openldap-bugs

openldap-bugs@openldap.org

1 participants
20965 discussions

Re: (ITS#5442) slapd_rq not locked before use bugfix
by hyc＠symas.com 30 Mar '08

30 Mar '08

rein(a)tollevik.no wrote: > On Sat, 29 Mar 2008, ando(a)sys-net.it wrote: >> rein(a)basefarm.no wrote: >>> I was seeing random failures of the test050-syncrepl-multimaster test. One of >>> the failures was that it went into a tight loop traversing a circular runqueue >>> it had managed to create in slapd_rq.task_list. It seems as this was caused by >>> missing mutex locks around accesses to slapd_rq, which the patch uploaded to >>> ftp://ftp.openldap.org/incoming/slapd_rq_lock.patch fixes. >>> >>> Before I applied this patch the test failed after being run a few times, with it >>> it has now passed 100 times and is still counting. >> locks in back-bdb/config.c should be pointless, as modifications to the >> configuration should only occur while all threads are paused. The rest >> makes sort of sense, but I'd leave it to Howard. Ignoring the ITS#5403 changes, I don't see anything here that isn't config-related, therefore it's all running single-threaded. Of the "relevant" changes in syncrepl.c, I note that three out of the four chunks of the patch are in code that is only run when using cn=config to delete a syncrepl configuration, and test050 never performs that operation. The remaining chunk only takes affect when adding syncrepl config, and again, slapd is single-threaded for that. I've also run test050 thru hundreds of iterations without any issue, without these patches. If there's a problem in test050, I don't believe it's in this code. -- -- Howard Chu Chief Architect, Symas Corp. http://www.symas.com Director, Highland Sun http://highlandsun.com/hyc/ Chief Architect, OpenLDAP http://www.openldap.org/project/

1 0

Re: (ITS#5442) slapd_rq not locked before use bugfix
by rein＠tollevik.no 29 Mar '08

29 Mar '08

On Sat, 29 Mar 2008, ando(a)sys-net.it wrote: > rein(a)basefarm.no wrote: > >> I was seeing random failures of the test050-syncrepl-multimaster test. One of >> the failures was that it went into a tight loop traversing a circular runqueue >> it had managed to create in slapd_rq.task_list. It seems as this was caused by >> missing mutex locks around accesses to slapd_rq, which the patch uploaded to >> ftp://ftp.openldap.org/incoming/slapd_rq_lock.patch fixes. >> >> Before I applied this patch the test failed after being run a few times, with it >> it has now passed 100 times and is still counting. > > locks in back-bdb/config.c should be pointless, as modifications to the > configuration should only occur while all threads are paused. The rest > makes sort of sense, but I'd leave it to Howard. That is probably true, it looks as if the places in config.c where locks really are required already had them. My patch adds locks everywhere slapd_rq was used without them, as I don't have enough knowledge of the code to know which functions guarranteed to only be used when threads are not running. I believe that the important patch is to syncrepl.c, but I found I it best to add locks everywhere just to be on the safe side. Looking at my copy of the patch it appears that another syncrepl.c patch which I was sure I had edited out has slipped through anyhow :-(. It is the first two modifications related to ldap_get_option, please disregard them in this bugreport. They have been reported in ITS#5403. Rein

1 0

Re: (ITS#5442) slapd_rq not locked before use bugfix
by ando＠sys-net.it 29 Mar '08

29 Mar '08

rein(a)basefarm.no wrote: > I was seeing random failures of the test050-syncrepl-multimaster test. One of > the failures was that it went into a tight loop traversing a circular runqueue > it had managed to create in slapd_rq.task_list. It seems as this was caused by > missing mutex locks around accesses to slapd_rq, which the patch uploaded to > ftp://ftp.openldap.org/incoming/slapd_rq_lock.patch fixes. > > Before I applied this patch the test failed after being run a few times, with it > it has now passed 100 times and is still counting. locks in back-bdb/config.c should be pointless, as modifications to the configuration should only occur while all threads are paused. The rest makes sort of sense, but I'd leave it to Howard. p. Ing. Pierangelo Masarati OpenLDAP Core Team SysNet s.r.l. via Dossi, 8 - 27100 Pavia - ITALIA http://www.sys-net.it --------------------------------------- Office: +39 02 23998309 Mobile: +39 333 4963172 Email: pierangelo.masarati(a)sys-net.it ---------------------------------------

