| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Copyright IBM Corporation, 2021 |
| 4 | * |
| 5 | * Author: Mike Rapoport <rppt@linux.ibm.com> |
| 6 | */ |
| 7 | |
| 8 | #include <linux/mm.h> |
| 9 | #include <linux/fs.h> |
| 10 | #include <linux/swap.h> |
| 11 | #include <linux/mount.h> |
| 12 | #include <linux/memfd.h> |
| 13 | #include <linux/bitops.h> |
| 14 | #include <linux/printk.h> |
| 15 | #include <linux/pagemap.h> |
| 16 | #include <linux/syscalls.h> |
| 17 | #include <linux/pseudo_fs.h> |
| 18 | #include <linux/secretmem.h> |
| 19 | #include <linux/set_memory.h> |
| 20 | #include <linux/sched/signal.h> |
| 21 | |
| 22 | #include <uapi/linux/magic.h> |
| 23 | |
| 24 | #include <asm/tlbflush.h> |
| 25 | |
| 26 | #include "internal.h" |
| 27 | |
| 28 | #undef pr_fmt |
| 29 | #define pr_fmt(fmt) "secretmem: " fmt |
| 30 | |
| 31 | /* |
| 32 | * Define mode and flag masks to allow validation of the system call |
| 33 | * parameters. |
| 34 | */ |
| 35 | #define SECRETMEM_MODE_MASK (0x0) |
| 36 | #define SECRETMEM_FLAGS_MASK SECRETMEM_MODE_MASK |
| 37 | |
| 38 | static bool secretmem_enable __ro_after_init; |
| 39 | module_param_named(enable, secretmem_enable, bool, 0400); |
| 40 | MODULE_PARM_DESC(secretmem_enable, |
| 41 | "Enable secretmem and memfd_secret(2) system call"); |
| 42 | |
| 43 | static atomic_t secretmem_users; |
| 44 | |
| 45 | bool secretmem_active(void) |
| 46 | { |
| 47 | return !!atomic_read(&secretmem_users); |
| 48 | } |
| 49 | |
| 50 | static vm_fault_t secretmem_fault(struct vm_fault *vmf) |
| 51 | { |
| 52 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
| 53 | struct inode *inode = file_inode(vmf->vma->vm_file); |
| 54 | pgoff_t offset = vmf->pgoff; |
| 55 | gfp_t gfp = vmf->gfp_mask; |
| 56 | unsigned long addr; |
| 57 | struct page *page; |
| 58 | vm_fault_t ret; |
| 59 | int err; |
| 60 | |
| 61 | if (((loff_t)vmf->pgoff << PAGE_SHIFT) >= i_size_read(inode)) |
| 62 | return vmf_error(-EINVAL); |
| 63 | |
| 64 | filemap_invalidate_lock_shared(mapping); |
| 65 | |
| 66 | retry: |
| 67 | page = find_lock_page(mapping, offset); |
| 68 | if (!page) { |
| 69 | page = alloc_page(gfp | __GFP_ZERO); |
| 70 | if (!page) { |
| 71 | ret = VM_FAULT_OOM; |
| 72 | goto out; |
| 73 | } |
| 74 | |
| 75 | err = set_direct_map_invalid_noflush(page); |
| 76 | if (err) { |
| 77 | put_page(page); |
| 78 | ret = vmf_error(err); |
| 79 | goto out; |
| 80 | } |
| 81 | |
| 82 | __SetPageUptodate(page); |
| 83 | err = add_to_page_cache_lru(page, mapping, offset, gfp); |
| 84 | if (unlikely(err)) { |
| 85 | put_page(page); |
| 86 | /* |
| 87 | * If a split of large page was required, it |
| 88 | * already happened when we marked the page invalid |
| 89 | * which guarantees that this call won't fail |
| 90 | */ |
| 91 | set_direct_map_default_noflush(page); |
| 92 | if (err == -EEXIST) |
| 93 | goto retry; |
| 94 | |
| 95 | ret = vmf_error(err); |
| 96 | goto out; |
| 97 | } |
| 98 | |
| 99 | addr = (unsigned long)page_address(page); |
