On Sat, 20 Aug 2022 17:11:05 +0530 Siddh Raman Pant wrote:
Syzkaller has reported a warning in iomap_iter(), which got triggered due to a call to iomap_iter_done() which has: WARN_ON_ONCE(iter->iomap.offset > iter->pos);
The warning was triggered because `pos` was being negative. I was having offset = 0, pos = -2950420705881096192.
This ridiculously negative value smells of an overflow, and sure it is.
The userspace can configure a loop using an ioctl call, wherein a configuration of type loop_config is passed (see lo_ioctl()'s case on line 1550 of drivers/block/loop.c). This proceeds to call loop_configure() which in turn calls loop_set_status_from_info() (see line 1050 of loop.c), passing &config->info which is of type loop_info64*. This function then sets the appropriate values, like the offset.
The problem here is loop_device has lo_offset of type loff_t (see line 52 of loop.c), which is typdef-chained to long long, whereas loop_info64 has lo_offset of type __u64 (see line 56 of include/uapi/linux/loop.h).
The function directly copies offset from info to the device as follows (See line 980 of loop.c): lo->lo_offset = info->lo_offset;
This results in the encountered overflow (in my case, the RHS was 15496323367828455424).
Thus, convert the type definitions in loop_info64 to their signed counterparts in order to match definitions in loop_device, and check for negative value during loop_set_status_from_info().
Bug report: https://syzkaller.appspot.com/bug?id=c620fe14aac810396d3c3edc9ad73848bf69a29... Reported-and-tested-by: syzbot+a8e049cd3abd342936b6@syzkaller.appspotmail.com Cc: stable@vger.kernel.org Signed-off-by: Siddh Raman Pant code@siddh.me>
Unless I am missing any other uses or quirks of UAPI loop_info64, I think this won't introduce regression, since if something is working, it will continue to work. If something does break, then it was relying on overflows, which is anyways an incorrect way to go about.
Also, it seems even the 32-bit compatibility structure uses the signed types (compat_loop_info uses compat_int_t which is s32), so this patch should be fine.
drivers/block/loop.c | 3 +++ include/uapi/linux/loop.h | 12 ++++++------ 2 files changed, 9 insertions(+), 6 deletions(-)
diff --git a/drivers/block/loop.c b/drivers/block/loop.c index e3c0ba93c1a3..4ca20ce3158d 100644 --- a/drivers/block/loop.c +++ b/drivers/block/loop.c @@ -977,6 +977,9 @@ loop_set_status_from_info(struct loop_device *lo, return -EINVAL; }
- if (info->lo_offset lo_sizelimit < 0)
return -EINVAL;- lo->lo_offset = info->lo_offset; lo->lo_sizelimit = info->lo_sizelimit; memcpy(lo->lo_file_name, info->lo_file_name, LO_NAME_SIZE);
diff --git a/include/uapi/linux/loop.h b/include/uapi/linux/loop.h index 6f63527dd2ed..973565f38f9d 100644 --- a/include/uapi/linux/loop.h +++ b/include/uapi/linux/loop.h @@ -53,12 +53,12 @@ struct loop_info64 { __u64 lo_device; /* ioctl r/o */ __u64 lo_inode; /* ioctl r/o */ __u64 lo_rdevice; /* ioctl r/o */
- __u64 lo_offset;
- __u64 lo_sizelimit;/* bytes, 0 == max available */
- __u32 lo_number; /* ioctl r/o */
- __u32 lo_encrypt_type; /* obsolete, ignored */
- __u32 lo_encrypt_key_size; /* ioctl w/o */
- __u32 lo_flags;
- __s64 lo_offset;
- __s64 lo_sizelimit;/* bytes, 0 == max available */
- __s32 lo_number; /* ioctl r/o */
- __s32 lo_encrypt_type; /* obsolete, ignored */
- __s32 lo_encrypt_key_size; /* ioctl w/o */
- __s32 lo_flags; __u8 lo_file_name[LO_NAME_SIZE]; __u8 lo_crypt_name[LO_NAME_SIZE]; __u8 lo_encrypt_key[LO_KEY_SIZE]; /* ioctl w/o */
-- 2.35.1
There has been discussion on syzkaller mailing list: https://groups.google.com/g/syzkaller-bugs/c/bg3ANn_7oJw/m/-MbtBx9cAwAJ
Reproducing the latest reply:
On Sun, 21 Aug 2022 11:59:05 +0530 Christoph Hellwig wrote:
On Thu, Aug 18, 2022 at 08:51:16PM +0530, Siddh Raman Pant wrote:
On Thu, 18 Aug 2022 20:20:02 +0530 Matthew Wilcox wrote:
I don't think changing these from u64 to s64 is the right way to go.
Why do you think so? Is there somnething I overlooked?
I think it won't intorduce regression, since if something is working, it will continue to work. If something does break, then they were relying on overflows, which is anyways an incorrect way to go about.
Well, for example userspace code expecting unsignedness of these types could break. So if we really think changing the types is so much preferred we'd need to audit common userspace first. Because of that I think the version proposed by willy is generally preferred.
Also, it seems even the 32-bit compatibility structure uses signed types.
We should probably fix that as well.
Thus, I will send a v2 once the discussion is resolved.
I had sent this patch because the discussion was stale for 2 days and Matthew seemed to be active on other email threads.
Thanks, Siddh