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
secretmem_active(void)45 bool secretmem_active(void)
46 {
47 return !!atomic_read(&secretmem_users);
48 }
49
secretmem_fault(struct vm_fault * vmf)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
secretmem_release(struct inode * inode,struct file * file)115 static int secretmem_release(struct inode *inode, struct file *file)
116 {
117 atomic_dec(&secretmem_users);
118 return 0;
119 }
120
secretmem_mmap(struct file * file,struct vm_area_struct * vma)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
vma_is_secretmem(struct vm_area_struct * vma)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
secretmem_isolate_page(struct page * page,isolate_mode_t mode)147 static bool secretmem_isolate_page(struct page *page, isolate_mode_t mode)
148 {
149 return false;
150 }
151
secretmem_migratepage(struct address_space * mapping,struct page * newpage,struct page * page,enum migrate_mode mode)152 static int secretmem_migratepage(struct address_space *mapping,
153 struct page *newpage, struct page *page,
154 enum migrate_mode mode)
155 {
156 return -EBUSY;
157 }
158
secretmem_freepage(struct page * page)159 static void secretmem_freepage(struct page *page)
160 {
161 set_direct_map_default_noflush(page);
162 clear_highpage(page);
163 }
164
165 const struct address_space_operations secretmem_aops = {
166 .set_page_dirty = __set_page_dirty_no_writeback,
167 .freepage = secretmem_freepage,
168 .migratepage = secretmem_migratepage,
169 .isolate_page = secretmem_isolate_page,
170 };
171
secretmem_setattr(struct user_namespace * mnt_userns,struct dentry * dentry,struct iattr * iattr)172 static int secretmem_setattr(struct user_namespace *mnt_userns,
173 struct dentry *dentry, struct iattr *iattr)
174 {
175 struct inode *inode = d_inode(dentry);
176 struct address_space *mapping = inode->i_mapping;
177 unsigned int ia_valid = iattr->ia_valid;
178 int ret;
179
180 filemap_invalidate_lock(mapping);
181
182 if ((ia_valid & ATTR_SIZE) && inode->i_size)
183 ret = -EINVAL;
184 else
185 ret = simple_setattr(mnt_userns, dentry, iattr);
186
187 filemap_invalidate_unlock(mapping);
188
189 return ret;
190 }
191
192 static const struct inode_operations secretmem_iops = {
193 .setattr = secretmem_setattr,
194 };
195
196 static struct vfsmount *secretmem_mnt;
197
secretmem_file_create(unsigned long flags)198 static struct file *secretmem_file_create(unsigned long flags)
199 {
200 struct file *file = ERR_PTR(-ENOMEM);
201 struct inode *inode;
202
203 inode = alloc_anon_inode(secretmem_mnt->mnt_sb);
204 if (IS_ERR(inode))
205 return ERR_CAST(inode);
206
207 file = alloc_file_pseudo(inode, secretmem_mnt, "secretmem",
208 O_RDWR, &secretmem_fops);
209 if (IS_ERR(file))
210 goto err_free_inode;
211
212 mapping_set_gfp_mask(inode->i_mapping, GFP_HIGHUSER);
213 mapping_set_unevictable(inode->i_mapping);
214
215 inode->i_op = &secretmem_iops;
216 inode->i_mapping->a_ops = &secretmem_aops;
217
218 /* pretend we are a normal file with zero size */
219 inode->i_mode |= S_IFREG;
220 inode->i_size = 0;
221
222 return file;
223
224 err_free_inode:
225 iput(inode);
226 return file;
227 }
228
SYSCALL_DEFINE1(memfd_secret,unsigned int,flags)229 SYSCALL_DEFINE1(memfd_secret, unsigned int, flags)
230 {
231 struct file *file;
232 int fd, err;
233
234 /* make sure local flags do not confict with global fcntl.h */
235 BUILD_BUG_ON(SECRETMEM_FLAGS_MASK & O_CLOEXEC);
236
237 if (!secretmem_enable)
238 return -ENOSYS;
239
240 if (flags & ~(SECRETMEM_FLAGS_MASK | O_CLOEXEC))
241 return -EINVAL;
242 if (atomic_read(&secretmem_users) < 0)
243 return -ENFILE;
244
245 fd = get_unused_fd_flags(flags & O_CLOEXEC);
246 if (fd < 0)
247 return fd;
248
249 file = secretmem_file_create(flags);
250 if (IS_ERR(file)) {
251 err = PTR_ERR(file);
252 goto err_put_fd;
253 }
254
255 file->f_flags |= O_LARGEFILE;
256
257 atomic_inc(&secretmem_users);
258 fd_install(fd, file);
259 return fd;
260
261 err_put_fd:
262 put_unused_fd(fd);
263 return err;
264 }
265
secretmem_init_fs_context(struct fs_context * fc)266 static int secretmem_init_fs_context(struct fs_context *fc)
267 {
268 return init_pseudo(fc, SECRETMEM_MAGIC) ? 0 : -ENOMEM;
269 }
270
271 static struct file_system_type secretmem_fs = {
272 .name = "secretmem",
273 .init_fs_context = secretmem_init_fs_context,
274 .kill_sb = kill_anon_super,
275 };
276
secretmem_init(void)277 static int secretmem_init(void)
278 {
279 int ret = 0;
280
281 if (!secretmem_enable)
282 return ret;
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 ret;
292 }
293 fs_initcall(secretmem_init);
294