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> 211da177e4SLinus Torvalds #include <linux/backing-dev.h> 221da177e4SLinus Torvalds #include <linux/hugetlb.h> 231da177e4SLinus Torvalds #include <linux/pagevec.h> 241da177e4SLinus Torvalds #include <linux/quotaops.h> 251da177e4SLinus Torvalds #include <linux/slab.h> 261da177e4SLinus Torvalds #include <linux/dnotify.h> 271da177e4SLinus Torvalds #include <linux/statfs.h> 281da177e4SLinus Torvalds #include <linux/security.h> 291da177e4SLinus Torvalds 301da177e4SLinus Torvalds #include <asm/uaccess.h> 311da177e4SLinus Torvalds 321da177e4SLinus Torvalds /* some random number */ 331da177e4SLinus Torvalds #define HUGETLBFS_MAGIC 0x958458f6 341da177e4SLinus Torvalds 351da177e4SLinus Torvalds static struct super_operations hugetlbfs_ops; 361da177e4SLinus Torvalds static struct address_space_operations hugetlbfs_aops; 371da177e4SLinus Torvalds struct file_operations hugetlbfs_file_operations; 381da177e4SLinus Torvalds static struct inode_operations hugetlbfs_dir_inode_operations; 391da177e4SLinus Torvalds static struct inode_operations hugetlbfs_inode_operations; 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds static struct backing_dev_info hugetlbfs_backing_dev_info = { 421da177e4SLinus Torvalds .ra_pages = 0, /* No readahead */ 431da177e4SLinus Torvalds .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK, 441da177e4SLinus Torvalds }; 451da177e4SLinus Torvalds 461da177e4SLinus Torvalds int sysctl_hugetlb_shm_group; 471da177e4SLinus Torvalds 481da177e4SLinus Torvalds static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma) 491da177e4SLinus Torvalds { 501da177e4SLinus Torvalds struct inode *inode = file->f_dentry->d_inode; 511da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 521da177e4SLinus Torvalds loff_t len, vma_len; 531da177e4SLinus Torvalds int ret; 541da177e4SLinus Torvalds 551da177e4SLinus Torvalds if ((vma->vm_flags & (VM_MAYSHARE | VM_WRITE)) == VM_WRITE) 561da177e4SLinus Torvalds return -EINVAL; 571da177e4SLinus Torvalds 581da177e4SLinus Torvalds if (vma->vm_pgoff & (HPAGE_SIZE / PAGE_SIZE - 1)) 591da177e4SLinus Torvalds return -EINVAL; 601da177e4SLinus Torvalds 611da177e4SLinus Torvalds if (vma->vm_start & ~HPAGE_MASK) 621da177e4SLinus Torvalds return -EINVAL; 631da177e4SLinus Torvalds 641da177e4SLinus Torvalds if (vma->vm_end & ~HPAGE_MASK) 651da177e4SLinus Torvalds return -EINVAL; 661da177e4SLinus Torvalds 671da177e4SLinus Torvalds if (vma->vm_end - vma->vm_start < HPAGE_SIZE) 681da177e4SLinus Torvalds return -EINVAL; 691da177e4SLinus Torvalds 701da177e4SLinus Torvalds vma_len = (loff_t)(vma->vm_end - vma->vm_start); 711da177e4SLinus Torvalds 721da177e4SLinus Torvalds down(&inode->i_sem); 731da177e4SLinus Torvalds file_accessed(file); 741da177e4SLinus Torvalds vma->vm_flags |= VM_HUGETLB | VM_RESERVED; 751da177e4SLinus Torvalds vma->vm_ops = &hugetlb_vm_ops; 761da177e4SLinus Torvalds 771da177e4SLinus Torvalds ret = -ENOMEM; 781da177e4SLinus Torvalds len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); 791da177e4SLinus Torvalds if (!(vma->vm_flags & VM_WRITE) && len > inode->i_size) 801da177e4SLinus Torvalds goto out; 811da177e4SLinus Torvalds 821da177e4SLinus Torvalds ret = hugetlb_prefault(mapping, vma); 831da177e4SLinus Torvalds if (ret) 841da177e4SLinus Torvalds goto out; 851da177e4SLinus Torvalds 861da177e4SLinus Torvalds if (inode->i_size < len) 871da177e4SLinus Torvalds inode->i_size = len; 881da177e4SLinus Torvalds out: 891da177e4SLinus Torvalds up(&inode->i_sem); 901da177e4SLinus Torvalds 911da177e4SLinus Torvalds return ret; 921da177e4SLinus Torvalds } 931da177e4SLinus Torvalds 941da177e4SLinus Torvalds /* 95508034a3SHugh Dickins * Called under down_write(mmap_sem). 961da177e4SLinus Torvalds */ 971da177e4SLinus Torvalds 981da177e4SLinus Torvalds #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA 991da177e4SLinus Torvalds unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, 1001da177e4SLinus Torvalds unsigned long len, unsigned long pgoff, unsigned long flags); 1011da177e4SLinus Torvalds #else 1021da177e4SLinus Torvalds static unsigned long 1031da177e4SLinus Torvalds hugetlb_get_unmapped_area(struct file *file, unsigned long addr, 1041da177e4SLinus Torvalds unsigned long len, unsigned long pgoff, unsigned long flags) 1051da177e4SLinus Torvalds { 1061da177e4SLinus Torvalds struct mm_struct *mm = current->mm; 1071da177e4SLinus Torvalds struct vm_area_struct *vma; 1081da177e4SLinus Torvalds unsigned long start_addr; 1091da177e4SLinus Torvalds 1101da177e4SLinus Torvalds if (len & ~HPAGE_MASK) 1111da177e4SLinus Torvalds return -EINVAL; 1121da177e4SLinus Torvalds if (len > TASK_SIZE) 1131da177e4SLinus Torvalds return -ENOMEM; 1141da177e4SLinus Torvalds 1151da177e4SLinus Torvalds if (addr) { 1161da177e4SLinus Torvalds addr = ALIGN(addr, HPAGE_SIZE); 1171da177e4SLinus Torvalds vma = find_vma(mm, addr); 1181da177e4SLinus Torvalds if (TASK_SIZE - len >= addr && 1191da177e4SLinus Torvalds (!vma || addr + len <= vma->vm_start)) 1201da177e4SLinus Torvalds return addr; 1211da177e4SLinus Torvalds } 1221da177e4SLinus Torvalds 1231da177e4SLinus Torvalds start_addr = mm->free_area_cache; 1241da177e4SLinus Torvalds 1251363c3cdSWolfgang Wander if (len <= mm->cached_hole_size) 1261363c3cdSWolfgang Wander start_addr = TASK_UNMAPPED_BASE; 1271363c3cdSWolfgang Wander 1281da177e4SLinus Torvalds full_search: 1291da177e4SLinus Torvalds addr = ALIGN(start_addr, HPAGE_SIZE); 1301da177e4SLinus Torvalds 1311da177e4SLinus Torvalds for (vma = find_vma(mm, addr); ; vma = vma->vm_next) { 1321da177e4SLinus Torvalds /* At this point: (!vma || addr < vma->vm_end). */ 1331da177e4SLinus Torvalds if (TASK_SIZE - len < addr) { 1341da177e4SLinus Torvalds /* 1351da177e4SLinus Torvalds * Start a new search - just in case we missed 1361da177e4SLinus Torvalds * some holes. 