]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
fs-verity: add data verification hooks for ->readpages()
authorEric Biggers <ebiggers@google.com>
Mon, 22 Jul 2019 16:26:22 +0000 (09:26 -0700)
committerEric Biggers <ebiggers@google.com>
Sun, 28 Jul 2019 23:59:16 +0000 (16:59 -0700)
Add functions that verify data pages that have been read from a
fs-verity file, against that file's Merkle tree.  These will be called
from filesystems' ->readpage() and ->readpages() methods.

Since data verification can block, a workqueue is provided for these
methods to enqueue verification work from their bio completion callback.

See the "Verifying data" section of
Documentation/filesystems/fsverity.rst for more information.

Reviewed-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Jaegeuk Kim <jaegeuk@kernel.org>
Signed-off-by: Eric Biggers <ebiggers@google.com>
fs/verity/Makefile
fs/verity/fsverity_private.h
fs/verity/init.c
fs/verity/open.c
fs/verity/verify.c [new file with mode: 0644]
include/linux/fsverity.h

index e6a8951c493a5ef40d0455ae3aa5eb193bc2addb..7fa628cd5eba244115574efa87ef6795bbee0873 100644 (file)
@@ -2,4 +2,5 @@
 
 obj-$(CONFIG_FS_VERITY) += hash_algs.o \
                           init.o \
-                          open.o
+                          open.o \
+                          verify.o
index c79746ff335e1430d897a360c5147ca19a94bb43..eaa2b3b93bbf6b82957a39e55750b70085af6eeb 100644 (file)
@@ -134,5 +134,10 @@ void fsverity_set_info(struct inode *inode, struct fsverity_info *vi);
 void fsverity_free_info(struct fsverity_info *vi);
 
 int __init fsverity_init_info_cache(void);
+void __init fsverity_exit_info_cache(void);
+
+/* verify.c */
+
+int __init fsverity_init_workqueue(void);
 
 #endif /* _FSVERITY_PRIVATE_H */
index fff1fd6343357d96ddba94d01e90f339125729a4..b593805aafcc8931dea04b9a38f2f8ed8a39737e 100644 (file)
@@ -41,7 +41,15 @@ static int __init fsverity_init(void)
        if (err)
                return err;
 
