Hi Coly,
Hmm, after a second thought, this problem sounds cant happen in the discard disabled reason:
because the seq is a random number,
get_random_bytes(&i->seq, sizeof(uint64_t));
So it's not possible to get a same random seq in last invalidated bucket and the new bucket.
But what about the power-cut case?
Yang
发件人："杨东升" <dongsheng.yang@xxxxxxxxxxxx>
发送日期：2020-09-16 14:19:46
收件人：colyli <colyli@xxxxxxx>
抄送人：linux-bcache <linux-bcache@xxxxxxxxxxxxxxx>
主题：Fw:About bcache-check>Resending with no HTML format ... ...
>
>
>Hi Coly and all,
> I found there is an error message in our testing:
>
>
>Sep 27 17:43:00 node-1 kernel: bcache: error on
>c2914b7e-d665-4ec1-80e1-272755de19ef: unsupported bset version at bucket
> 58290, block 0, 40818810 keys, disabling caching
>
>
>I checked the code in bch_btree_node_read_done() around this message:
>
> 214 for (;
> 215 b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
> 216 i = write_block(b)) {
> 217 err = "unsupported bset version";
> 218 if (i->version > BCACHE_BSET_VERSION)
> 219 goto err;
> 220
>The problem is we found the i->seq is what we expected for this btree_node, but the version is not BCACHE_BSET_VERSION (1)
>
>
>
>I think there would be two reasons to cause this messages:
>(1) cache discard is not enabled.
> When we allocate a bucket, if we dont enable discard, there could be some outdated data in this bucket,
>
>and there is possibility that the location of i->seq is equal to what we expected,
>
>but that's really not an bset at all, so we will found version, magic and bset_csum are all unexpected,
>
>currently we will goto err and stop cache_set.
>
>
>(2) power-cut.
> When we are doing btree_node_write, if there is a power-cut happen, we could write a partial btree.
>
>
>But when we meet this kind of problems, we cant use this cache device. There is no tool to recovery from this kind of problem.
>
>I think I can cook a bcache-check in bcache-tools, something like fsck. to check this kind of problem
>
>and allow user to repair it, warning on user force-repaire is risky.
>
>
>
>Please help to point out if there is something I am missing.
>
>
>
>Thanx
>Dongsheng
>
>
>