1 0

Re: (ITS#5340) REP_ENTRY_MODIFIABLE bug in dynlist
by ando＠sys-net.it 28 Mar '08

28 Mar '08

Hallvard B Furuseth wrote: > ando(a)sys-net.it wrote: > > [about REP_ENTRY_MUSTRELEASE] >> it is not >> clear what happens when a callback chain is interrupted by >> slap_null_cb() or similar, without getting to slap_send_search_entry(). >> This seems to indicate that callbacks should always provide a last >> resort means to release the resources they set; if read-only, by keeping >> track of what they sent; if modifiable, by freeing the contents of >> rs->sr_* if not NULL, setting it to NULL to prevent further cleanup. > > That sounds cumbersome, I hope slapd could take care of that somehow. Yes, it could: for example, by providing a helper like slap_entry2modifiable() that takes s SlapReply, checks the flags and does nothing if the entry is already modifiable, or copies it if not, releasing the original one if REP_ENTRY_MUSTRELEASE was set. > But I don't see how the be_release() code can work now. It sounds like > be->be_release() functions must check (how?) that the entry was created > by 'be', and otherwise pass it on to the next overlay/backend or > otherwise to entry_free(). Might involve mucking with op->o_bd and > sr_entry->e_private, I suppose. Except maybe I'm missing some existing > magic since slapd doesn't regularly crash... Yes, but that's trivial: e_private must be NULL for temporary entries, and copying the entry loses it (no one is supposed to muck with it expect the entry's creator). And the appropriate o_bd of a (non-modifiable) entry can be easily computed from the entry's DN. > be->be_release() does receive entries that were not created by 'be' or > at least not with be->be_fetch(), see openldap-devel thread 'slapd API' > in mar 2008. Might be a bug, but I'm not familiar with that code. >> Similarly, the existence of REP_ENTRY_MUSTBEFREED is not totally clear: >> in principle as soon as REP_ENTRY_MODIFYABLE is set, it should imply >> REP_ENTRY_MUSTBEFREED; the only difference in the semantics of the two >> is that REP_ENTRY_MODIFYABLE without REP_ENTRY_MUSTBEFREED implies that >> the callback that set the former will take care of freeing the entry; >> however, other callbacks may further modify it, so freeing temporary >> data should probably be left to the final handler. > > That's not my impression. MODIFIABLE would be that other modules than > the creator can modify the entry - but the creator might still be the > one who will free it. MUSTBEFREED is that the entry must be > entry_free()ed - the creator will not do it (or not do it unless that > flag is set). They clearly mean different things; my point is that as soon as an entry is modifiable it is not read-only and thus is a temporary, and thus will eventually need to be freed. So the point is who is actually going to free it. A copy might be created by an overlay after receiving a read-only entry, but the same overlay might not actually perform the copy if it receives the entry from another overlay that already copied it, or from a proxy backend. However, after the entry is copied the overlay will have no means to determined who actually created the copy. This might be an issue depending on the order cleanup handlers are called (didn't check what order they're called). My point is that temporary entries need to be freed at some point; who frees them should not be relevant... > So, if I'm getting this right... > > A backend must expect an entry to change or be freed if it sends the > entry with REP_ENTRY_<MUSTRELEASE, MUSTBEFREED or MODIFIABLE>, or if > it passes through a slap_callback.sc_response. Right now, a backend expects the entry to change if sent with MODIFIABLE; to be released if sent with MUSTBERELEASED; to be freed if sent with MUSTBEFREED. Modifications could occur to a MODIFIABLE entry when passing through a sc_callback(). Otherwise, if the entry is not MODIFIABLE, a copy will be created if the callback needs to modify the entry; after copying, the entry will be released if MUSTBERELEASED was set. When a copy is created, the callback that creates the copy must also clear the MUSTBERELEASED flag. Optionally, it may set the MUSTBEFREED flag if it doesn't intend to take care of cleaning the entry up after it's done. > back-ldif does not: it uses e_name/e_nname _after_ sending with with > REP_ENTRY_MODIFIABLE. That should be considered a bug. > Nor overlay retcode, it calls retcode_entry_response(,,,rs->sr_entry) > which makes pointers into sr_entry without checking if those flags are > set. If I'm getting that code correctly. I haven't tested. I havent' checked, but that might be a bug as well. > Others apparent problems (also not tested, I've just browsed the code): > > Overlays that obey and reset MUSTBEFREED or MUSTRELEASE, do not > necessarily clear or set MODIFIABLE when setting a new entry. > translucent does not, even though it is a careful one which > obeys both MUSTBEFREED and MUSTRELEASE. > > > I'm not sure when the code must clear the entry_related flags. > Some code which sets new entries seem to assume they are zero, > but some code which sets sr_entry=NULL does not clear the flags. > > There are sr_flags bits which are not about the entry, in particular > REP_MATCHED_MUSTBEFREED and REP_MATCHED_MUSTBEFREED. Looks like these > flags can get lost when but some code sets sr_flags = <something> > rather sr_flags |= <something> or sr_flags ~= ~<something>. It seems to me that we should provide "smart" handlers to deal with preparing sr_entry for modification, and to take care of cleaning things up as appropriate. Those helpers should then be consistently used in the code. p. Ing. Pierangelo Masarati OpenLDAP Core Team SysNet s.r.l. via Dossi, 8 - 27100 Pavia - ITALIA http://www.sys-net.it --------------------------------------- Office: +39 02 23998309 Mobile: +39 333 4963172 Email: pierangelo.masarati(a)sys-net.it ---------------------------------------

