On 14 Jan 2021, at 21:32, Pierre Rogier <progier(a)redhat.com>
wrote:
Hi William,
> It's a scenario we will need to fix via your BE work because of the MVCC
transaction model that
> LMDB will force us to adopt :)
As I see things in the early phases the lmdb read txn will probably only be managed at
the db plugin level rather than at backend level. That means that we will have the same
inconsistency risk than today (i.e as if using bdb and the implicit txn).
The txn model redesign you are speaking about should only occur in one of the last phases
(once bdb does no more coexists with lmdb).
It must be done because it could provide a serious performance boost for read operations
(IMHO, In most cases we could avoid to duplicate the db data)
But we should not do it while bdb is still around because of the risk of lock issue and
excessive retries.
Yep, agreed. It will be needed for a large read performance boost, but just to prevent
exactly this kind of issue. We should be able to move to a model where everything is
always within a transaction.
We could introduce it earlier and have the read txns be a no-op for bdb and continue using
the implied transactions that we currently have, but also perhaps there is then no benefit
to doing this earlier :)
Note I put a phasing section in
https://directory.fedoraproject.org/docs/389ds/design/backend-redesign-ph...
explaining that. But I guess I should move it within Ludwig's document that englobs
it.
Pierre
On Thu, Jan 14, 2021 at 12:01 AM William Brown <wbrown(a)suse.de> wrote:
> On 13 Jan 2021, at 21:24, Pierre Rogier <progier(a)redhat.com> wrote:
>
> Thank you Willian,
> So far your scenario (entry found when reading base entry but no more existing when
computing the candidates) is the only one that matches the symptoms.
It's a scenario we will need to fix via your BE work because of the MVCC transaction
model that LMDB will force us to adopt :)
> And that triggered a thought:
> We cannot do anything for SUBTREE and ONE_LEVEL searches
> because the fact that the base entry id is not in the candidate may be normal
> but IMHO we should improve the BASE search case.
> In this case the candidate list is directly set to the base entry id
> ==> if the candidate entry (in ldbm_back_next_search_entry) is not found and the
scope is BASE then we should return a LDAP_NO_SUCH_ENTRY error ..
I suspect that Mark has seen this email and submitted a PR to resolve this exact case :)
>
> Pierre
>
>
> On Wed, Jan 13, 2021 at 1:45 AM William Brown <wbrown(a)suse.de> wrote:
> Hey there,
>
>
https://github.com/389ds/389-ds-base/pull/4525/files
>
> I had a look and I can see a few possible contributing factors, but without a core
and the exact state I can't be sure if this is correct. It's all just hypothetical
from reading the code.
>
>
> The crash is in deref_do_deref_attr() which is called as part of deref_pre_entry().
This is the SLAPI_PLUGIN_PRE_ENTRY_FN which is called by
"./ldap/servers/slapd/result.c:1488: rc = plugin_call_plugins(pb,
SLAPI_PLUGIN_PRE_ENTRY_FN);"
>
>
> I think what's important here is that the search is conducted in
./ldap/servers/slapd/opshared.c:818 rc = (*be->be_search)(pb); Is *not* in a
transaction. That means that while the single search in be_search() is consistent due to
an implied transaction, the subsequent search in deref_pre_entry() is likely conducted in
a seperate transaction. This allows for other operations to potentially interleave and
cause changes - modrdn or delete would certainly be candidates to cause a DN to be remove
between these two points. It would be extremely hard to reproduce as a race condition of
course.
>
>
> A question you asked is why don't we get a "no such entry" error or
similar? I think that this is because build_candidate_list in ldbm_search.c doesn't
actually create an error if the base_candidates list is empty, because an IDL is allocated
with a value of 0 (no matching entries). this allows the search to proceed, and there are
no errors, and the result set is set to NULL with size 0. I can't see where
LDAP_NO_SUCH_OBJECT is set in this process, but without looking further into it, my
suspicion is that entries of size 0 WONT return an error condition to internal_search_pb,
so it's valid for this to be empty.
>
> Anyway, again, this is just reading the code for 20 minutes, and is not a complete
in depth investigation, but maybe it's some ideas about what happened?
>
> Hope it helps :)
>
>
>
> —
> Sincerely,
>
> William Brown
>
> Senior Software Engineer, 389 Directory Server
> SUSE Labs, Australia
> _______________________________________________
> 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
> To unsubscribe send an email to 389-devel-leave(a)lists.fedoraproject.org
> Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
> List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
> List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
>
>
> --
> --
>
> 389 Directory Server Development Team
> _______________________________________________
> 389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
> To unsubscribe send an email to 389-devel-leave(a)lists.fedoraproject.org
> Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
> List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
> List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
—
Sincerely,
William Brown
Senior Software Engineer, 389 Directory Server
SUSE Labs, Australia
_______________________________________________
389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
To unsubscribe send an email to 389-devel-leave(a)lists.fedoraproject.org
Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
--
--
389 Directory Server Development Team
_______________________________________________
389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
To unsubscribe send an email to 389-devel-leave(a)lists.fedoraproject.org
Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
List Guidelines:
https://fedoraproject.org/wiki/Mailing_list_guidelines
List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
—
Sincerely,
William Brown
Senior Software Engineer, 389 Directory Server
SUSE Labs, Australia