| 100 | flush_tlb_kernel_range(addr, addr + PAGE_SIZE); |
| 101 | } |
| 102 | |
| 103 | vmf->page = page; |
| 104 | ret = VM_FAULT_LOCKED; |
| 105 | |
| 106 | out: |
| 107 | filemap_invalidate_unlock_shared(mapping); |
| 108 | return ret; |
| 109 | } |
| 110 | |
| 111 | static const struct vm_operations_struct secretmem_vm_ops = { |
| 112 | .fault = secretmem_fault, |
| 113 | }; |
| 114 | |
| 115 | static int secretmem_release(struct inode *inode, struct file *file) |
| 116 | { |
| 117 | atomic_dec(&secretmem_users); |
| 118 | return 0; |
| 119 | } |
| 120 | |
| 121 | static int secretmem_mmap(struct file *file, struct vm_area_struct *vma) |
| 122 | { |
| 123 | unsigned long len = vma->vm_end - vma->vm_start; |
| 124 | |
| 125 | if ((vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) == 0) |
| 126 | return -EINVAL; |
| 127 | |
| 128 | if (mlock_future_check(vma->vm_mm, vma->vm_flags | VM_LOCKED, len)) |
| 129 | return -EAGAIN; |
| 130 | |
| 131 | vma->vm_flags |= VM_LOCKED | VM_DONTDUMP; |
| 132 | vma->vm_ops = &secretmem_vm_ops; |
| 133 | |
| 134 | return 0; |
| 135 | } |
| 136 | |
| 137 | bool vma_is_secretmem(struct vm_area_struct *vma) |
| 138 | { |
| 139 | return vma->vm_ops == &secretmem_vm_ops; |
| 140 | } |
| 141 | |
| 142 | static const struct file_operations secretmem_fops = { |
| 143 | .release = secretmem_release, |
| 144 | .mmap = secretmem_mmap, |
| 145 | }; |
| 146 | |
| 147 | static int secretmem_migrate_folio(struct address_space *mapping, |
| 148 | struct folio *dst, struct folio *src, enum migrate_mode mode) |
| 149 | { |
| 150 | return -EBUSY; |
| 151 | } |
| 152 | |
| 153 | static void secretmem_free_folio(struct folio *folio) |
| 154 | { |
| 155 | set_direct_map_default_noflush(&folio->page); |
| 156 | folio_zero_segment(folio, 0, folio_size(folio)); |
| 157 | } |
| 158 | |
| 159 | const struct address_space_operations secretmem_aops = { |
| 160 | .dirty_folio = noop_dirty_folio, |
| 161 | .free_folio = secretmem_free_folio, |
| 162 | .migrate_folio = secretmem_migrate_folio, |
| 163 | }; |
| 164 | |
| 165 | static int secretmem_setattr(struct user_namespace *mnt_userns, |
| 166 | struct dentry *dentry, struct iattr *iattr) |
| 167 | { |
| 168 | struct inode *inode = d_inode(dentry); |
| 169 | struct address_space *mapping = inode->i_mapping; |
| 170 | unsigned int ia_valid = iattr->ia_valid; |
| 171 | int ret; |
| 172 | |
| 173 | filemap_invalidate_lock(mapping); |
| 174 | |
| 175 | if ((ia_valid & ATTR_SIZE) && inode->i_size) |
| 176 | ret = -EINVAL; |
| 177 | else |
| 178 | ret = simple_setattr(mnt_userns, dentry, iattr); |
| 179 | |
| 180 | filemap_invalidate_unlock(mapping); |
| 181 | |
| 182 | return ret; |
| 183 | } |
| 184 | |
| 185 | static const struct inode_operations secretmem_iops = { |
| 186 | .setattr = secretmem_setattr, |
| 187 | }; |
| 188 | |
| 189 | static struct vfsmount *secretmem_mnt; |
| 190 | |
| 191 | static struct file *secretmem_file_create(unsigned long flags) |
| 192 | { |
| 193 | struct file *file; |
