7021da177e4SLinus Torvalds return ret; 7031da177e4SLinus Torvalds } 7041da177e4SLinus Torvalds 7051da177e4SLinus Torvalds void hugetlb_put_quota(struct address_space *mapping) 7061da177e4SLinus Torvalds { 7071da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(mapping->host->i_sb); 7081da177e4SLinus Torvalds 7091da177e4SLinus Torvalds if (sbinfo->free_blocks > -1) { 7101da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 7111da177e4SLinus Torvalds sbinfo->free_blocks++; 7121da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 7131da177e4SLinus Torvalds } 7141da177e4SLinus Torvalds } 7151da177e4SLinus Torvalds 716454e2398SDavid Howells static int hugetlbfs_get_sb(struct file_system_type *fs_type, 717454e2398SDavid Howells int flags, const char *dev_name, void *data, struct vfsmount *mnt) 7181da177e4SLinus Torvalds { 719454e2398SDavid Howells return get_sb_nodev(fs_type, flags, data, hugetlbfs_fill_super, mnt); 7201da177e4SLinus Torvalds } 7211da177e4SLinus Torvalds 7221da177e4SLinus Torvalds static struct file_system_type hugetlbfs_fs_type = { 7231da177e4SLinus Torvalds .name = "hugetlbfs", 7241da177e4SLinus Torvalds .get_sb = hugetlbfs_get_sb, 7251da177e4SLinus Torvalds .kill_sb = kill_litter_super, 7261da177e4SLinus Torvalds }; 7271da177e4SLinus Torvalds 7281da177e4SLinus Torvalds static struct vfsmount *hugetlbfs_vfsmount; 7291da177e4SLinus Torvalds 7301da177e4SLinus Torvalds static int can_do_hugetlb_shm(void) 7311da177e4SLinus Torvalds { 7321da177e4SLinus Torvalds return likely(capable(CAP_IPC_LOCK) || 7331da177e4SLinus Torvalds in_group_p(sysctl_hugetlb_shm_group) || 7341da177e4SLinus Torvalds can_do_mlock()); 7351da177e4SLinus Torvalds } 7361da177e4SLinus Torvalds 7371da177e4SLinus Torvalds struct file *hugetlb_zero_setup(size_t size) 7381da177e4SLinus Torvalds { 7391da177e4SLinus Torvalds int error = -ENOMEM; 7401da177e4SLinus Torvalds struct file *file; 7411da177e4SLinus Torvalds struct inode *inode; 7421da177e4SLinus Torvalds struct dentry *dentry, *root; 7431da177e4SLinus Torvalds struct qstr quick_string; 7441da177e4SLinus Torvalds char buf[16]; 745bba1e9b2SChen, Kenneth W static atomic_t counter; 7461da177e4SLinus Torvalds 7471da177e4SLinus Torvalds if (!can_do_hugetlb_shm()) 7481da177e4SLinus Torvalds return ERR_PTR(-EPERM); 7491da177e4SLinus Torvalds 7501da177e4SLinus Torvalds if (!user_shm_lock(size, current->user)) 7511da177e4SLinus Torvalds return ERR_PTR(-ENOMEM); 7521da177e4SLinus Torvalds 7531da177e4SLinus Torvalds root = hugetlbfs_vfsmount->mnt_root; 754bba1e9b2SChen, Kenneth W snprintf(buf, 16, "%u", atomic_inc_return(&counter)); 7551da177e4SLinus Torvalds quick_string.name = buf; 7561da177e4SLinus Torvalds quick_string.len = strlen(quick_string.name); 7571da177e4SLinus Torvalds quick_string.hash = 0; 7581da177e4SLinus Torvalds dentry = d_alloc(root, &quick_string); 7591da177e4SLinus Torvalds if (!dentry) 7601da177e4SLinus Torvalds goto out_shm_unlock; 7611da177e4SLinus Torvalds 7621da177e4SLinus Torvalds error = -ENFILE; 7631da177e4SLinus Torvalds file = get_empty_filp(); 7641da177e4SLinus Torvalds if (!file) 7651da177e4SLinus Torvalds goto out_dentry; 7661da177e4SLinus Torvalds 7671da177e4SLinus Torvalds error = -ENOSPC; 7681da177e4SLinus Torvalds