Skip to contentSkip to conversation

hfsplus: fix uninit-value in hfsplus_cat_build_record

PATCH In review [email protected] 2 participants 2 messages Nov 20, 2025

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 hfsplus: fix uninit-value in hfsplus_cat_build_record

From: Shaurya Rane <[email protected]>

The root cause is in hfsplus_cat_build_record(), which builds catalog entries using the union hfsplus_cat_entry. This union contains three members with significantly different sizes:

  struct hfsplus_cat_folder folder;    (88 bytes)
  struct hfsplus_cat_file file;        (248 bytes)
  struct hfsplus_cat_thread thread;    (520 bytes)

The function was only zeroing the specific member being used (folder or file), not the entire union. This left significant uninitialized data:

For folders: 520 - 88 = 432 bytes uninitialized
For files: 520 - 248 = 272 bytes uninitialized

This uninitialized data was then written to disk via hfs_brec_insert(), read back through the loop device, and eventually copied to userspace via filemap_read(), resulting in a leak of kernel stack memory. Fix this by zeroing the entire union before initializing the specific member. This ensures no uninitialized bytes remain.

---
Changes in v2:
- Corrected format of Fixes tag
- Removed extra blank line before Signed-off-by

 fs/hfsplus/catalog.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
fs/hfsplus/catalog.c +4 -2
index 02c1eee4a4b8..4d42e7139f3b 100644
@@ -111,7 +111,8 @@ static int hfsplus_cat_build_record(hfsplus_cat_entry *entry,
 		struct hfsplus_cat_folder *folder;
 
 		folder = &entry->folder;
-		memset(folder, 0, sizeof(*folder));
+		/* Zero the entire union to avoid leaking uninitialized data */
+		memset(entry, 0, sizeof(*entry));
 		folder->type = cpu_to_be16(HFSPLUS_FOLDER);
 		if (test_bit(HFSPLUS_SB_HFSX, &sbi->flags))
 			folder->flags |= cpu_to_be16(HFSPLUS_HAS_FOLDER_COUNT);
@@ -130,7 +131,8 @@ static int hfsplus_cat_build_record(hfsplus_cat_entry *entry,
 		struct hfsplus_cat_file *file;
 
 		file = &entry->file;
-		memset(file, 0, sizeof(*file));
+		/* Zero the entire union to avoid leaking uninitialized data */
+		memset(entry, 0, sizeof(*entry));
 		file->type = cpu_to_be16(HFSPLUS_FILE);
 		file->flags = cpu_to_be16(HFSPLUS_FILE_THREAD_EXISTS);
 		file->id = cpu_to_be32(cnid);
Signature · 1 line
2.34.1
#1 lore raw
Quoted 3 lines · On Fri, 2025-11-21 at 00:16 +0530, [email protected] wrote:

From: Shaurya Rane <[email protected]>

The root cause is in hfsplus_cat_build_record(), which builds catalog

Commit message sounds slightly confusing. The root cause of what?

Quoted 7 lines

entries using the union hfsplus_cat_entry. This union contains three members with significantly different sizes:

  struct hfsplus_cat_folder folder;    (88 bytes)
  struct hfsplus_cat_file file;        (248 bytes)
  struct hfsplus_cat_thread thread;    (520 bytes)
Quoted 7 lines

The function was only zeroing the specific member being used (folder or file), not the entire union. This left significant uninitialized data:

For folders: 520 - 88 = 432 bytes uninitialized
For files: 520 - 248 = 272 bytes uninitialized

This uninitialized data was then written to disk via hfs_brec_insert(),

I like the zeroing of whole union. But hfs_brec_insert() operates by key_len and entry_len:

hfs_bnode_write(node, fd->search_key, data_off, key_len); hfs_bnode_write(node, entry, data_off + key_len, entry_len);

This values should prevent from writing something out of size folder, file or thread record. Even if we are zeroing the entire union, then we could write something that not fit to the size of folder, file or thread record. It sounds to me that we will corrupt the record anyway because it sounds that key_len and entry_len are incorrect. And if they are incorrect, then we have much bigger issue here.

So, I am not completely sure that it is complete fix of the issue. Could you please justify that my point is incorrect here?

It will be good to be sure that FSCK tool is happy. However, if volume is corrupted intentionally, then it's hard to use the FSCK tool.

Thanks,
Slava.

Quoted 40 lines

read back through the loop device, and eventually copied to userspace via filemap_read(), resulting in a leak of kernel stack memory. Fix this by zeroing the entire union before initializing the specific member. This ensures no uninitialized bytes remain.

---
Changes in v2:
- Corrected format of Fixes tag
- Removed extra blank line before Signed-off-by

 fs/hfsplus/catalog.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
fs/hfsplus/catalog.c +4 -2
index 02c1eee4a4b8..4d42e7139f3b 100644
@@ -111,7 +111,8 @@ static int hfsplus_cat_build_record(hfsplus_cat_entry *entry,
 		struct hfsplus_cat_folder *folder;
 
 		folder = &entry->folder;
-		memset(folder, 0, sizeof(*folder));
+		/* Zero the entire union to avoid leaking uninitialized data */
+		memset(entry, 0, sizeof(*entry));
 		folder->type = cpu_to_be16(HFSPLUS_FOLDER);
 		if (test_bit(HFSPLUS_SB_HFSX, &sbi->flags))
 			folder->flags |= cpu_to_be16(HFSPLUS_HAS_FOLDER_COUNT);
@@ -130,7 +131,8 @@ static int hfsplus_cat_build_record(hfsplus_cat_entry *entry,
 		struct hfsplus_cat_file *file;
 
 		file = &entry->file;
-		memset(file, 0, sizeof(*file));
+		/* Zero the entire union to avoid leaking uninitialized data */
+		memset(entry, 0, sizeof(*entry));
 		file->type = cpu_to_be16(HFSPLUS_FILE);
 		file->flags = cpu_to_be16(HFSPLUS_FILE_THREAD_EXISTS);
 		file->id = cpu_to_be32(cnid);