| 194 | struct inode *inode; |
| 195 | const char *anon_name = "[secretmem]"; |
| 196 | const struct qstr qname = QSTR_INIT(anon_name, strlen(anon_name)); |
| 197 | int err; |
| 198 | |
| 199 | inode = alloc_anon_inode(secretmem_mnt->mnt_sb); |
| 200 | if (IS_ERR(inode)) |
| 201 | return ERR_CAST(inode); |
| 202 | |
| 203 | err = security_inode_init_security_anon(inode, &qname, NULL); |
| 204 | if (err) { |
| 205 | file = ERR_PTR(err); |
| 206 | goto err_free_inode; |
| 207 | } |
| 208 | |
| 209 | file = alloc_file_pseudo(inode, secretmem_mnt, "secretmem", |
| 210 | O_RDWR, &secretmem_fops); |
| 211 | if (IS_ERR(file)) |
| 212 | goto err_free_inode; |
| 213 | |
| 214 | mapping_set_gfp_mask(inode->i_mapping, GFP_HIGHUSER); |
| 215 | mapping_set_unevictable(inode->i_mapping); |
| 216 | |
| 217 | inode->i_op = &secretmem_iops; |
| 218 | inode->i_mapping->a_ops = &secretmem_aops; |
| 219 | |
| 220 | /* pretend we are a normal file with zero size */ |
| 221 | inode->i_mode |= S_IFREG; |
| 222 | inode->i_size = 0; |
| 223 | |
| 224 | return file; |
| 225 | |
| 226 | err_free_inode: |
| 227 | iput(inode); |
| 228 | return file; |
| 229 | } |
| 230 | |
| 231 | SYSCALL_DEFINE1(memfd_secret, unsigned int, flags) |
| 232 | { |
| 233 | struct file *file; |
| 234 | int fd, err; |
| 235 | |
| 236 | /* make sure local flags do not confict with global fcntl.h */ |
| 237 | BUILD_BUG_ON(SECRETMEM_FLAGS_MASK & O_CLOEXEC); |
| 238 | |
| 239 | if (!secretmem_enable) |
| 240 | return -ENOSYS; |
| 241 | |
| 242 | if (flags & ~(SECRETMEM_FLAGS_MASK | O_CLOEXEC)) |
| 243 | return -EINVAL; |
| 244 | if (atomic_read(&secretmem_users) < 0) |
| 245 | return -ENFILE; |
| 246 | |
| 247 | fd = get_unused_fd_flags(flags & O_CLOEXEC); |
| 248 | if (fd < 0) |
| 249 | return fd; |
| 250 | |
| 251 | file = secretmem_file_create(flags); |
| 252 | if (IS_ERR(file)) { |
| 253 | err = PTR_ERR(file); |
| 254 | goto err_put_fd; |
| 255 | } |
| 256 | |
| 257 | file->f_flags |= O_LARGEFILE; |
| 258 | |
| 259 | atomic_inc(&secretmem_users); |
| 260 | fd_install(fd, file); |
| 261 | return fd; |
| 262 | |
| 263 | err_put_fd: |
| 264 | put_unused_fd(fd); |
| 265 | return err; |
| 266 | } |
| 267 | |
| 268 | static int secretmem_init_fs_context(struct fs_context *fc) |
| 269 | { |
| 270 | return init_pseudo(fc, SECRETMEM_MAGIC) ? 0 : -ENOMEM; |
| 271 | } |
| 272 | |
| 273 | static struct file_system_type secretmem_fs = { |
| 274 | .name = "secretmem", |
| 275 | .init_fs_context = secretmem_init_fs_context, |
| 276 | .kill_sb = kill_anon_super, |
| 277 | }; |
| 278 | |
| 279 | static int __init secretmem_init(void) |
| 280 | { |
| 281 | if (!secretmem_enable) |
| 282 | return 0; |
| 283 | |
| 284 | secretmem_mnt = kern_mount(&secretmem_fs); |
| 285 | if (IS_ERR(secretmem_mnt)) |
| 286 | return PTR_ERR(secretmem_mnt); |
| 287 | |
| 288 | /* prevent secretmem mappings from ever getting PROT_EXEC */ |
| 289 | secretmem_mnt->mnt_flags |= MNT_NOEXEC; |
| 290 | |
| 291 | return 0; |
| 292 | } |
| 293 | fs_initcall(secretmem_init); |