inode = hugetlbfs_get_inode(root->d_sb, current->fsuid, 7691da177e4SLinus Torvalds current->fsgid, S_IFREG | S_IRWXUGO, 0); 7701da177e4SLinus Torvalds if (!inode) 7711da177e4SLinus Torvalds goto out_file; 7721da177e4SLinus Torvalds 773b45b5bd6SDavid Gibson error = -ENOMEM; 774a43a8c39SChen, Kenneth W if (hugetlb_reserve_pages(inode, 0, size >> HPAGE_SHIFT)) 775b45b5bd6SDavid Gibson goto out_inode; 776b45b5bd6SDavid Gibson 7771da177e4SLinus Torvalds d_instantiate(dentry, inode); 7781da177e4SLinus Torvalds inode->i_size = size; 7791da177e4SLinus Torvalds inode->i_nlink = 0; 780b39424e2SJosef Sipek file->f_path.mnt = mntget(hugetlbfs_vfsmount); 781b39424e2SJosef Sipek file->f_path.dentry = dentry; 7821da177e4SLinus Torvalds file->f_mapping = inode->i_mapping; 7831da177e4SLinus Torvalds file->f_op = &hugetlbfs_file_operations; 7841da177e4SLinus Torvalds file->f_mode = FMODE_WRITE | FMODE_READ; 7851da177e4SLinus Torvalds return file; 7861da177e4SLinus Torvalds 787b45b5bd6SDavid Gibson out_inode: 788b45b5bd6SDavid Gibson iput(inode); 7891da177e4SLinus Torvalds out_file: 7901da177e4SLinus Torvalds put_filp(file); 7911da177e4SLinus Torvalds out_dentry: 7921da177e4SLinus Torvalds dput(dentry); 7931da177e4SLinus Torvalds out_shm_unlock: 7941da177e4SLinus Torvalds user_shm_unlock(size, current->user); 7951da177e4SLinus Torvalds return ERR_PTR(error); 7961da177e4SLinus Torvalds } 7971da177e4SLinus Torvalds 7981da177e4SLinus Torvalds static int __init init_hugetlbfs_fs(void) 7991da177e4SLinus Torvalds { 8001da177e4SLinus Torvalds int error; 8011da177e4SLinus Torvalds struct vfsmount *vfsmount; 8021da177e4SLinus Torvalds 8031da177e4SLinus Torvalds hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache", 8041da177e4SLinus Torvalds sizeof(struct hugetlbfs_inode_info), 8051da177e4SLinus Torvalds 0, 0, init_once, NULL); 8061da177e4SLinus Torvalds if (hugetlbfs_inode_cachep == NULL) 8071da177e4SLinus Torvalds return -ENOMEM; 8081da177e4SLinus Torvalds 8091da177e4SLinus Torvalds error = register_filesystem(&hugetlbfs_fs_type); 8101da177e4SLinus Torvalds if (error) 8111da177e4SLinus Torvalds goto out; 8121da177e4SLinus Torvalds 8131da177e4SLinus Torvalds vfsmount = kern_mount(&hugetlbfs_fs_type); 8141da177e4SLinus Torvalds 8151da177e4SLinus Torvalds if (!IS_ERR(vfsmount)) { 8161da177e4SLinus Torvalds hugetlbfs_vfsmount = vfsmount; 8171da177e4SLinus Torvalds return 0; 8181da177e4SLinus Torvalds } 8191da177e4SLinus Torvalds 8201da177e4SLinus Torvalds error = PTR_ERR(vfsmount); 8211da177e4SLinus Torvalds 8221da177e4SLinus Torvalds out: 8231da177e4SLinus Torvalds if (error) 8241da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 8251da177e4SLinus Torvalds return error; 8261da177e4SLinus Torvalds } 8271da177e4SLinus Torvalds 8281da177e4SLinus Torvalds static void __exit exit_hugetlbfs_fs(void) 8291da177e4SLinus Torvalds { 8301da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 8311da177e4SLinus Torvalds unregister_filesystem(&hugetlbfs_fs_type); 8321da177e4SLinus Torvalds } 8331da177e4SLinus Torvalds 8341da177e4SLinus Torvalds module_init(init_hugetlbfs_fs) 8351da177e4SLinus Torvalds module_exit(exit_hugetlbfs_fs) 8361da177e4SLinus Torvalds 8371da177e4SLinus Torvalds MODULE_LICENSE("GPL"); 838