[sheepdog] question about the epoch field in reading peer request header

[Hitoshi Mitake]

At Sat, 30 May 2015 01:12:12 +0800,
=?ISO-8859-1?B?QmluZ3BlbmcgWmh1?= wrote:
> 
> [1  <multipart/alternative (8bit)>]
> [1.1  <text/plain; ISO-8859-1 (base64)>]
> Hi all:
>     I have a problem about the epoch information in SD_OP_READ_PEER request header. I'm not sure whether I misunderstand the code or it is a bug.
>     When we recover a erasure code object in recovery, we need to read the remaining replicas firstly to rebuild the lost replica. In function read_erasure_object(), we init  SD_OP_READ_PEER request header by the following code:
> 
> 
> 	sd_init_req(&hdr, SD_OP_READ_PEER);
> 	hdr.epoch = epoch;
> 	hdr.flags = SD_FLAG_CMD_RECOVERY;
> 	hdr.data_length = rlen;
> 	hdr.obj.oid = oid;
> 	hdr.obj.tgt_epoch = tgt_epoch;
> 	hdr.obj.ec_index = idx;
> 
> 
> 
>     I think hdr.epoch is current epoch of the cluster and hdr.obj.tgt_epoch is the historical epoch from which we want to read the stale replica. The target node will call peer_read_obj() to process SD_OP_READ_PEER request. Peer_read_obj() set  iocb.epoch = hdr->epoch then pass iocb to sd_store->read(). In default_read(), we use iocb->epoch < sys_epoch() to  judge whether the request is againt the older epoch which needs to read replica from the stale directory. I think we use the wrong epoch here. We should use hdr.obj.tgt_epoch rather than hdr.epoch to make the judgement. Can anyone answer my question?

Hi Bingpeng,

Thanks for your pointing! Your argument seems to be right. We need to
use tgt_epoch. The problem seems to be shared with recovyer process of
ordinal replicated object.

Could you send a patch for solving this problem?

Thanks,
Hitoshi

> 
> 
> Thanks.
> Bingpeng
> [1.2  <text/html; ISO-8859-1 (base64)>]
> 
> [2  <text/plain; us-ascii (7bit)>]
> -- 
> sheepdog mailing list
> sheepdog at lists.wpkg.org
> https://lists.wpkg.org/mailman/listinfo/sheepdog