On 9/24/2024 4:56 PM, Thorsten Blum wrote:
On 24. Sep 2024, at 21:33, Thorsten Blum <thorsten.blum@xxxxxxxxx> wrote:
On 24. Sep 2024, at 20:05, Tom Talpey <tom@xxxxxxxxxx> wrote:
On 9/24/2024 6:22 AM, Thorsten Blum wrote:
Add the __counted_by_le compiler attribute to the flexible array member
Chunks to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and
CONFIG_FORTIFY_SOURCE.
Read Chunks[0] after checking that ChunkCount is not 0.
Compile-tested only.
Signed-off-by: Thorsten Blum <thorsten.blum@xxxxxxxxx>
---
fs/smb/server/smb2pdu.c | 2 +-
fs/smb/server/smb2pdu.h | 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/fs/smb/server/smb2pdu.c b/fs/smb/server/smb2pdu.c
index 461c4fc682ac..0670bdf3e167 100644
--- a/fs/smb/server/smb2pdu.c
+++ b/fs/smb/server/smb2pdu.c
@@ -7565,7 +7565,6 @@ static int fsctl_copychunk(struct ksmbd_work *work,
ci_rsp->TotalBytesWritten =
cpu_to_le32(ksmbd_server_side_copy_max_total_size());
- chunks = (struct srv_copychunk *)&ci_req->Chunks[0];
chunk_count = le32_to_cpu(ci_req->ChunkCount);
if (chunk_count == 0)
goto out;
@@ -7579,6 +7578,7 @@ static int fsctl_copychunk(struct ksmbd_work *work,
return -EINVAL;
}
+ chunks = (struct srv_copychunk *)&ci_req->Chunks[0];
for (i = 0; i < chunk_count; i++) {
if (le32_to_cpu(chunks[i].Length) == 0 ||
le32_to_cpu(chunks[i].Length) > ksmbd_server_side_copy_max_chunk_size())
diff --git a/fs/smb/server/smb2pdu.h b/fs/smb/server/smb2pdu.h
index 73aff20e22d0..f01121dbf358 100644
--- a/fs/smb/server/smb2pdu.h
+++ b/fs/smb/server/smb2pdu.h
@@ -194,7 +194,7 @@ struct copychunk_ioctl_req {
__le64 ResumeKey[3];
__le32 ChunkCount;
__le32 Reserved;
- __u8 Chunks[]; /* array of srv_copychunk */
+ __u8 Chunks[] __counted_by_le(ChunkCount); /* array of srv_copychunk */
} __packed;
This isn't correct. The u8 is just a raw buffer, copychunk structs are
marshaled into it, and they're 24 bytes each.
Hm, I see.
How does this for-loop work then? It iterates over ci_req->ChunkCount
and expects a srv_copychunk at each ci_req->Chunks[i]?
for (i = 0; i < chunk_count; i++) {
if (le32_to_cpu(chunks[i].Length) == 0 ||
le32_to_cpu(chunks[i].Length) > ksmbd_server_side_copy_max_chunk_size())
break;
total_size_written += le32_to_cpu(chunks[i].Length);
}
Never mind, I just realized that the pointer arithmetic takes the array
offset into account.
A srv_copychunk takes up 24 bytes and therefore 24 __u8[] slots. The
__counted_by annotation is essentially off by a factor of 24.
Right, and the fix is probably obvious, but the copychunk code has
yielded a few surprises in the past, so I'd really like to see any
change tested... Maybe Steve has a ready-to-go section in his set.
Tom.