On Wed, Nov 04, 2020 at 10:19:06AM +0800, Chao Yu wrote:
> Eric reported a ioctl bug in below link:
>
> https://lore.kernel.org/linux-f2fs-devel/20201103032234.GB2875@sol.localdomain/
>
> That said, on some 32-bit architectures, u64 has only 32-bit alignment,
> notably i386 and x86_32, so that size of struct f2fs_gc_range compiled
> in x86_32 is 20 bytes, however the size in x86_64 is 24 bytes, binary
> compiled in x86_32 can not call F2FS_IOC_GARBAGE_COLLECT_RANGE successfully
> due to mismatched value of ioctl command in betweeen binary and f2fs
> module, similarly, F2FS_IOC_MOVE_RANGE will fail too.
>
> In this patch we introduce two ioctls for compatibility of above special
> 32-bit binary:
> - F2FS_IOC32_GARBAGE_COLLECT_RANGE
> - F2FS_IOC32_MOVE_RANGE
>
> Signed-off-by: Chao Yu <yuch...@huawei.com>
> ---
>
> Jaegeuk, Eric,
>
> I have no 32-bit machine now, so I don't run any test on this patch,
> please take a look at this RFC patch first.
You can test this by running a 32-bit binary on a machine with a 64-bit kernel.
E.g. on x86_64, compile a binary with 'gcc -m32'.
> #ifdef CONFIG_COMPAT
> +static int f2fs_compat_ioc_gc_range(struct file *file, unsigned long arg)
> +{
> + struct compat_f2fs_gc_range __user *urange;
> + struct f2fs_gc_range range;
> + int err;
> +
> + urange = compat_ptr(arg);
> + err = get_user(range.sync, &urange->sync);
> + err |= get_user(range.start, &urange->start);
> + err |= get_user(range.len, &urange->len);
> + if (err)
> + return -EFAULT;
> + if (unlikely(f2fs_cp_error(F2FS_I_SB(file_inode(file)))))
> + return -EIO;
> + if (!f2fs_is_checkpoint_ready(F2FS_I_SB(file_inode(file))))
> + return -ENOSPC;
> + return f2fs_ioc_gc_range(file, (unsigned long)&range);
> +}
This won't work because f2fs_ioc_gc_range() expects a user pointer. You'll need
to make the native and compat versions do the copy from user separately, and
have them call a helper function that takes a pointer to the argument in kernel
memory.
- Eric