+       err = fsverity_init_workqueue();
+       if (err)
+               goto err_exit_info_cache;
+
        pr_debug("Initialized fs-verity\n");
        return 0;
+
+err_exit_info_cache:
+       fsverity_exit_info_cache();
+       return err;
 }
 late_initcall(fsverity_init)
index 2cb2fe8082bff435c27abb1055c8ee969aab904c..3636a1ed8e2c41e87ee8149f0693f48cf1399f7e 100644 (file)
@@ -337,3 +337,9 @@ int __init fsverity_init_info_cache(void)
                return -ENOMEM;
        return 0;
 }
+
+void __init fsverity_exit_info_cache(void)
+{
+       kmem_cache_destroy(fsverity_info_cachep);
+       fsverity_info_cachep = NULL;
+}
diff --git a/fs/verity/verify.c b/fs/verity/verify.c
new file mode 100644 (file)
index 0000000..62ab8f6
--- /dev/null
@@ -0,0 +1,275 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * fs/verity/verify.c: data verification functions, i.e. hooks for ->readpages()
+ *
+ * Copyright 2019 Google LLC
+ */
+
+#include "fsverity_private.h"
+
+#include <crypto/hash.h>
+#include <linux/bio.h>
+#include <linux/ratelimit.h>
+
+static struct workqueue_struct *fsverity_read_workqueue;
+
+/**
+ * hash_at_level() - compute the location of the block's hash at the given level
+ *
+ * @params:    (in) the Merkle tree parameters
+ * @dindex:    (in) the index of the data block being verified
+ * @level:     (in) the level of hash we want (0 is leaf level)
+ * @hindex:    (out) the index of the hash block containing the wanted hash
+ * @hoffset:   (out) the byte offset to the wanted hash within the hash block
+ */
+static void hash_at_level(const struct merkle_tree_params *params,
+                         pgoff_t dindex, unsigned int level, pgoff_t *hindex,
+                         unsigned int *hoffset)
+{
+       pgoff_t position;
+
+       /* Offset of the hash within the level's region, in hashes */
+       position = dindex >> (level * params->log_arity);
+
+       /* Index of the hash block in the tree overall */
+       *hindex = params->level_start[level] + (position >> params->log_arity);
+
+       /* Offset of the wanted hash (in bytes) within the hash block */
+       *hoffset = (position & ((1 << params->log_arity) - 1)) <<
+                  (params->log_blocksize - params->log_arity);
+}
+
+/* Extract a hash from a hash page */
+static void extract_hash(struct page *hpage, unsigned int hoffset,
+                        unsigned int hsize, u8 *out)
+{
+       void *virt = kmap_atomic(hpage);
+
+       memcpy(out, virt + hoffset, hsize);
+       kunmap_atomic(virt);
+}
+
+static inline int cmp_hashes(const struct fsverity_info *vi,
+                            const u8 *want_hash, const u8 *real_hash,
+                            pgoff_t index, int level)
+{
+       const unsigned int hsize = vi->tree_params.digest_size;
+
+       if (memcmp(want_hash, real_hash, hsize) == 0)
+               return 0;
+
+       fsverity_err(vi->inode,
+                    "FILE CORRUPTED! index=%lu, level=%d, want_hash=%s:%*phN, real_hash=%s:%*phN",
+                    index, level,
+                    vi->tree_params.hash_alg->name, hsize, want_hash,
+                    vi->tree_params.hash_alg->name, hsize, real_hash);
+       return -EBADMSG;
+}
+
+/*
+ * Verify a single data page against the file's Merkle tree.
+ *
+ * In principle, we need to verify the entire path to the root node.  However,
+ * for efficiency the filesystem may cache the hash pages.  Therefore we need
+ * only ascend the tree until an already-verified page is seen, as indicated by
+ * the PageChecked bit being set; then verify the path to that page.
+ *
+ * This code currently only supports the case where the verity block size is
+ * equal to PAGE_SIZE.  Doing otherwise would be possible but tricky, since we
+ * wouldn't be able to use the PageChecked bit.
+ *
+ * Note that multiple processes may race to verify a hash page and mark it
+ * Checked, but it doesn't matter; the result will be the same either way.
+ *
+ * Return: true if the page is valid, else false.
+ */
+static bool verify_page(struct inode *inode, const struct fsverity_info *vi,
+                       struct ahash_request *req, struct page *data_page)
+{
+       const struct merkle_tree_params *params = &vi->tree_params;
+       const unsigned int hsize = params->digest_size;
+       const pgoff_t index = data_page->index;
+       int level;
+       u8 _want_hash[FS_VERITY_MAX_DIGEST_SIZE];
+       const u8 *want_hash;
+       u8 real_hash[FS_VERITY_MAX_DIGEST_SIZE];
+       struct page *hpages[FS_VERITY_MAX_LEVELS];
+       unsigned int hoffsets[FS_VERITY_MAX_LEVELS];
+       int err;
+
+       if (WARN_ON_ONCE(!PageLocked(data_page) || PageUptodate(data_page)))
+               return false;
+
+       pr_debug_ratelimited("Verifying data page %lu...\n", index);
+
+       /*
+        * Starting at the leaf level, ascend the tree saving hash pages along
+        * the way until we find a verified hash page, indicated by PageChecked;
+        * or until we reach the root.
+        */
+       for (level = 0; level < params->num_levels; level++) {
+               pgoff_t hindex;
+               unsigned int hoffset;
+               struct page *hpage;
+
+               hash_at_level(params, index, level, &hindex, &hoffset);
+
+               pr_debug_ratelimited("Level %d: hindex=%lu, hoffset=%u\n",
+                                    level, hindex, hoffset);
+
+               hpage = inode->i_sb->s_vop->read_merkle_tree_page(inode,
+                                                                 hindex);
+               if (IS_ERR(hpage)) {
+                       err = PTR_ERR(hpage);
+                       fsverity_err(inode,
+                                    "Error %d reading Merkle tree page %lu",
+                                    err, hindex);
+                       goto out;
+               }
+
+               if (PageChecked(hpage)) {
+                       extract_hash(hpage, hoffset, hsize, _want_hash);
+                       want_hash = _want_hash;
+                       put_page(hpage);
+                       pr_debug_ratelimited("Hash page already checked, want %s:%*phN\n",
+                                            params->hash_alg->name,
