On Wed, Jul 9, 2025, at 15:48, Naresh Kamboju wrote:
> On Tue, 8 Jul 2025 at 18:28, Naresh Kamboju <naresh.kamboju@xxxxxxxxxx> wrote:
>>
>> Regressions were observed while testing LTP syscalls cachestat01 and
>> other related tests on the next-20250702 Linux kernel across several devices.
>>
>> The issue appears to be related to the inability to configure /dev/loop0
>> via the LOOP_SET_STATUS ioctl, which returned EOPNOTSUPP
>> (Operation not supported). This results in a TBROK condition,
>> causing the test to fail.
>
> Anders, bisected this down to this commit id,
> # first bad commit:
> [9eb22f7fedfc9eb1b7f431a5359abd4d15b0b0cd]
> fs: add ioctl to query metadata and protection info capabilities
I see the problem now in
+ if (_IOC_NR(cmd) == _IOC_NR(FS_IOC_GETLBMD_CAP))
+ return blk_get_meta_cap(bdev, cmd, argp);
+
This only compares _IOC_NR() but not _IOC_TYPE, so LOOP_SET_STATUS
is treated the same as FS_IOC_GETLBMD_CAP, since both use '2' in
the lower 8 bit.
include/uapi/linux/fs.h:#define FS_IOC_GETLBMD_CAP _IOWR(0x15, 2, struct logical_block_metadata_cap)
include/uapi/linux/loop.h:#define LOOP_SET_STATUS 0x4C02
I checked a couple of other drivers using _IOC_NR(), and it seems
that they many of them have the same bug, e.g.:
drivers/accel/habanalabs/common/habanalabs_ioctl.c
drivers/block/ublk_drv.c
drivers/dma-buf/dma-heap.c
Arnd