1371da177e4SLinus Torvalds */ 1381da177e4SLinus Torvalds if (start_addr != TASK_UNMAPPED_BASE) { 1391da177e4SLinus Torvalds start_addr = TASK_UNMAPPED_BASE; 1401da177e4SLinus Torvalds goto full_search; 1411da177e4SLinus Torvalds } 1421da177e4SLinus Torvalds return -ENOMEM; 1431da177e4SLinus Torvalds } 1441da177e4SLinus Torvalds 1451da177e4SLinus Torvalds if (!vma || addr + len <= vma->vm_start) 1461da177e4SLinus Torvalds return addr; 1471da177e4SLinus Torvalds addr = ALIGN(vma->vm_end, HPAGE_SIZE); 1481da177e4SLinus Torvalds } 1491da177e4SLinus Torvalds } 1501da177e4SLinus Torvalds #endif 1511da177e4SLinus Torvalds 1521da177e4SLinus Torvalds /* 1531da177e4SLinus Torvalds * Read a page. Again trivial. If it didn't already exist 1541da177e4SLinus Torvalds * in the page cache, it is zero-filled. 1551da177e4SLinus Torvalds */ 1561da177e4SLinus Torvalds static int hugetlbfs_readpage(struct file *file, struct page * page) 1571da177e4SLinus Torvalds { 1581da177e4SLinus Torvalds unlock_page(page); 1591da177e4SLinus Torvalds return -EINVAL; 1601da177e4SLinus Torvalds } 1611da177e4SLinus Torvalds 1621da177e4SLinus Torvalds static int hugetlbfs_prepare_write(struct file *file, 1631da177e4SLinus Torvalds struct page *page, unsigned offset, unsigned to) 1641da177e4SLinus Torvalds { 1651da177e4SLinus Torvalds return -EINVAL; 1661da177e4SLinus Torvalds } 1671da177e4SLinus Torvalds 1681da177e4SLinus Torvalds static int hugetlbfs_commit_write(struct file *file, 1691da177e4SLinus Torvalds struct page *page, unsigned offset, unsigned to) 1701da177e4SLinus Torvalds { 1711da177e4SLinus Torvalds return -EINVAL; 1721da177e4SLinus Torvalds } 1731da177e4SLinus Torvalds 1741da177e4SLinus Torvalds static void huge_pagevec_release(struct pagevec *pvec) 1751da177e4SLinus Torvalds { 1761da177e4SLinus Torvalds int i; 1771da177e4SLinus Torvalds 1781da177e4SLinus Torvalds for (i = 0; i < pagevec_count(pvec); ++i) 1791da177e4SLinus Torvalds put_page(pvec->pages[i]); 1801da177e4SLinus Torvalds 1811da177e4SLinus Torvalds pagevec_reinit(pvec); 1821da177e4SLinus Torvalds } 1831da177e4SLinus Torvalds 1841da177e4SLinus Torvalds static void truncate_huge_page(struct page *page) 1851da177e4SLinus Torvalds { 1861da177e4SLinus Torvalds clear_page_dirty(page); 1871da177e4SLinus Torvalds ClearPageUptodate(page); 1881da177e4SLinus Torvalds remove_from_page_cache(page); 1891da177e4SLinus Torvalds put_page(page); 1901da177e4SLinus Torvalds } 1911da177e4SLinus Torvalds 1921da177e4SLinus Torvalds static void truncate_hugepages(struct address_space *mapping, loff_t lstart) 1931da177e4SLinus Torvalds { 1941da177e4SLinus Torvalds const pgoff_t start = lstart >> HPAGE_SHIFT; 1951da177e4SLinus Torvalds struct pagevec pvec; 1961da177e4SLinus Torvalds pgoff_t next; 1971da177e4SLinus Torvalds int i; 1981da177e4SLinus Torvalds 1991da177e4SLinus Torvalds pagevec_init(&pvec, 0); 2001da177e4SLinus Torvalds next = start; 2011da177e4SLinus Torvalds while (1) { 2021da177e4SLinus Torvalds if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) { 2031da177e4SLinus Torvalds if (next == start) 2041da177e4SLinus Torvalds break; 2051da177e4SLinus Torvalds next = start; 2061da177e4SLinus Torvalds continue; 2071da177e4SLinus Torvalds } 2081da177e4SLinus Torvalds 2091da177e4SLinus Torvalds for (i = 0; i < pagevec_count(&pvec); ++i) { 2101da177e4SLinus Torvalds struct page *page = pvec.pages[i]; 2111da177e4SLinus Torvalds 2121da177e4SLinus Torvalds lock_page(page); 2131da177e4SLinus Torvalds if (page->index > next) 2141da177e4SLinus Torvalds next = page->index; 2151da177e4SLinus Torvalds ++next; 2161da177e4SLinus Torvalds truncate_huge_page(page); 2171da177e4SLinus Torvalds unlock_page(page); 2181da177e4SLinus Torvalds hugetlb_put_quota(mapping); 2191da177e4SLinus Torvalds } 2201da177e4SLinus Torvalds huge_pagevec_release(&pvec); 2211da177e4SLinus Torvalds } 2221da177e4SLinus Torvalds BUG_ON(!lstart && mapping->nrpages); 2231da177e4SLinus Torvalds } 2241da177e4SLinus Torvalds 2251da177e4SLinus Torvalds static void hugetlbfs_delete_inode(struct inode *inode) 2261da177e4SLinus Torvalds { 227149f4211SChristoph Hellwig if (inode->i_data.nrpages) 228149f4211SChristoph Hellwig truncate_hugepages(&inode->i_data, 0); 229149f4211SChristoph Hellwig clear_inode(inode); 230149f4211SChristoph Hellwig } 231149f4211SChristoph Hellwig 232149f4211SChristoph Hellwig static void hugetlbfs_do_delete_inode(struct inode *inode) 233149f4211SChristoph Hellwig { 234149f4211SChristoph Hellwig struct super_operations *op = inode->i_sb->s_op; 235149f4211SChristoph Hellwig 2361da177e4SLinus Torvalds list_del_init(&inode->i_list); 2371da177e4SLinus Torvalds list_del_init(&inode->i_sb_list); 2381da177e4SLinus Torvalds inode->i_state |= I_FREEING; 2391da177e4SLinus