+                                            hsize, want_hash);
+                       goto descend;
+               }
+               pr_debug_ratelimited("Hash page not yet checked\n");
+               hpages[level] = hpage;
+               hoffsets[level] = hoffset;
+       }
+
+       want_hash = vi->root_hash;
+       pr_debug("Want root hash: %s:%*phN\n",
+                params->hash_alg->name, hsize, want_hash);
+descend:
+       /* Descend the tree verifying hash pages */
+       for (; level > 0; level--) {
+               struct page *hpage = hpages[level - 1];
+               unsigned int hoffset = hoffsets[level - 1];
+
+               err = fsverity_hash_page(params, inode, req, hpage, real_hash);
+               if (err)
+                       goto out;
+               err = cmp_hashes(vi, want_hash, real_hash, index, level - 1);
+               if (err)
+                       goto out;
+               SetPageChecked(hpage);
+               extract_hash(hpage, hoffset, hsize, _want_hash);
+               want_hash = _want_hash;
+               put_page(hpage);
+               pr_debug("Verified hash page at level %d, now want %s:%*phN\n",
+                        level - 1, params->hash_alg->name, hsize, want_hash);
+       }
+
+       /* Finally, verify the data page */
+       err = fsverity_hash_page(params, inode, req, data_page, real_hash);
+       if (err)
+               goto out;
+       err = cmp_hashes(vi, want_hash, real_hash, index, -1);
+out:
+       for (; level > 0; level--)
+               put_page(hpages[level - 1]);
+
+       return err == 0;
+}
+
+/**
+ * fsverity_verify_page() - verify a data page
+ *
+ * Verify a page that has just been read from a verity file.  The page must be a
+ * pagecache page that is still locked and not yet uptodate.
+ *
+ * Return: true if the page is valid, else false.
+ */
+bool fsverity_verify_page(struct page *page)
+{
+       struct inode *inode = page->mapping->host;
+       const struct fsverity_info *vi = inode->i_verity_info;
+       struct ahash_request *req;
+       bool valid;
+
+       req = ahash_request_alloc(vi->tree_params.hash_alg->tfm, GFP_NOFS);
+       if (unlikely(!req))
+               return false;
+
+       valid = verify_page(inode, vi, req, page);
+
+       ahash_request_free(req);
+
+       return valid;
+}
+EXPORT_SYMBOL_GPL(fsverity_verify_page);
+
+#ifdef CONFIG_BLOCK
+/**
+ * fsverity_verify_bio() - verify a 'read' bio that has just completed
+ *
+ * Verify a set of pages that have just been read from a verity file.  The pages
+ * must be pagecache pages that are still locked and not yet uptodate.  Pages
+ * that fail verification are set to the Error state.  Verification is skipped
+ * for pages already in the Error state, e.g. due to fscrypt decryption failure.
+ *
+ * This is a helper function for use by the ->readpages() method of filesystems
+ * that issue bios to read data directly into the page cache.  Filesystems that
+ * populate the page cache without issuing bios (e.g. non block-based
+ * filesystems) must instead call fsverity_verify_page() directly on each page.
+ * All filesystems must also call fsverity_verify_page() on holes.
+ */
+void fsverity_verify_bio(struct bio *bio)
+{
+       struct inode *inode = bio_first_page_all(bio)->mapping->host;
+       const struct fsverity_info *vi = inode->i_verity_info;
+       struct ahash_request *req;
+       struct bio_vec *bv;
+       struct bvec_iter_all iter_all;
+
+       req = ahash_request_alloc(vi->tree_params.hash_alg->tfm, GFP_NOFS);
+       if (unlikely(!req)) {
+               bio_for_each_segment_all(bv, bio, iter_all)
+                       SetPageError(bv->bv_page);
+               return;
+       }
+
+       bio_for_each_segment_all(bv, bio, iter_all) {
+               struct page *page = bv->bv_page;
+
+               if (!PageError(page) && !verify_page(inode, vi, req, page))
+                       SetPageError(page);
+       }
+
+       ahash_request_free(req);
+}
+EXPORT_SYMBOL_GPL(fsverity_verify_bio);
+#endif /* CONFIG_BLOCK */
+
+/**
+ * fsverity_enqueue_verify_work() - enqueue work on the fs-verity workqueue
+ *
+ * Enqueue verification work for asynchronous processing.
+ */
+void fsverity_enqueue_verify_work(struct work_struct *work)
+{
+       queue_work(fsverity_read_workqueue, work);
+}
+EXPORT_SYMBOL_GPL(fsverity_enqueue_verify_work);
+
+int __init fsverity_init_workqueue(void)
+{
+       /*
+        * Use an unbound workqueue to allow bios to be verified in parallel
+        * even when they happen to complete on the same CPU.  This sacrifices
+        * locality, but it's worthwhile since hashing is CPU-intensive.
+        *
+        * Also use a high-priority workqueue to prioritize verification work,
+        * which blocks reads from completing, over regular application tasks.
+        */
+       fsverity_read_workqueue = alloc_workqueue("fsverity_read_queue",
+                                                 WQ_UNBOUND | WQ_HIGHPRI,
+                                                 num_online_cpus());
+       if (!fsverity_read_workqueue)
+               return -ENOMEM;
+       return 0;
+}
index cbd0f84e1620f7b4b0d20db9931f86b73203537d..95c257cd7ff02909c54c6e944d0ce057d1f883f6 100644 (file)
@@ -33,6 +33,23 @@ struct fsverity_operations {
         */
        int (*get_verity_descriptor)(struct inode *inode, void *buf,
                                     size_t bufsize);
+
+       /**
+        * Read a Merkle tree page of the given inode.
+        *
+        * @inode: the inode
+        * @index: 0-based index of the page within the Merkle tree
+        *
+        * This can be called at any time on an open verity file, as well as
+        * between ->begin_enable_verity() and ->end_enable_verity().  It may be
+        * called by multiple processes concurrently, even with the same page.
+        *
+        * Note that this must retrieve a *page*, not necessarily a *block*.
+        *
+        * Return: the page on success, ERR_PTR() on failure
+        */
+       struct page *(*read_merkle_tree_page)(struct inode *inode,
+                                             pgoff_t index);
 };
 