1 0

Re: (ITS#5443) Multiple identical attibutes break syncrepl process fatally?
by ando＠sys-net.it 28 Mar '08

28 Mar '08

marian.eichholz(a)freenet.ag wrote: > We use openldap as mail service directory with some 8 Mio objects on several > replicas. > For openldap 2.4.x we have to migrate from slurpd to syncrepl. > We got a working syncrepl provider als slurpd consumer (slapadd -q, 36 hrs) > > So I try to get a blank DB up by syncrepl only (yes, it is not at all > performant, but informative) > > The process kind of breaks after a couple of minutes and some 44.000 objects > (8.000.000 expected). Tracing it on the consumer side (-d 16384), I see > something like this after an entry: > > syncrepl_message_to_entry: rid=001 mods check (forwardto: value #4 provided more > than once) > > Indeed, the entry to come has three "forwardto:" Attributes with the same value > (and other forwardto-attributes, too). This makes no greate sense at the > application level, but until now it has been perfectly OK for the directory, and > the LDAP-API did not complain about the attribute modification, neither did the > slurpd. It is a violation of RFC 4512, section 2.2, which OpenLDAP 2.4 conforms to. > This leads to some questions and suggestions: > > - the provider does not log anything with -d 16384, no error, no nothing. Could > it do some useful logging about successful and failing replication sessions? What's -16384? Since OpenLDAP 2.3 you can use strings to identify each log subsystem (16384 == 0x4000 == "sync"). The error occurs when the consumer tries to manipulate the data it receives. The producer has nothing to do with it, since it assumes that data contained in it already passed sanity checks when they were stored. How incorrect data got stored into the producer is a totally different business, and the producer-side replication process should not muck with it. > - the consumer does not log anything that can explain, why the remaining objects > are not read, either. A bit of warning/logging could help the hopeful admin, > probably. A sync error occurred, which prevented sync'in from continuing. This error is logged by the "sync" subsystem. As far as I understand from reading the code, the error (at least, a replication error) should be logged also by the "any" subsystem, which means that as soon as any logging is enabled, you get a message logged. > - why is one problematic object lethal for the whole rest of the objects, since > future modifications keep to be incorporated? Is this lack of robustness more a > bug or a feature? If inconsistent data is received, synchronization is supposed to stop. In fact, continuing may result in an inconsistent state. The fact that the stop is caused by a real error, and the fact that fixing the error allows synchronization to recover doesn't sound like lack of robustness to me. It sounds more about wisdom. > - are identical attributes really forbidden with LDAP? RFC 4512, Section 2.2 > - what could one do, to prevent unskillful "editors" of the master node to kill > the replication processes for the whole replication cluster? Besides adding a > checking/filtering API layer, of course. Slapd has sanity checks for this. Slapadd doesn't, since it is supposed to be operated only with consistent data, as resulting from slapcat. You might have slapadd'ed inconsistent data to the producer. In the end, I don't see how this ITS involves a bug in synchronization software. The fact your producer got corrupted by inconsistent data might have been caused by a bug in the software, however your analysis does not give a clear indication of how it happened. If it happened by slapadd, then it's a known (and desired, and documented) limitation of the software. Unless you can reproduce it, I'd consider this ITS closed. p. Ing. Pierangelo Masarati OpenLDAP Core Team SysNet s.r.l. via Dossi, 8 - 27100 Pavia - ITALIA http://www.sys-net.it --------------------------------------- Office: +39 02 23998309 Mobile: +39 333 4963172 Email: pierangelo.masarati(a)sys-net.it ---------------------------------------