Torvalds inodes_stat.nr_inodes--; 2401da177e4SLinus Torvalds spin_unlock(&inode_lock); 2411da177e4SLinus Torvalds 2421da177e4SLinus Torvalds security_inode_delete(inode); 2431da177e4SLinus Torvalds 244149f4211SChristoph Hellwig if (op->delete_inode) { 245149f4211SChristoph Hellwig void (*delete)(struct inode *) = op->delete_inode; 246149f4211SChristoph Hellwig if (!is_bad_inode(inode)) 247149f4211SChristoph Hellwig DQUOT_INIT(inode); 248149f4211SChristoph Hellwig /* Filesystems implementing their own 249149f4211SChristoph Hellwig * s_op->delete_inode are required to call 250149f4211SChristoph Hellwig * truncate_inode_pages and clear_inode() 251149f4211SChristoph Hellwig * internally 252149f4211SChristoph Hellwig */ 253149f4211SChristoph Hellwig delete(inode); 254149f4211SChristoph Hellwig } else { 255149f4211SChristoph Hellwig truncate_inode_pages(&inode->i_data, 0); 2561da177e4SLinus Torvalds clear_inode(inode); 257149f4211SChristoph Hellwig } 258149f4211SChristoph Hellwig 259149f4211SChristoph Hellwig spin_lock(&inode_lock); 260149f4211SChristoph Hellwig hlist_del_init(&inode->i_hash); 261149f4211SChristoph Hellwig spin_unlock(&inode_lock); 262149f4211SChristoph Hellwig wake_up_inode(inode); 263149f4211SChristoph Hellwig if (inode->i_state != I_CLEAR) 264149f4211SChristoph Hellwig BUG(); 2651da177e4SLinus Torvalds destroy_inode(inode); 2661da177e4SLinus Torvalds } 2671da177e4SLinus Torvalds 2681da177e4SLinus Torvalds static void hugetlbfs_forget_inode(struct inode *inode) 2691da177e4SLinus Torvalds { 2701da177e4SLinus Torvalds struct super_block *super_block = inode->i_sb; 2711da177e4SLinus Torvalds 2721da177e4SLinus Torvalds if (hlist_unhashed(&inode->i_hash)) 2731da177e4SLinus Torvalds goto out_truncate; 2741da177e4SLinus Torvalds 2751da177e4SLinus Torvalds if (!(inode->i_state & (I_DIRTY|I_LOCK))) { 2761da177e4SLinus Torvalds list_del(&inode->i_list); 2771da177e4SLinus Torvalds list_add(&inode->i_list, &inode_unused); 2781da177e4SLinus Torvalds } 2791da177e4SLinus Torvalds inodes_stat.nr_unused++; 2801da177e4SLinus Torvalds if (!super_block || (super_block->s_flags & MS_ACTIVE)) { 2811da177e4SLinus Torvalds spin_unlock(&inode_lock); 2821da177e4SLinus Torvalds return; 2831da177e4SLinus Torvalds } 2841da177e4SLinus Torvalds 2851da177e4SLinus Torvalds /* write_inode_now() ? */ 2861da177e4SLinus Torvalds inodes_stat.nr_unused--; 2871da177e4SLinus Torvalds hlist_del_init(&inode->i_hash); 2881da177e4SLinus Torvalds out_truncate: 2891da177e4SLinus Torvalds list_del_init(&inode->i_list); 2901da177e4SLinus Torvalds list_del_init(&inode->i_sb_list); 2911da177e4SLinus Torvalds inode->i_state |= I_FREEING; 2921da177e4SLinus Torvalds inodes_stat.nr_inodes--; 2931da177e4SLinus Torvalds spin_unlock(&inode_lock); 2941da177e4SLinus Torvalds if (inode->i_data.nrpages) 2951da177e4SLinus Torvalds truncate_hugepages(&inode->i_data, 0); 2961da177e4SLinus Torvalds 2971da177e4SLinus Torvalds clear_inode(inode); 2981da177e4SLinus Torvalds destroy_inode(inode); 2991da177e4SLinus Torvalds } 3001da177e4SLinus Torvalds 3011da177e4SLinus Torvalds static void hugetlbfs_drop_inode(struct inode *inode) 3021da177e4SLinus Torvalds { 3031da177e4SLinus Torvalds if (!inode->i_nlink) 304149f4211SChristoph Hellwig hugetlbfs_do_delete_inode(inode); 3051da177e4SLinus Torvalds else 3061da177e4SLinus Torvalds hugetlbfs_forget_inode(inode); 3071da177e4SLinus Torvalds } 3081da177e4SLinus Torvalds 3091da177e4SLinus Torvalds /* 3101da177e4SLinus Torvalds * h_pgoff is in HPAGE_SIZE units. 3111da177e4SLinus Torvalds * vma->vm_pgoff is in PAGE_SIZE units. 3121da177e4SLinus Torvalds */ 3131da177e4SLinus Torvalds static inline void 3141da177e4SLinus Torvalds hugetlb_vmtruncate_list(struct prio_tree_root *root, unsigned long h_pgoff) 3151da177e4SLinus Torvalds { 3161da177e4SLinus Torvalds struct vm_area_struct *vma; 3171da177e4SLinus Torvalds struct prio_tree_iter iter; 3181da177e4SLinus Torvalds 3191da177e4SLinus Torvalds vma_prio_tree_foreach(vma, &iter, root, h_pgoff, ULONG_MAX) { 3201da177e4SLinus Torvalds unsigned long h_vm_pgoff; 3211da177e4SLinus Torvalds unsigned long v_offset; 3221da177e4SLinus Torvalds 3231da177e4SLinus Torvalds h_vm_pgoff = vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT); 3241da177e4SLinus Torvalds v_offset = (h_pgoff - h_vm_pgoff) << HPAGE_SHIFT; 3251da177e4SLinus Torvalds /* 3261da177e4SLinus Torvalds * Is this VMA fully outside the truncation point? 3271da177e4SLinus Torvalds */ 3281da177e4SLinus Torvalds if (h_vm_pgoff >= h_pgoff) 3291da177e4SLinus Torvalds v_offset = 0; 3301da177e4SLinus Torvalds 331508034a3SHugh Dickins unmap_hugepage_range(vma, 332508034a3SHugh Dickins vma->vm_start + v_offset, vma->vm_end); 3331da177e4SLinus Torvalds } 3341da177e4SLinus Torvalds } 3351da177e4SLinus Torvalds 3361da177e4SLinus Torvalds /* 3371da177e4SLinus Torvalds * Expanding truncates are not allowed. 