On 2019-8-18 10:53, Matthew Wilcox wrote:
> On Sun, Aug 18, 2019 at 10:32:45AM +0800, Gao Xiang wrote:
>> On Sat, Aug 17, 2019 at 07:20:55PM -0700, Matthew Wilcox wrote:
>>> On Sun, Aug 18, 2019 at 09:56:31AM +0800, Gao Xiang wrote:
>>>> @@ -82,8 +82,12 @@ static int erofs_readdir(struct file *f, struct dir_context *ctx)
>>>> unsigned int nameoff, maxsize;
>>>>
>>>> dentry_page = read_mapping_page(mapping, i, NULL);
>>>> - if (IS_ERR(dentry_page))
>>>> - continue;
>>>> + if (IS_ERR(dentry_page)) {
>>>> + errln("fail to readdir of logical block %u of nid %llu",
>>>> + i, EROFS_V(dir)->nid);
>>>> + err = PTR_ERR(dentry_page);
>>>> + break;
>>>
>>> I don't think you want to use the errno that came back from
>>> read_mapping_page() (which is, I think, always going to be -EIO).
>>> Rather you want -EFSCORRUPTED, at least if I understand the recent
>>> patches to ext2/ext4/f2fs/xfs/...
>>
>> Thanks for your reply and noticing this. :)
>>
>> Yes, as I talked with you about read_mapping_page() in a xfs related
>> topic earlier, I think I fully understand what returns here.
>>
>> I actually had some concern about that before sending out this patch.
>> You know the status is
>> PG_uptodate is not set and PG_error is set here.
>>
>> But we cannot know it is actually a disk read error or due to
>> corrupted images (due to lack of page flags or some status, and
>> I think it could be a waste of page structure space for such
>> corrupted image or disk error)...
>>
>> And some people also like propagate errors from insiders...
>> (and they could argue about err = -EFSCORRUPTED as well..)
>>
>> I'd like hear your suggestion about this after my words above?
>> still return -EFSCORRUPTED?
>
> I don't think it matters whether it's due to a disk error or a corrupted
> image. We can't read the directory entry, so we should probably return
> -EFSCORRUPTED. Thinking about it some more, read_mapping_page() can
> also return -ENOMEM, so it should probably look something like this:
>
> err = 0;
> if (dentry_page == ERR_PTR(-ENOMEM))
> err = -ENOMEM;
> else if (IS_ERR(dentry_page)) {
> errln("fail to readdir of logical block %u of nid %llu",
> i, EROFS_V(dir)->nid);
> err = -EFSCORRUPTED;
Well, if there is real IO error happen under filesystem, we should return -EIO
instead of EFSCORRUPTED?
The right fix may be that doing sanity check on on-disk blkaddr, and return
-EFSCORRUPTED if the blkaddr is invalid and propagate the error to its caller
erofs_readdir(), IIUC below error info.
> [36297.354090] attempt to access beyond end of device
> [36297.354098] loop17: rw=0, want=29887428984, limit=1953128
> [36297.354107] attempt to access beyond end of device
> [36297.354109] loop17: rw=0, want=29887428480, limit=1953128
> [36301.827234] attempt to access beyond end of device
> [36301.827243] loop17: rw=0, want=29887428480, limit=1953128
Thanks,
> }
>
> if (err)
> break;
>