 #ifdef CONFIG_FS_VERITY
@@ -49,6 +66,12 @@ extern int fsverity_file_open(struct inode *inode, struct file *filp);
 extern int fsverity_prepare_setattr(struct dentry *dentry, struct iattr *attr);
 extern void fsverity_cleanup_inode(struct inode *inode);
 
+/* verify.c */
+
+extern bool fsverity_verify_page(struct page *page);
+extern void fsverity_verify_bio(struct bio *bio);
+extern void fsverity_enqueue_verify_work(struct work_struct *work);
+
 #else /* !CONFIG_FS_VERITY */
 
 static inline struct fsverity_info *fsverity_get_info(const struct inode *inode)
@@ -73,6 +96,39 @@ static inline void fsverity_cleanup_inode(struct inode *inode)
 {
 }
 
+/* verify.c */
+
+static inline bool fsverity_verify_page(struct page *page)
+{
+       WARN_ON(1);
+       return false;
+}
+
+static inline void fsverity_verify_bio(struct bio *bio)
+{
+       WARN_ON(1);
+}
+
+static inline void fsverity_enqueue_verify_work(struct work_struct *work)
+{
+       WARN_ON(1);
+}
+
 #endif /* !CONFIG_FS_VERITY */
 
+/**
+ * fsverity_active() - do reads from the inode need to go through fs-verity?
+ *
+ * This checks whether ->i_verity_info has been set.
+ *
+ * Filesystems call this from ->readpages() to check whether the pages need to
+ * be verified or not.  Don't use IS_VERITY() for this purpose; it's subject to
+ * a race condition where the file is being read concurrently with
+ * FS_IOC_ENABLE_VERITY completing.  (S_VERITY is set before ->i_verity_info.)
+ */
+static inline bool fsverity_active(const struct inode *inode)
+{
+       return fsverity_get_info(inode) != NULL;
+}
+
 #endif /* _LINUX_FSVERITY_H */