3381da177e4SLinus Torvalds */ 3391da177e4SLinus Torvalds static int hugetlb_vmtruncate(struct inode *inode, loff_t offset) 3401da177e4SLinus Torvalds { 3411da177e4SLinus Torvalds unsigned long pgoff; 3421da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 3431da177e4SLinus Torvalds 3441da177e4SLinus Torvalds if (offset > inode->i_size) 3451da177e4SLinus Torvalds return -EINVAL; 3461da177e4SLinus Torvalds 3471da177e4SLinus Torvalds BUG_ON(offset & ~HPAGE_MASK); 3481da177e4SLinus Torvalds pgoff = offset >> HPAGE_SHIFT; 3491da177e4SLinus Torvalds 3501da177e4SLinus Torvalds inode->i_size = offset; 3511da177e4SLinus Torvalds spin_lock(&mapping->i_mmap_lock); 3521da177e4SLinus Torvalds if (!prio_tree_empty(&mapping->i_mmap)) 3531da177e4SLinus Torvalds hugetlb_vmtruncate_list(&mapping->i_mmap, pgoff); 3541da177e4SLinus Torvalds spin_unlock(&mapping->i_mmap_lock); 3551da177e4SLinus Torvalds truncate_hugepages(mapping, offset); 3561da177e4SLinus Torvalds return 0; 3571da177e4SLinus Torvalds } 3581da177e4SLinus Torvalds 3591da177e4SLinus Torvalds static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr) 3601da177e4SLinus Torvalds { 3611da177e4SLinus Torvalds struct inode *inode = dentry->d_inode; 3621da177e4SLinus Torvalds int error; 3631da177e4SLinus Torvalds unsigned int ia_valid = attr->ia_valid; 3641da177e4SLinus Torvalds 3651da177e4SLinus Torvalds BUG_ON(!inode); 3661da177e4SLinus Torvalds 3671da177e4SLinus Torvalds error = inode_change_ok(inode, attr); 3681da177e4SLinus Torvalds if (error) 3691da177e4SLinus Torvalds goto out; 3701da177e4SLinus Torvalds 3711da177e4SLinus Torvalds if (ia_valid & ATTR_SIZE) { 3721da177e4SLinus Torvalds error = -EINVAL; 3731da177e4SLinus Torvalds if (!(attr->ia_size & ~HPAGE_MASK)) 3741da177e4SLinus Torvalds error = hugetlb_vmtruncate(inode, attr->ia_size); 3751da177e4SLinus Torvalds if (error) 3761da177e4SLinus Torvalds goto out; 3771da177e4SLinus Torvalds attr->ia_valid &= ~ATTR_SIZE; 3781da177e4SLinus Torvalds } 3791da177e4SLinus Torvalds error = inode_setattr(inode, attr); 3801da177e4SLinus Torvalds out: 3811da177e4SLinus Torvalds return error; 3821da177e4SLinus Torvalds } 3831da177e4SLinus Torvalds 3841da177e4SLinus Torvalds static struct inode *hugetlbfs_get_inode(struct super_block *sb, uid_t uid, 3851da177e4SLinus Torvalds gid_t gid, int mode, dev_t dev) 3861da177e4SLinus Torvalds { 3871da177e4SLinus Torvalds struct inode *inode; 3881da177e4SLinus Torvalds 3891da177e4SLinus Torvalds inode = new_inode(sb); 3901da177e4SLinus Torvalds if (inode) { 3911da177e4SLinus Torvalds struct hugetlbfs_inode_info *info; 3921da177e4SLinus Torvalds inode->i_mode = mode; 3931da177e4SLinus Torvalds inode->i_uid = uid; 3941da177e4SLinus Torvalds inode->i_gid = gid; 3951da177e4SLinus Torvalds inode->i_blksize = HPAGE_SIZE; 3961da177e4SLinus Torvalds inode->i_blocks = 0; 3971da177e4SLinus Torvalds inode->i_mapping->a_ops = &hugetlbfs_aops; 3981da177e4SLinus Torvalds inode->i_mapping->backing_dev_info =&hugetlbfs_backing_dev_info; 3991da177e4SLinus Torvalds inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME; 4001da177e4SLinus Torvalds info = HUGETLBFS_I(inode); 4011da177e4SLinus Torvalds mpol_shared_policy_init(&info->policy); 4021da177e4SLinus Torvalds switch (mode & S_IFMT) { 4031da177e4SLinus Torvalds default: 4041da177e4SLinus Torvalds init_special_inode(inode, mode, dev); 4051da177e4SLinus Torvalds break; 4061da177e4SLinus Torvalds case S_IFREG: 4071da177e4SLinus Torvalds inode->i_op = &hugetlbfs_inode_operations; 4081da177e4SLinus Torvalds inode->i_fop = &hugetlbfs_file_operations; 4091da177e4SLinus Torvalds break; 4101da177e4SLinus Torvalds case S_IFDIR: 4111da177e4SLinus Torvalds inode->i_op = &hugetlbfs_dir_inode_operations; 4121da177e4SLinus Torvalds inode->i_fop = &simple_dir_operations; 4131da177e4SLinus Torvalds 4141da177e4SLinus Torvalds /* directory inodes start off with i_nlink == 2 (for "." entry) */ 4151da177e4SLinus Torvalds inode->i_nlink++; 4161da177e4SLinus Torvalds break; 4171da177e4SLinus Torvalds case S_IFLNK: 4181da177e4SLinus Torvalds inode->i_op = &page_symlink_inode_operations; 4191da177e4SLinus Torvalds break; 4201da177e4SLinus Torvalds } 4211da177e4SLinus Torvalds } 4221da177e4SLinus Torvalds return inode; 4231da177e4SLinus Torvalds } 4241da177e4SLinus Torvalds 4251da177e4SLinus Torvalds /* 4261da177e4SLinus Torvalds * File creation. Allocate an inode, and we're done.. 4271da177e4SLinus Torvalds */ 4281da177e4SLinus Torvalds static int hugetlbfs_mknod(struct inode *dir, 4291da177e4SLinus Torvalds struct dentry *dentry, int mode, dev_t dev) 4301da177e4SLinus Torvalds { 4311da177e4SLinus Torvalds struct inode *inode; 4321da177e4SLinus Torvalds int error = -ENOSPC; 4331da177e4SLinus Torvalds gid_t gid; 4341da177e4SLinus Torvalds 4351da177e4SLinus Torvalds if (dir->i_mode & S_ISGID) { 4361da177e4SLinus Torvalds gid = dir->i_gid; 4371da177e4SLinus Torvalds if (S_ISDIR(mode)) 4381da177e4SLinus Torvalds mode |= S_ISGID; 4391da177e4SLinus Torvalds } else { 4401da177e4SLinus Torvalds gid = current->fsgid; 4411da177e4SLinus Torvalds } 4421da177e4SLinus Torvalds inode = hugetlbfs_get_inode(dir->i_sb, current->fsuid, gid, mode, dev); 4431da177e4SLinus Torvalds if (inode) { 4441da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 4451da177e4SLinus Torvalds d_instantiate(dentry, inode); 4461da177e4SLinus Torvalds dget(dentry); /* Extra count - pin the dentry in core */ 4471da177e4SLinus Torvalds error = 0; 4481da177e4SLinus Torvalds } 4491da177e4SLinus Torvalds return error; 4501da177e4SLinus Torvalds } 4511da177e4SLinus Torvalds 4521da177e4SLinus Torvalds static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, int mode) 4531da177e4SLinus Torvalds { 