On Fri, Mar 22, 2013 at 5:59 AM, Boaz Harrosh <bharrosh@xxxxxxxxxxx> wrote:
> On 03/21/2013 07:06 PM, Peng Tao wrote:
>> On Thu, Mar 21, 2013 at 10:13:00AM -0400, Trond Myklebust wrote:
>> <snip>
>>> @@ -1458,7 +1479,6 @@ static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
>>> dprintk("pnfs write error = %d\n", hdr->pnfs_error);
>>> if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
>>> PNFS_LAYOUTRET_ON_ERROR) {
>>> - clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
>> Hi Trond and Boaz,
>>
>> If object layout requires layout being committed before returned (as fixed in
>> the 3/3 patch), is it a potential problem to directly return layout here as
>> well? e.g., if one lseg is successfully written and pending layoutcommit,
>> then another lseg of the same file failed read/write, then layout will be
>> returned w/o layoutcommit. For blocklayout, it is a potential data corruption
>> and that's why block layout doesn't set PNFS_LAYOUTRET_ON_ERROR bit. So
>> I'm wondering if object will suffer from the same issue?
>>
>
> Hi Tao
>
> No, not at all. The objects layout has error reported as part of the
> layout_return OPT. With exact devices that failed and why. In fact the
> data should not be "committed" per ce, but a recovery process must be
> preformed because we know that not all data of a stripe was committed
> including parity, and the raid5 check-some is surly wrong. This is why
> there is an error bit in layout_commit OPT to denote that this is not
> a true commit and that there is an Error report on the way, for those
> clients that must always lo_commit before lo_return even on Error.
>
Thanks for explaining. I learned more about objects :)
Cheers,
Tao
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
