xref: /openbmc/linux/fs/verity/enable.c (revision 1edd0337)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Ioctl to enable verity on a file
4  *
5  * Copyright 2019 Google LLC
6  */
7 
8 #include "fsverity_private.h"
9 
10 #include <linux/mount.h>
11 #include <linux/sched/signal.h>
12 #include <linux/uaccess.h>
13 
14 struct block_buffer {
15 	u32 filled;
16 	u8 *data;
17 };
18 
19 /* Hash a block, writing the result to the next level's pending block buffer. */
20 static int hash_one_block(struct inode *inode,
21 			  const struct merkle_tree_params *params,
22 			  struct ahash_request *req, struct block_buffer *cur)
23 {
24 	struct block_buffer *next = cur + 1;
25 	int err;
26 
27 	/* Zero-pad the block if it's shorter than the block size. */
28 	memset(&cur->data[cur->filled], 0, params->block_size - cur->filled);
29 
30 	err = fsverity_hash_block(params, inode, req, virt_to_page(cur->data),
31 				  offset_in_page(cur->data),
32 				  &next->data[next->filled]);
33 	if (err)
34 		return err;
35 	next->filled += params->digest_size;
36 	cur->filled = 0;
37 	return 0;
38 }
39 
40 static int write_merkle_tree_block(struct inode *inode, const u8 *buf,
41 				   unsigned long index,
42 				   const struct merkle_tree_params *params)
43 {
44 	u64 pos = (u64)index << params->log_blocksize;
45 	int err;
46 
47 	err = inode->i_sb->s_vop->write_merkle_tree_block(inode, buf, pos,
48 							  params->block_size);
49 	if (err)
50 		fsverity_err(inode, "Error %d writing Merkle tree block %lu",
51 			     err, index);
52 	return err;
53 }
54 
55 /*
56  * Build the Merkle tree for the given file using the given parameters, and
57  * return the root hash in @root_hash.
58  *
59  * The tree is written to a filesystem-specific location as determined by the
60  * ->write_merkle_tree_block() method.  However, the blocks that comprise the
61  * tree are the same for all filesystems.
62  */
63 static int build_merkle_tree(struct file *filp,
64 			     const struct merkle_tree_params *params,
65 			     u8 *root_hash)
66 {
67 	struct inode *inode = file_inode(filp);
68 	const u64 data_size = inode->i_size;
69 	const int num_levels = params->num_levels;
70 	struct ahash_request *req;
71 	struct block_buffer _buffers[1 + FS_VERITY_MAX_LEVELS + 1] = {};
72 	struct block_buffer *buffers = &_buffers[1];
73 	unsigned long level_offset[FS_VERITY_MAX_LEVELS];
74 	int level;
75 	u64 offset;
76 	int err;
77 
78 	if (data_size == 0) {
79 		/* Empty file is a special case; root hash is all 0's */
80 		memset(root_hash, 0, params->digest_size);
81 		return 0;
82 	}
83 
84 	/* This allocation never fails, since it's mempool-backed. */
85 	req = fsverity_alloc_hash_request(params->hash_alg, GFP_KERNEL);
86 
87 	/*
88 	 * Allocate the block buffers.  Buffer "-1" is for data blocks.
89 	 * Buffers 0 <= level < num_levels are for the actual tree levels.
90 	 * Buffer 'num_levels' is for the root hash.
91 	 */
92 	for (level = -1; level < num_levels; level++) {
93 		buffers[level].data = kzalloc(params->block_size, GFP_KERNEL);
94 		if (!buffers[level].data) {
95 			err = -ENOMEM;
96 			goto out;
97 		}
98 	}
99 	buffers[num_levels].data = root_hash;
100 
101 	BUILD_BUG_ON(sizeof(level_offset) != sizeof(params->level_start));
102 	memcpy(level_offset, params->level_start, sizeof(level_offset));
103 
104 	/* Hash each data block, also hashing the tree blocks as they fill up */
105 	for (offset = 0; offset < data_size; offset += params->block_size) {
106 		ssize_t bytes_read;
107 		loff_t pos = offset;
108 
109 		buffers[-1].filled = min_t(u64, params->block_size,
110 					   data_size - offset);
111 		bytes_read = __kernel_read(filp, buffers[-1].data,
112 					   buffers[-1].filled, &pos);
113 		if (bytes_read < 0) {
114 			err = bytes_read;
115 			fsverity_err(inode, "Error %d reading file data", err);
116 			goto out;
117 		}
118 		if (bytes_read != buffers[-1].filled) {
119 			err = -EINVAL;
120 			fsverity_err(inode, "Short read of file data");
121 			goto out;
122 		}
123 		err = hash_one_block(inode, params, req, &buffers[-1]);
124 		if (err)
125 			goto out;
126 		for (level = 0; level < num_levels; level++) {
127 			if (buffers[level].filled + params->digest_size <=
128 			    params->block_size) {
129 				/* Next block at @level isn't full yet */
130 				break;
131 			}
132 			/* Next block at @level is full */