4541da177e4SLinus Torvalds int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0); 4551da177e4SLinus Torvalds if (!retval) 4561da177e4SLinus Torvalds dir->i_nlink++; 4571da177e4SLinus Torvalds return retval; 4581da177e4SLinus Torvalds } 4591da177e4SLinus Torvalds 4601da177e4SLinus Torvalds static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, int mode, struct nameidata *nd) 4611da177e4SLinus Torvalds { 4621da177e4SLinus Torvalds return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0); 4631da177e4SLinus Torvalds } 4641da177e4SLinus Torvalds 4651da177e4SLinus Torvalds static int hugetlbfs_symlink(struct inode *dir, 4661da177e4SLinus Torvalds struct dentry *dentry, const char *symname) 4671da177e4SLinus Torvalds { 4681da177e4SLinus Torvalds struct inode *inode; 4691da177e4SLinus Torvalds int error = -ENOSPC; 4701da177e4SLinus Torvalds gid_t gid; 4711da177e4SLinus Torvalds 4721da177e4SLinus Torvalds if (dir->i_mode & S_ISGID) 4731da177e4SLinus Torvalds gid = dir->i_gid; 4741da177e4SLinus Torvalds else 4751da177e4SLinus Torvalds gid = current->fsgid; 4761da177e4SLinus Torvalds 4771da177e4SLinus Torvalds inode = hugetlbfs_get_inode(dir->i_sb, current->fsuid, 4781da177e4SLinus Torvalds gid, S_IFLNK|S_IRWXUGO, 0); 4791da177e4SLinus Torvalds if (inode) { 4801da177e4SLinus Torvalds int l = strlen(symname)+1; 4811da177e4SLinus Torvalds error = page_symlink(inode, symname, l); 4821da177e4SLinus Torvalds if (!error) { 4831da177e4SLinus Torvalds d_instantiate(dentry, inode); 4841da177e4SLinus Torvalds dget(dentry); 4851da177e4SLinus Torvalds } else 4861da177e4SLinus Torvalds iput(inode); 4871da177e4SLinus Torvalds } 4881da177e4SLinus Torvalds dir->i_ctime = dir->i_mtime = CURRENT_TIME; 4891da177e4SLinus Torvalds 4901da177e4SLinus Torvalds return error; 4911da177e4SLinus Torvalds } 4921da177e4SLinus Torvalds 4931da177e4SLinus Torvalds /* 4941da177e4SLinus Torvalds * For direct-IO reads into hugetlb pages 4951da177e4SLinus Torvalds */ 4961da177e4SLinus Torvalds static int hugetlbfs_set_page_dirty(struct page *page) 4971da177e4SLinus Torvalds { 4981da177e4SLinus Torvalds return 0; 4991da177e4SLinus Torvalds } 5001da177e4SLinus Torvalds 5011da177e4SLinus Torvalds static int hugetlbfs_statfs(struct super_block *sb, struct kstatfs *buf) 5021da177e4SLinus Torvalds { 5031da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb); 5041da177e4SLinus Torvalds 5051da177e4SLinus Torvalds buf->f_type = HUGETLBFS_MAGIC; 5061da177e4SLinus Torvalds buf->f_bsize = HPAGE_SIZE; 5071da177e4SLinus Torvalds if (sbinfo) { 5081da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 5091da177e4SLinus Torvalds buf->f_blocks = sbinfo->max_blocks; 5101da177e4SLinus Torvalds buf->f_bavail = buf->f_bfree = sbinfo->free_blocks; 5111da177e4SLinus Torvalds buf->f_files = sbinfo->max_inodes; 5121da177e4SLinus Torvalds buf->f_ffree = sbinfo->free_inodes; 5131da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 5141da177e4SLinus Torvalds } 5151da177e4SLinus Torvalds buf->f_namelen = NAME_MAX; 5161da177e4SLinus Torvalds return 0; 5171da177e4SLinus Torvalds } 5181da177e4SLinus Torvalds 5191da177e4SLinus Torvalds static void hugetlbfs_put_super(struct super_block *sb) 5201da177e4SLinus Torvalds { 5211da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb); 5221da177e4SLinus Torvalds 5231da177e4SLinus Torvalds if (sbi) { 5241da177e4SLinus Torvalds sb->s_fs_info = NULL; 5251da177e4SLinus Torvalds kfree(sbi); 5261da177e4SLinus Torvalds } 5271da177e4SLinus Torvalds } 5281da177e4SLinus Torvalds 52996527980SChristoph Hellwig static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo) 53096527980SChristoph Hellwig { 53196527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 53296527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 53396527980SChristoph Hellwig if (unlikely(!sbinfo->free_inodes)) { 53496527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 53596527980SChristoph Hellwig return 0; 53696527980SChristoph Hellwig } 53796527980SChristoph Hellwig sbinfo->free_inodes--; 53896527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 53996527980SChristoph Hellwig } 54096527980SChristoph Hellwig 54196527980SChristoph Hellwig return 1; 54296527980SChristoph Hellwig } 54396527980SChristoph Hellwig 54496527980SChristoph Hellwig static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo) 54596527980SChristoph Hellwig { 54696527980SChristoph Hellwig if (sbinfo->free_inodes >= 0) { 54796527980SChristoph Hellwig spin_lock(&sbinfo->stat_lock); 54896527980SChristoph Hellwig sbinfo->free_inodes++; 54996527980SChristoph Hellwig spin_unlock(&sbinfo->stat_lock); 55096527980SChristoph Hellwig } 55196527980SChristoph Hellwig } 55296527980SChristoph Hellwig 55396527980SChristoph Hellwig 5541da177e4SLinus Torvalds static kmem_cache_t *hugetlbfs_inode_cachep; 5551da177e4SLinus Torvalds 5561da177e4SLinus Torvalds static struct inode *hugetlbfs_alloc_inode(struct super_block *sb) 5571da177e4SLinus Torvalds { 55896527980SChristoph Hellwig struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb); 5591da177e4SLinus Torvalds struct hugetlbfs_inode_info *p; 5601da177e4SLinus Torvalds 56196527980SChristoph Hellwig if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo))) 5621da177e4SLinus Torvalds return NULL; 56396527980SChristoph Hellwig