1 0

(ITS#5443) Multiple identical attibutes break syncrepl process fatally?
by marian.eichholz＠freenet.ag 28 Mar '08

28 Mar '08

Full_Name: Marian Eichholz Version: 2.4.8 OS: Linux URL: Submission from: (NULL) (194.97.7.65) We use openldap as mail service directory with some 8 Mio objects on several replicas. For openldap 2.4.x we have to migrate from slurpd to syncrepl. We got a working syncrepl provider als slurpd consumer (slapadd -q, 36 hrs) So I try to get a blank DB up by syncrepl only (yes, it is not at all performant, but informative) The process kind of breaks after a couple of minutes and some 44.000 objects (8.000.000 expected). Tracing it on the consumer side (-d 16384), I see something like this after an entry: syncrepl_message_to_entry: rid=001 mods check (forwardto: value #4 provided more than once) Indeed, the entry to come has three "forwardto:" Attributes with the same value (and other forwardto-attributes, too). This makes no greate sense at the application level, but until now it has been perfectly OK for the directory, and the LDAP-API did not complain about the attribute modification, neither did the slurpd. This leads to some questions and suggestions: - the provider does not log anything with -d 16384, no error, no nothing. Could it do some useful logging about successful and failing replication sessions? - the consumer does not log anything that can explain, why the remaining objects are not read, either. A bit of warning/logging could help the hopeful admin, probably. - why is one problematic object lethal for the whole rest of the objects, since future modifications keep to be incorporated? Is this lack of robustness more a bug or a feature? - are identical attributes really forbidden with LDAP? - what could one do, to prevent unskillful "editors" of the master node to kill the replication processes for the whole replication cluster? Besides adding a checking/filtering API layer, of course. Thank You in advance. Please let me know, if I can provide You with something useful information about our issue(s).

1 0

(ITS#5442) slapd_rq not locked before use bugfix
by rein＠basefarm.no 28 Mar '08

28 Mar '08

Full_Name: Rein Tollevik Version: CVS head OS: linux and solaris URL: ftp://ftp.openldap.org/incoming/slapd_rq_lock.patch Submission from: (NULL) (81.93.160.250) I was seeing random failures of the test050-syncrepl-multimaster test. One of the failures was that it went into a tight loop traversing a circular runqueue it had managed to create in slapd_rq.task_list. It seems as this was caused by missing mutex locks around accesses to slapd_rq, which the patch uploaded to ftp://ftp.openldap.org/incoming/slapd_rq_lock.patch fixes. Before I applied this patch the test failed after being run a few times, with it it has now passed 100 times and is still counting. Rein Tollevik Basefarm AS

1 0

(ITS#5441) Avoid unneccessary time() syscall in auditlog overlay
by rein＠basefarm.no 28 Mar '08

28 Mar '08

Full_Name: Rein Tollevik Version: CVS head OS: linux and solaris URL: ftp://ftp.openldap.org/incoming/ Submission from: (NULL) (81.93.160.250) Below is a minor enhancement to the auditlog overlay, don't fetch the time() until it is known that it is needed. Which it isn't on all those search responses that passes through it.. Rein Tollevik Basefarm AS Index: OpenLDAP/servers/slapd/overlays/auditlog.c diff -u OpenLDAP/servers/slapd/overlays/auditlog.c:1.1.1.8 OpenLDAP/servers/slapd/overlays/auditlog.c:1.2 --- OpenLDAP/servers/slapd/overlays/auditlog.c:1.1.1.8 Sat Mar 22 16:48:17 2008 +++ OpenLDAP/servers/slapd/overlays/auditlog.c Wed Mar 26 22:58:55 2008 @@ -74,7 +74,7 @@ Modifications *m; struct berval *b, *who = NULL; char *what, *suffix; - long stamp = slap_get_time(); + time_t stamp; int i; if ( rs->sr_err != LDAP_SUCCESS ) return SLAP_CB_CONTINUE; @@ -125,6 +125,7 @@ return SLAP_CB_CONTINUE; } + stamp = slap_get_time(); fprintf(f, "# %s %ld %s%s%s\n", what, stamp, suffix, who ? " " : "", who ? who->bv_val : "");