133 
134 			err = hash_one_block(inode, params, req,
135 					     &buffers[level]);
136 			if (err)
137 				goto out;
138 			err = write_merkle_tree_block(inode,
139 						      buffers[level].data,
140 						      level_offset[level],
141 						      params);
142 			if (err)
143 				goto out;
144 			level_offset[level]++;
145 		}
146 		if (fatal_signal_pending(current)) {
147 			err = -EINTR;
148 			goto out;
149 		}
150 		cond_resched();
151 	}
152 	/* Finish all nonempty pending tree blocks. */
153 	for (level = 0; level < num_levels; level++) {
154 		if (buffers[level].filled != 0) {
155 			err = hash_one_block(inode, params, req,
156 					     &buffers[level]);
157 			if (err)
158 				goto out;
159 			err = write_merkle_tree_block(inode,
160 						      buffers[level].data,
161 						      level_offset[level],
162 						      params);
163 			if (err)
164 				goto out;
165 		}
166 	}
167 	/* The root hash was filled by the last call to hash_one_block(). */
168 	if (WARN_ON(buffers[num_levels].filled != params->digest_size)) {
169 		err = -EINVAL;
170 		goto out;
171 	}
172 	err = 0;
173 out:
174 	for (level = -1; level < num_levels; level++)
175 		kfree(buffers[level].data);
176 	fsverity_free_hash_request(params->hash_alg, req);
177 	return err;
178 }
179 
180 static int enable_verity(struct file *filp,
181 			 const struct fsverity_enable_arg *arg)
182 {
183 	struct inode *inode = file_inode(filp);
184 	const struct fsverity_operations *vops = inode->i_sb->s_vop;
185 	struct merkle_tree_params params = { };
186 	struct fsverity_descriptor *desc;
187 	size_t desc_size = struct_size(desc, signature, arg->sig_size);
188 	struct fsverity_info *vi;
189 	int err;
190 
191 	/* Start initializing the fsverity_descriptor */
192 	desc = kzalloc(desc_size, GFP_KERNEL);
193 	if (!desc)
194 		return -ENOMEM;
195 	desc->version = 1;
196 	desc->hash_algorithm = arg->hash_algorithm;
197 	desc->log_blocksize = ilog2(arg->block_size);
198 
199 	/* Get the salt if the user provided one */
200 	if (arg->salt_size &&
201 	    copy_from_user(desc->salt, u64_to_user_ptr(arg->salt_ptr),
202 			   arg->salt_size)) {
203 		err = -EFAULT;
204 		goto out;
205 	}
206 	desc->salt_size = arg->salt_size;
207 
208 	/* Get the signature if the user provided one */
209 	if (arg->sig_size &&
210 	    copy_from_user(desc->signature, u64_to_user_ptr(arg->sig_ptr),
211 			   arg->sig_size)) {
212 		err = -EFAULT;
213 		goto out;
214 	}
215 	desc->sig_size = cpu_to_le32(arg->sig_size);
216 
217 	desc->data_size = cpu_to_le64(inode->i_size);
218 
219 	/* Prepare the Merkle tree parameters */
220 	err = fsverity_init_merkle_tree_params(&params, inode,
221 					       arg->hash_algorithm,
222 					       desc->log_blocksize,
223 					       desc->salt, desc->salt_size);
224 	if (err)
225 		goto out;
226 
227 	/*
228 	 * Start enabling verity on this file, serialized by the inode lock.
229 	 * Fail if verity is already enabled or is already being enabled.
230 	 */
231 	inode_lock(inode);
232 	if (IS_VERITY(inode))
233 		err = -EEXIST;
234 	else
235 		err = vops->begin_enable_verity(filp);
236 	inode_unlock(inode);
237 	if (err)
238 		goto out;
239 
240 	/*
241 	 * Build the Merkle tree.  Don't hold the inode lock during this, since
242 	 * on huge files this may take a very long time and we don't want to
243 	 * force unrelated syscalls like chown() to block forever.  We don't
244 	 * need the inode lock here because deny_write_access() already prevents
245 	 * the file from being written to or truncated, and we still serialize
246 	 * ->begin_enable_verity() and ->end_enable_verity() using the inode
247 	 * lock and only allow one process to be here at a time on a given file.
248 	 */
249 	BUILD_BUG_ON(sizeof(desc->root_hash) < FS_VERITY_MAX_DIGEST_SIZE);
250 	err = build_merkle_tree(filp, &params, desc->root_hash);
251 	if (err) {
252 		fsverity_err(inode, "Error %d building Merkle tree", err);
253 		goto rollback;
254 	}
255 
256 	/*
257 	 * Create the fsverity_info.  Don't bother trying to save work by
258 	 * reusing the merkle_tree_params from above.  Instead, just create the
259 	 * fsverity_info from the fsverity_descriptor as if it were just loaded
260 	 * from disk.  This is simpler, and it serves as an extra check that the
261 	 * metadata we're writing is valid before actually enabling verity.