p = kmem_cache_alloc(hugetlbfs_inode_cachep, SLAB_KERNEL); 56496527980SChristoph Hellwig if (unlikely(!p)) { 56596527980SChristoph Hellwig hugetlbfs_inc_free_inodes(sbinfo); 56696527980SChristoph Hellwig return NULL; 5671da177e4SLinus Torvalds } 56896527980SChristoph Hellwig return &p->vfs_inode; 5691da177e4SLinus Torvalds } 5701da177e4SLinus Torvalds 5711da177e4SLinus Torvalds static void hugetlbfs_destroy_inode(struct inode *inode) 5721da177e4SLinus Torvalds { 57396527980SChristoph Hellwig hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb)); 5741da177e4SLinus Torvalds mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy); 5751da177e4SLinus Torvalds kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode)); 5761da177e4SLinus Torvalds } 5771da177e4SLinus Torvalds 5781da177e4SLinus Torvalds static struct address_space_operations hugetlbfs_aops = { 5791da177e4SLinus Torvalds .readpage = hugetlbfs_readpage, 5801da177e4SLinus Torvalds .prepare_write = hugetlbfs_prepare_write, 5811da177e4SLinus Torvalds .commit_write = hugetlbfs_commit_write, 5821da177e4SLinus Torvalds .set_page_dirty = hugetlbfs_set_page_dirty, 5831da177e4SLinus Torvalds }; 5841da177e4SLinus Torvalds 58596527980SChristoph Hellwig 58696527980SChristoph Hellwig static void init_once(void *foo, kmem_cache_t *cachep, unsigned long flags) 58796527980SChristoph Hellwig { 58896527980SChristoph Hellwig struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo; 58996527980SChristoph Hellwig 59096527980SChristoph Hellwig if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) == 59196527980SChristoph Hellwig SLAB_CTOR_CONSTRUCTOR) 59296527980SChristoph Hellwig inode_init_once(&ei->vfs_inode); 59396527980SChristoph Hellwig } 59496527980SChristoph Hellwig 5951da177e4SLinus Torvalds struct file_operations hugetlbfs_file_operations = { 5961da177e4SLinus Torvalds .mmap = hugetlbfs_file_mmap, 5971da177e4SLinus Torvalds .fsync = simple_sync_file, 5981da177e4SLinus Torvalds .get_unmapped_area = hugetlb_get_unmapped_area, 5991da177e4SLinus Torvalds }; 6001da177e4SLinus Torvalds 6011da177e4SLinus Torvalds static struct inode_operations hugetlbfs_dir_inode_operations = { 6021da177e4SLinus Torvalds .create = hugetlbfs_create, 6031da177e4SLinus Torvalds .lookup = simple_lookup, 6041da177e4SLinus Torvalds .link = simple_link, 6051da177e4SLinus Torvalds .unlink = simple_unlink, 6061da177e4SLinus Torvalds .symlink = hugetlbfs_symlink, 6071da177e4SLinus Torvalds .mkdir = hugetlbfs_mkdir, 6081da177e4SLinus Torvalds .rmdir = simple_rmdir, 6091da177e4SLinus Torvalds .mknod = hugetlbfs_mknod, 6101da177e4SLinus Torvalds .rename = simple_rename, 6111da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 6121da177e4SLinus Torvalds }; 6131da177e4SLinus Torvalds 6141da177e4SLinus Torvalds static struct inode_operations hugetlbfs_inode_operations = { 6151da177e4SLinus Torvalds .setattr = hugetlbfs_setattr, 6161da177e4SLinus Torvalds }; 6171da177e4SLinus Torvalds 6181da177e4SLinus Torvalds static struct super_operations hugetlbfs_ops = { 6191da177e4SLinus Torvalds .alloc_inode = hugetlbfs_alloc_inode, 6201da177e4SLinus Torvalds .destroy_inode = hugetlbfs_destroy_inode, 6211da177e4SLinus Torvalds .statfs = hugetlbfs_statfs, 622149f4211SChristoph Hellwig .delete_inode = hugetlbfs_delete_inode, 6231da177e4SLinus Torvalds .drop_inode = hugetlbfs_drop_inode, 6241da177e4SLinus Torvalds .put_super = hugetlbfs_put_super, 6251da177e4SLinus Torvalds }; 6261da177e4SLinus Torvalds 6271da177e4SLinus Torvalds static int 6281da177e4SLinus Torvalds hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig) 6291da177e4SLinus Torvalds { 6301da177e4SLinus Torvalds char *opt, *value, *rest; 6311da177e4SLinus Torvalds 6321da177e4SLinus Torvalds if (!options) 6331da177e4SLinus Torvalds return 0; 6341da177e4SLinus Torvalds while ((opt = strsep(&options, ",")) != NULL) { 6351da177e4SLinus Torvalds if (!*opt) 6361da177e4SLinus Torvalds continue; 6371da177e4SLinus Torvalds 6381da177e4SLinus Torvalds value = strchr(opt, '='); 6391da177e4SLinus Torvalds if (!value || !*value) 6401da177e4SLinus Torvalds return -EINVAL; 6411da177e4SLinus Torvalds else 6421da177e4SLinus Torvalds *value++ = '\0'; 6431da177e4SLinus Torvalds 6441da177e4SLinus Torvalds if (!strcmp(opt, "uid")) 6451da177e4SLinus Torvalds pconfig->uid = simple_strtoul(value, &value, 0); 6461da177e4SLinus Torvalds else if (!strcmp(opt, "gid")) 6471da177e4SLinus Torvalds pconfig->gid = simple_strtoul(value, &value, 0); 6481da177e4SLinus Torvalds else if (!strcmp(opt, "mode")) 6491da177e4SLinus Torvalds pconfig->mode = simple_strtoul(value,&value,0) & 0777U; 6501da177e4SLinus Torvalds else if (!strcmp(opt, "size")) { 6511da177e4SLinus Torvalds unsigned long long size = memparse(value, &rest); 6521da177e4SLinus Torvalds if (*rest == '%') { 6531da177e4SLinus Torvalds size <<= HPAGE_SHIFT; 6541da177e4SLinus Torvalds size *= max_huge_pages; 6551da177e4SLinus Torvalds do_div(size, 100); 6561da177e4SLinus Torvalds rest++; 6571da177e4SLinus Torvalds } 6581da177e4SLinus Torvalds size &= HPAGE_MASK; 6591da177e4SLinus Torvalds pconfig->nr_blocks = (size >> HPAGE_SHIFT); 6601da177e4SLinus Torvalds value = rest; 6611da177e4SLinus Torvalds } else if (!strcmp(opt,"nr_inodes")) { 6621da177e4SLinus Torvalds pconfig->nr_inodes = memparse(value, &rest); 