1 0

Re: (ITS#5439) syncprov race condition seg. fault
by h.b.furuseth＠usit.uio.no 28 Mar '08

28 Mar '08

rein(a)basefarm.no writes: > It now looks to me as if the entire rs->sr_entry was released an > reused, (...) Could this be ITS#5340? -- Hallvard

1 0

Re: (ITS#5340) REP_ENTRY_MODIFIABLE bug in dynlist
by h.b.furuseth＠usit.uio.no 28 Mar '08

28 Mar '08

ando(a)sys-net.it wrote: [about REP_ENTRY_MUSTRELEASE] > it is not > clear what happens when a callback chain is interrupted by > slap_null_cb() or similar, without getting to slap_send_search_entry(). > This seems to indicate that callbacks should always provide a last > resort means to release the resources they set; if read-only, by keeping > track of what they sent; if modifiable, by freeing the contents of > rs->sr_* if not NULL, setting it to NULL to prevent further cleanup. That sounds cumbersome, I hope slapd could take care of that somehow. But I don't see how the be_release() code can work now. It sounds like be->be_release() functions must check (how?) that the entry was created by 'be', and otherwise pass it on to the next overlay/backend or otherwise to entry_free(). Might involve mucking with op->o_bd and sr_entry->e_private, I suppose. Except maybe I'm missing some existing magic since slapd doesn't regularly crash... be->be_release() does receive entries that were not created by 'be' or at least not with be->be_fetch(), see openldap-devel thread 'slapd API' in mar 2008. > Similarly, the existence of REP_ENTRY_MUSTBEFREED is not totally clear: > in principle as soon as REP_ENTRY_MODIFYABLE is set, it should imply > REP_ENTRY_MUSTBEFREED; the only difference in the semantics of the two > is that REP_ENTRY_MODIFYABLE without REP_ENTRY_MUSTBEFREED implies that > the callback that set the former will take care of freeing the entry; > however, other callbacks may further modify it, so freeing temporary > data should probably be left to the final handler. That's not my impression. MODIFIABLE would be that other modules than the creator can modify the entry - but the creator might still be the one who will free it. MUSTBEFREED is that the entry must be entry_free()ed - the creator will not do it (or not do it unless that flag is set). So, if I'm getting this right... A backend must expect an entry to change or be freed if it sends the entry with REP_ENTRY_<MUSTRELEASE, MUSTBEFREED or MODIFIABLE>, or if it passes through a slap_callback.sc_response. back-ldif does not: it uses e_name/e_nname _after_ sending with with REP_ENTRY_MODIFIABLE. Nor overlay retcode, it calls retcode_entry_response(,,,rs->sr_entry) which makes pointers into sr_entry without checking if those flags are set. If I'm getting that code correctly. I haven't tested. Others apparent problems (also not tested, I've just browsed the code): Overlays that obey and reset MUSTBEFREED or MUSTRELEASE, do not necessarily clear or set MODIFIABLE when setting a new entry. translucent does not, even though it is a careful one which obeys both MUSTBEFREED and MUSTRELEASE. I'm not sure when the code must clear the entry_related flags. Some code which sets new entries seem to assume they are zero, but some code which sets sr_entry=NULL does not clear the flags. There are sr_flags bits which are not about the entry, in particular REP_MATCHED_MUSTBEFREED and REP_MATCHED_MUSTBEFREED. Looks like these flags can get lost when but some code sets sr_flags = <something> rather sr_flags |= <something> or sr_flags ~= ~<something>. -- Hallvard

1 0

Jump to page:

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

2006

openldap-bugs