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 = 1;
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 folio *folio;
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 folio = filemap_lock_folio(mapping, offset);
68 if (IS_ERR(folio)) {
69 folio = folio_alloc(gfp | __GFP_ZERO, 0);
70 if (!folio) {
71 ret = VM_FAULT_OOM;
72 goto out;
73 }
74
75 err = set_direct_map_invalid_noflush(folio_page(folio, 0));
76 if (err) {
77 folio_put(folio);
78 ret = vmf_error(err);
79 goto out;
80 }
81
82 __folio_mark_uptodate(folio);
83 err = filemap_add_folio(mapping, folio, offset, gfp);
84 if (unlikely(err)) {
85 /*
86 * If a split of large page was required, it
87 * already happened when we marked the page invalid
88 * which guarantees that this call won't fail
89 */
90 set_direct_map_default_noflush(folio_page(folio, 0));
91 folio_put(folio);
92 if (err == -EEXIST)
93 goto retry;
94
95 ret = vmf_error(err);
96 goto out;
97 }
98
99 addr = (unsigned long)folio_address(folio);
100 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
101 }
102
103 vmf->page = folio_file_page(folio, vmf->pgoff);
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_prepare(struct vm_area_desc * desc)121 static int secretmem_mmap_prepare(struct vm_area_desc *desc)
122 {
123 const unsigned long len = vma_desc_size(desc);
124
125 if (!vma_desc_test_flags(desc, VMA_SHARED_BIT, VMA_MAYSHARE_BIT))
126 return -EINVAL;
127
128 vma_desc_set_flags(desc, VMA_LOCKED_BIT, VMA_DONTDUMP_BIT);
129 if (!mlock_future_ok(desc->mm, /*is_vma_locked=*/ true, len))
130 return -EAGAIN;
131 desc->vm_ops = &secretmem_vm_ops;
132
133 return 0;
134 }
135
vma_is_secretmem(struct vm_area_struct * vma)136 bool vma_is_secretmem(struct vm_area_struct *vma)
137 {
138 return vma->vm_ops == &secretmem_vm_ops;
139 }
140
141 static const struct file_operations secretmem_fops = {
142 .release = secretmem_release,
143 .mmap_prepare = secretmem_mmap_prepare,
144 };
145
secretmem_migrate_folio(struct address_space * mapping,struct folio * dst,struct folio * src,enum migrate_mode mode)146 static int secretmem_migrate_folio(struct address_space *mapping,
147 struct folio *dst, struct folio *src, enum migrate_mode mode)
148 {
149 return -EBUSY;
150 }
151
secretmem_free_folio(struct folio * folio)152 static void secretmem_free_folio(struct folio *folio)
153 {
154 set_direct_map_default_noflush(folio_page(folio, 0));
155 folio_zero_segment(folio, 0, folio_size(folio));
156 }
157
158 const struct address_space_operations secretmem_aops = {
159 .dirty_folio = noop_dirty_folio,
160 .free_folio = secretmem_free_folio,
161 .migrate_folio = secretmem_migrate_folio,
162 };
163
secretmem_setattr(struct mnt_idmap * idmap,struct dentry * dentry,struct iattr * iattr)164 static int secretmem_setattr(struct mnt_idmap *idmap,
165 struct dentry *dentry, struct iattr *iattr)
166 {
167 struct inode *inode = d_inode(dentry);
168 struct address_space *mapping = inode->i_mapping;
169 unsigned int ia_valid = iattr->ia_valid;
170 int ret;
171
172 filemap_invalidate_lock(mapping);
173
174 if ((ia_valid & ATTR_SIZE) && inode->i_size)
175 ret = -EINVAL;
176 else
177 ret = simple_setattr(idmap, dentry, iattr);
178
179 filemap_invalidate_unlock(mapping);
180
181 return ret;
182 }
183
184 static const struct inode_operations secretmem_iops = {
185 .setattr = secretmem_setattr,
186 };
187
188 static struct vfsmount *secretmem_mnt;
189
secretmem_file_create(unsigned long flags)190 static struct file *secretmem_file_create(unsigned long flags)
191 {
192 struct file *file;
193 struct inode *inode;
194 const char *anon_name = "[secretmem]";
195
196 inode = anon_inode_make_secure_inode(secretmem_mnt->mnt_sb, anon_name, NULL);
197 if (IS_ERR(inode))
198 return ERR_CAST(inode);
199
200 file = alloc_file_pseudo(inode, secretmem_mnt, "secretmem",
201 O_RDWR | O_LARGEFILE, &secretmem_fops);
202 if (IS_ERR(file))
203 goto err_free_inode;
204
205 mapping_set_gfp_mask(inode->i_mapping, GFP_HIGHUSER);
206 mapping_set_unevictable(inode->i_mapping);
207
208 inode->i_op = &secretmem_iops;
209 inode->i_mapping->a_ops = &secretmem_aops;
210
211 /* pretend we are a normal file with zero size */
212 inode->i_mode |= S_IFREG;
213 inode->i_size = 0;
214
215 atomic_inc(&secretmem_users);
216
217 return file;
218
219 err_free_inode:
220 iput(inode);
221 return file;
222 }
223
SYSCALL_DEFINE1(memfd_secret,unsigned int,flags)224 SYSCALL_DEFINE1(memfd_secret, unsigned int, flags)
225 {
226 /* make sure local flags do not conflict with global fcntl.h */
227 BUILD_BUG_ON(SECRETMEM_FLAGS_MASK & O_CLOEXEC);
228
229 if (!secretmem_enable || !can_set_direct_map())
230 return -ENOSYS;
231
232 if (flags & ~(SECRETMEM_FLAGS_MASK | O_CLOEXEC))
233 return -EINVAL;
234 if (atomic_read(&secretmem_users) < 0)
235 return -ENFILE;
236
237 return FD_ADD(flags & O_CLOEXEC, secretmem_file_create(flags));
238 }
239
secretmem_init_fs_context(struct fs_context * fc)240 static int secretmem_init_fs_context(struct fs_context *fc)
241 {
242 struct pseudo_fs_context *ctx;
243
244 ctx = init_pseudo(fc, SECRETMEM_MAGIC);
245 if (!ctx)
246 return -ENOMEM;
247
248 fc->s_iflags |= SB_I_NOEXEC;
249 fc->s_iflags |= SB_I_NODEV;
250 return 0;
251 }
252
253 static struct file_system_type secretmem_fs = {
254 .name = "secretmem",
255 .init_fs_context = secretmem_init_fs_context,
256 .kill_sb = kill_anon_super,
257 };
258
secretmem_init(void)259 static int __init secretmem_init(void)
260 {
261 if (!secretmem_enable || !can_set_direct_map())
262 return 0;
263
264 secretmem_mnt = kern_mount(&secretmem_fs);
265 if (IS_ERR(secretmem_mnt))
266 return PTR_ERR(secretmem_mnt);
267
268 return 0;
269 }
270 fs_initcall(secretmem_init);
271