6631da177e4SLinus Torvalds value = rest; 6641da177e4SLinus Torvalds } else 6651da177e4SLinus Torvalds return -EINVAL; 6661da177e4SLinus Torvalds 6671da177e4SLinus Torvalds if (*value) 6681da177e4SLinus Torvalds return -EINVAL; 6691da177e4SLinus Torvalds } 6701da177e4SLinus Torvalds return 0; 6711da177e4SLinus Torvalds } 6721da177e4SLinus Torvalds 6731da177e4SLinus Torvalds static int 6741da177e4SLinus Torvalds hugetlbfs_fill_super(struct super_block *sb, void *data, int silent) 6751da177e4SLinus Torvalds { 6761da177e4SLinus Torvalds struct inode * inode; 6771da177e4SLinus Torvalds struct dentry * root; 6781da177e4SLinus Torvalds int ret; 6791da177e4SLinus Torvalds struct hugetlbfs_config config; 6801da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo; 6811da177e4SLinus Torvalds 6821da177e4SLinus Torvalds config.nr_blocks = -1; /* No limit on size by default */ 6831da177e4SLinus Torvalds config.nr_inodes = -1; /* No limit on number of inodes by default */ 6841da177e4SLinus Torvalds config.uid = current->fsuid; 6851da177e4SLinus Torvalds config.gid = current->fsgid; 6861da177e4SLinus Torvalds config.mode = 0755; 6871da177e4SLinus Torvalds ret = hugetlbfs_parse_options(data, &config); 6881da177e4SLinus Torvalds 6891da177e4SLinus Torvalds if (ret) 6901da177e4SLinus Torvalds return ret; 6911da177e4SLinus Torvalds 6921da177e4SLinus Torvalds sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL); 6931da177e4SLinus Torvalds if (!sbinfo) 6941da177e4SLinus Torvalds return -ENOMEM; 6951da177e4SLinus Torvalds sb->s_fs_info = sbinfo; 6961da177e4SLinus Torvalds spin_lock_init(&sbinfo->stat_lock); 6971da177e4SLinus Torvalds sbinfo->max_blocks = config.nr_blocks; 6981da177e4SLinus Torvalds sbinfo->free_blocks = config.nr_blocks; 6991da177e4SLinus Torvalds sbinfo->max_inodes = config.nr_inodes; 7001da177e4SLinus Torvalds sbinfo->free_inodes = config.nr_inodes; 7011da177e4SLinus Torvalds sb->s_maxbytes = MAX_LFS_FILESIZE; 7021da177e4SLinus Torvalds sb->s_blocksize = HPAGE_SIZE; 7031da177e4SLinus Torvalds sb->s_blocksize_bits = HPAGE_SHIFT; 7041da177e4SLinus Torvalds sb->s_magic = HUGETLBFS_MAGIC; 7051da177e4SLinus Torvalds sb->s_op = &hugetlbfs_ops; 7061da177e4SLinus Torvalds sb->s_time_gran = 1; 7071da177e4SLinus Torvalds inode = hugetlbfs_get_inode(sb, config.uid, config.gid, 7081da177e4SLinus Torvalds S_IFDIR | config.mode, 0); 7091da177e4SLinus Torvalds if (!inode) 7101da177e4SLinus Torvalds goto out_free; 7111da177e4SLinus Torvalds 7121da177e4SLinus Torvalds root = d_alloc_root(inode); 7131da177e4SLinus Torvalds if (!root) { 7141da177e4SLinus Torvalds iput(inode); 7151da177e4SLinus Torvalds goto out_free; 7161da177e4SLinus Torvalds } 7171da177e4SLinus Torvalds sb->s_root = root; 7181da177e4SLinus Torvalds return 0; 7191da177e4SLinus Torvalds out_free: 7201da177e4SLinus Torvalds kfree(sbinfo); 7211da177e4SLinus Torvalds return -ENOMEM; 7221da177e4SLinus Torvalds } 7231da177e4SLinus Torvalds 7241da177e4SLinus Torvalds int hugetlb_get_quota(struct address_space *mapping) 7251da177e4SLinus Torvalds { 7261da177e4SLinus Torvalds int ret = 0; 7271da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(mapping->host->i_sb); 7281da177e4SLinus Torvalds 7291da177e4SLinus Torvalds if (sbinfo->free_blocks > -1) { 7301da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 7311da177e4SLinus Torvalds if (sbinfo->free_blocks > 0) 7321da177e4SLinus Torvalds sbinfo->free_blocks--; 7331da177e4SLinus Torvalds else 7341da177e4SLinus Torvalds ret = -ENOMEM; 7351da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 7361da177e4SLinus Torvalds } 7371da177e4SLinus Torvalds 7381da177e4SLinus Torvalds return ret; 7391da177e4SLinus Torvalds } 7401da177e4SLinus Torvalds 7411da177e4SLinus Torvalds void hugetlb_put_quota(struct address_space *mapping) 7421da177e4SLinus Torvalds { 7431da177e4SLinus Torvalds struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(mapping->host->i_sb); 7441da177e4SLinus Torvalds 7451da177e4SLinus Torvalds if (sbinfo->free_blocks > -1) { 7461da177e4SLinus Torvalds spin_lock(&sbinfo->stat_lock); 7471da177e4SLinus Torvalds sbinfo->free_blocks++; 7481da177e4SLinus Torvalds spin_unlock(&sbinfo->stat_lock); 7491da177e4SLinus Torvalds } 7501da177e4SLinus Torvalds } 7511da177e4SLinus Torvalds 7521da177e4SLinus Torvalds static struct super_block *hugetlbfs_get_sb(struct file_system_type *fs_type, 7531da177e4SLinus Torvalds int flags, const char *dev_name, void *data) 7541da177e4SLinus Torvalds { 7551da177e4SLinus Torvalds return get_sb_nodev(fs_type, flags, data, hugetlbfs_fill_super); 7561da177e4SLinus Torvalds } 7571da177e4SLinus Torvalds 7581da177e4SLinus Torvalds static struct file_system_type hugetlbfs_fs_type = { 7591da177e4SLinus Torvalds .name = "hugetlbfs", 7601da177e4SLinus Torvalds .get_sb = hugetlbfs_get_sb, 7611da177e4SLinus Torvalds .kill_sb = kill_litter_super, 7621da177e4SLinus Torvalds }; 7631da177e4SLinus Torvalds 7641da177e4SLinus Torvalds static struct vfsmount *hugetlbfs_vfsmount; 7651da177e4SLinus Torvalds 7661da177e4SLinus Torvalds /* 7671da177e4SLinus Torvalds * Return the next identifier for a shm file 7681da177e4SLinus Torvalds */ 7691da177e4SLinus Torvalds static unsigned long hugetlbfs_counter(void) 7701da177e4SLinus Torvalds { 7711da177e4SLinus Torvalds static