262 	 */
263 	vi = fsverity_create_info(inode, desc);
264 	if (IS_ERR(vi)) {
265 		err = PTR_ERR(vi);
266 		goto rollback;
267 	}
268 
269 	/*
270 	 * Tell the filesystem to finish enabling verity on the file.
271 	 * Serialized with ->begin_enable_verity() by the inode lock.
272 	 */
273 	inode_lock(inode);
274 	err = vops->end_enable_verity(filp, desc, desc_size, params.tree_size);
275 	inode_unlock(inode);
276 	if (err) {
277 		fsverity_err(inode, "%ps() failed with err %d",
278 			     vops->end_enable_verity, err);
279 		fsverity_free_info(vi);
280 	} else if (WARN_ON(!IS_VERITY(inode))) {
281 		err = -EINVAL;
282 		fsverity_free_info(vi);
283 	} else {
284 		/* Successfully enabled verity */
285 
286 		/*
287 		 * Readers can start using ->i_verity_info immediately, so it
288 		 * can't be rolled back once set.  So don't set it until just
289 		 * after the filesystem has successfully enabled verity.
290 		 */
291 		fsverity_set_info(inode, vi);
292 	}
293 out:
294 	kfree(params.hashstate);
295 	kfree(desc);
296 	return err;
297 
298 rollback:
299 	inode_lock(inode);
300 	(void)vops->end_enable_verity(filp, NULL, 0, params.tree_size);
301 	inode_unlock(inode);
302 	goto out;
303 }
304 
305 /**
306  * fsverity_ioctl_enable() - enable verity on a file
307  * @filp: file to enable verity on
308  * @uarg: user pointer to fsverity_enable_arg
309  *
310  * Enable fs-verity on a file.  See the "FS_IOC_ENABLE_VERITY" section of
311  * Documentation/filesystems/fsverity.rst for the documentation.
312  *
313  * Return: 0 on success, -errno on failure
314  */
315 int fsverity_ioctl_enable(struct file *filp, const void __user *uarg)
316 {
317 	struct inode *inode = file_inode(filp);
318 	struct fsverity_enable_arg arg;
319 	int err;
320 
321 	if (copy_from_user(&arg, uarg, sizeof(arg)))
322 		return -EFAULT;
323 
324 	if (arg.version != 1)
325 		return -EINVAL;
326 
327 	if (arg.__reserved1 ||
328 	    memchr_inv(arg.__reserved2, 0, sizeof(arg.__reserved2)))
329 		return -EINVAL;
330 
331 	if (!is_power_of_2(arg.block_size))
332 		return -EINVAL;
333 
334 	if (arg.salt_size > sizeof_field(struct fsverity_descriptor, salt))
335 		return -EMSGSIZE;
336 
337 	if (arg.sig_size > FS_VERITY_MAX_SIGNATURE_SIZE)
338 		return -EMSGSIZE;
339 
340 	/*
341 	 * Require a regular file with write access.  But the actual fd must
342 	 * still be readonly so that we can lock out all writers.  This is
343 	 * needed to guarantee that no writable fds exist to the file once it
344 	 * has verity enabled, and to stabilize the data being hashed.
345 	 */
346 
347 	err = file_permission(filp, MAY_WRITE);
348 	if (err)
349 		return err;
350 
351 	if (IS_APPEND(inode))
352 		return -EPERM;
353 
354 	if (S_ISDIR(inode->i_mode))
355 		return -EISDIR;
356 
357 	if (!S_ISREG(inode->i_mode))
358 		return -EINVAL;
359 
360 	err = mnt_want_write_file(filp);
361 	if (err) /* -EROFS */
362 		return err;
363 
364 	err = deny_write_access(filp);
365 	if (err) /* -ETXTBSY */
366 		goto out_drop_write;
367 
368 	err = enable_verity(filp, &arg);
369 
370 	/*
371 	 * We no longer drop the inode's pagecache after enabling verity.  This
372 	 * used to be done to try to avoid a race condition where pages could be
373 	 * evicted after being used in the Merkle tree construction, then
374 	 * re-instantiated by a concurrent read.  Such pages are unverified, and
375 	 * the backing storage could have filled them with different content, so
376 	 * they shouldn't be used to fulfill reads once verity is enabled.
377 	 *
378 	 * But, dropping the pagecache has a big performance impact, and it
379 	 * doesn't fully solve the race condition anyway.  So for those reasons,
380 	 * and also because this race condition isn't very important relatively
381 	 * speaking (especially for small-ish files, where the chance of a page
382 	 * being used, evicted, *and* re-instantiated all while enabling verity
383 	 * is quite small), we no longer drop the inode's pagecache.
384 	 */
385 
386 	/*
387 	 * allow_write_access() is needed to pair with deny_write_access().
388 	 * Regardless, the filesystem won't allow writing to verity files.
389 	 */
390 	allow_write_access(filp);
391 out_drop_write:
392 	mnt_drop_write_file(filp);
393 	return err;
394 }
395 EXPORT_SYMBOL_GPL(fsverity_ioctl_enable);
396