DEFINE_SPINLOCK(lock); 7721da177e4SLinus Torvalds static unsigned long counter; 7731da177e4SLinus Torvalds unsigned long ret; 7741da177e4SLinus Torvalds 7751da177e4SLinus Torvalds spin_lock(&lock); 7761da177e4SLinus Torvalds ret = ++counter; 7771da177e4SLinus Torvalds spin_unlock(&lock); 7781da177e4SLinus Torvalds return ret; 7791da177e4SLinus Torvalds } 7801da177e4SLinus Torvalds 7811da177e4SLinus Torvalds static int can_do_hugetlb_shm(void) 7821da177e4SLinus Torvalds { 7831da177e4SLinus Torvalds return likely(capable(CAP_IPC_LOCK) || 7841da177e4SLinus Torvalds in_group_p(sysctl_hugetlb_shm_group) || 7851da177e4SLinus Torvalds can_do_mlock()); 7861da177e4SLinus Torvalds } 7871da177e4SLinus Torvalds 7881da177e4SLinus Torvalds struct file *hugetlb_zero_setup(size_t size) 7891da177e4SLinus Torvalds { 7901da177e4SLinus Torvalds int error = -ENOMEM; 7911da177e4SLinus Torvalds struct file *file; 7921da177e4SLinus Torvalds struct inode *inode; 7931da177e4SLinus Torvalds struct dentry *dentry, *root; 7941da177e4SLinus Torvalds struct qstr quick_string; 7951da177e4SLinus Torvalds char buf[16]; 7961da177e4SLinus Torvalds 7971da177e4SLinus Torvalds if (!can_do_hugetlb_shm()) 7981da177e4SLinus Torvalds return ERR_PTR(-EPERM); 7991da177e4SLinus Torvalds 8001da177e4SLinus Torvalds if (!is_hugepage_mem_enough(size)) 8011da177e4SLinus Torvalds return ERR_PTR(-ENOMEM); 8021da177e4SLinus Torvalds 8031da177e4SLinus Torvalds if (!user_shm_lock(size, current->user)) 8041da177e4SLinus Torvalds return ERR_PTR(-ENOMEM); 8051da177e4SLinus Torvalds 8061da177e4SLinus Torvalds root = hugetlbfs_vfsmount->mnt_root; 8071da177e4SLinus Torvalds snprintf(buf, 16, "%lu", hugetlbfs_counter()); 8081da177e4SLinus Torvalds quick_string.name = buf; 8091da177e4SLinus Torvalds quick_string.len = strlen(quick_string.name); 8101da177e4SLinus Torvalds quick_string.hash = 0; 8111da177e4SLinus Torvalds dentry = d_alloc(root, &quick_string); 8121da177e4SLinus Torvalds if (!dentry) 8131da177e4SLinus Torvalds goto out_shm_unlock; 8141da177e4SLinus Torvalds 8151da177e4SLinus Torvalds error = -ENFILE; 8161da177e4SLinus Torvalds file = get_empty_filp(); 8171da177e4SLinus Torvalds if (!file) 8181da177e4SLinus Torvalds goto out_dentry; 8191da177e4SLinus Torvalds 8201da177e4SLinus Torvalds error = -ENOSPC; 8211da177e4SLinus Torvalds inode = hugetlbfs_get_inode(root->d_sb, current->fsuid, 8221da177e4SLinus Torvalds current->fsgid, S_IFREG | S_IRWXUGO, 0); 8231da177e4SLinus Torvalds if (!inode) 8241da177e4SLinus Torvalds goto out_file; 8251da177e4SLinus Torvalds 8261da177e4SLinus Torvalds d_instantiate(dentry, inode); 8271da177e4SLinus Torvalds inode->i_size = size; 8281da177e4SLinus Torvalds inode->i_nlink = 0; 8291da177e4SLinus Torvalds file->f_vfsmnt = mntget(hugetlbfs_vfsmount); 8301da177e4SLinus Torvalds file->f_dentry = dentry; 8311da177e4SLinus Torvalds file->f_mapping = inode->i_mapping; 8321da177e4SLinus Torvalds file->f_op = &hugetlbfs_file_operations; 8331da177e4SLinus Torvalds file->f_mode = FMODE_WRITE | FMODE_READ; 8341da177e4SLinus Torvalds return file; 8351da177e4SLinus Torvalds 8361da177e4SLinus Torvalds out_file: 8371da177e4SLinus Torvalds put_filp(file); 8381da177e4SLinus Torvalds out_dentry: 8391da177e4SLinus Torvalds dput(dentry); 8401da177e4SLinus Torvalds out_shm_unlock: 8411da177e4SLinus Torvalds user_shm_unlock(size, current->user); 8421da177e4SLinus Torvalds return ERR_PTR(error); 8431da177e4SLinus Torvalds } 8441da177e4SLinus Torvalds 8451da177e4SLinus Torvalds static int __init init_hugetlbfs_fs(void) 8461da177e4SLinus Torvalds { 8471da177e4SLinus Torvalds int error; 8481da177e4SLinus Torvalds struct vfsmount *vfsmount; 8491da177e4SLinus Torvalds 8501da177e4SLinus Torvalds hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache", 8511da177e4SLinus Torvalds sizeof(struct hugetlbfs_inode_info), 8521da177e4SLinus Torvalds 0, 0, init_once, NULL); 8531da177e4SLinus Torvalds if (hugetlbfs_inode_cachep == NULL) 8541da177e4SLinus Torvalds return -ENOMEM; 8551da177e4SLinus Torvalds 8561da177e4SLinus Torvalds error = register_filesystem(&hugetlbfs_fs_type); 8571da177e4SLinus Torvalds if (error) 8581da177e4SLinus Torvalds goto out; 8591da177e4SLinus Torvalds 8601da177e4SLinus Torvalds vfsmount = kern_mount(&hugetlbfs_fs_type); 8611da177e4SLinus Torvalds 8621da177e4SLinus Torvalds if (!IS_ERR(vfsmount)) { 8631da177e4SLinus Torvalds hugetlbfs_vfsmount = vfsmount; 8641da177e4SLinus Torvalds return 0; 8651da177e4SLinus Torvalds } 8661da177e4SLinus Torvalds 8671da177e4SLinus Torvalds error = PTR_ERR(vfsmount); 8681da177e4SLinus Torvalds 8691da177e4SLinus Torvalds out: 8701da177e4SLinus Torvalds if (error) 8711da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 8721da177e4SLinus Torvalds return error; 8731da177e4SLinus Torvalds } 8741da177e4SLinus Torvalds 8751da177e4SLinus Torvalds static void __exit exit_hugetlbfs_fs(void) 8761da177e4SLinus Torvalds { 8771da177e4SLinus Torvalds kmem_cache_destroy(hugetlbfs_inode_cachep); 8781da177e4SLinus Torvalds unregister_filesystem(&hugetlbfs_fs_type); 8791da177e4SLinus Torvalds } 8801da177e4SLinus Torvalds 8811da177e4SLinus Torvalds module_init(init_hugetlbfs_fs) 8821da177e4SLinus Torvalds module_exit(exit_hugetlbfs_fs) 8831da177e4SLinus Torvalds 8841da177e4SLinus Torvalds MODULE_LICENSE("GPL"); 885