| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Factor out most of ulog_add_update into a helper function named
store_update, and make ulog_add_update just responsible for assigning
a serial number and timestamp to the update before storing it. In
ulog_replay, use store_update and ulog_finish_update to add each
update to the ulog in addition to replaying it to the database. Don't
use incr_ret->lastentry to set kdb_last_sno/kdb_last_time, since it
will have been set properly by adding the individual updates; instead,
just reinitialize the ulog on error.
Slave ulogs use serial numbers provided from upstream, and thus do not
always begin at serial number 1 after a header reset. As a result, we
must: (A) in store_update, detect the first update (for which we must
assign kdb_first_sno/kdb_first_time) by comparing kdb_num to 0,
instead of by comparing the serial number to 1; (B) in store_update,
detect that we are overwriting the first update by comparing kdb_num
to ulogentries, instead of comparing the serial number to ulogentries;
and (C) in ulog_map, detect that ulogentries changed by verifying the
first and last serial number and timestamp against the actual ulog
entries, rather than simply comparing kdb_last_sno to kdb_num.
Based on code submitted by Richard Basch.
ticket: 7855
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Get rid of the caller parameter. The kproplog semantics (without -R)
for mapping the ulog are simple and almost completely different from
other users of the ulog, so implement them as a static helper in
kproplog. With hierarchical iprop, kpropd will need the same
semantics as FKCOMMAND and FKADMIND, which were already identical.
Get rid of the db_args parameter, since ulog_map no longer opens the
database after #7552.
Remove an inoperative lseek() call when creating a new ulog file.
Rename ulog_filesize to filesize and compute it from scratch each time
we use it, for easier analysis. If kdb_hmagic is zero, init the ulog
header but don't skip the rest of the function; it's possible that we
need to expand the ulog file. Remove an unneeded conditional before
calling extend_file_to for an existing ulog.
ticket: 7855
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Always lock the ulog when accessing it. We can currently get away
with some laxness on iprop slaves because they are mostly synchronous,
but hierarchical iprop will allow master and slave operations to take
place concurrently, requiring more strict locking.
Add new functions ulog_get_last and ulog_set_last, which access the
ulog header with locking, and use them in kdb5_util and kpropd. Add
locking to ulog_replay and ulog_init_header.
ulog_lock and ulog_sync_header are no longer used outside of kdb_log.c
after these changes, so make them static functions and remove the
ulog_ prefix. Add an unlock_ulog function for clarity.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add a new function ulog_get_sno_status, which checks a serial number
and timestamp against the ulog for currency. Use it in kdb5_util dump
and in ulog_get_entries. Adjust parse_iprop_header's contract in
dump.c to better match the ulog_get_sno_status contract.
This change causes some minor behavior differences. kadmind will
check for an empty ulog unless the last serial number matches exactly,
and will never set lastentry when returning UPDATE_FULL_RESYNC_NEEDED
(which was pointless). kdb5_util dump will recognize a dump file as
current if it exactly matches the last serial number, even if the ulog
is empty; it will be more robust in the presence of non-monotonic
clocks; and it will properly lock around the ulog access.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Since we are no longer treating the update log like a journal (#7552),
we don't need two-stage update logging. In kdb5.c, add an update log
entry after each DB change in one step, without getting an explicit
lock. In kdb_log.c, combine ulog_add_update with ulog_finish_update,
and make ulog_add_update lock the ulog internally.
This change avoids deadlock by removing the only cases where the ulog
is locked before the DB.
ticket: 7861
|
| |
|
|
|
|
|
|
| |
ulog_get_entries does not access the KDB, only the ulog, so it does
not need a KDB lock; its read lock on the ulog is sufficient to
prevent logged updates from happening while it is running. There is
no reason to serialize against unlogged KDB updates such as those
performed by the KDC, as those do not affect the ulog.
|
| |
|
|
|
|
|
| |
Always map the ulog with FKCOMMAND, not FKLOAD which reinitializes the
ulog. Don't reinitialize the ulog until just before calling
krb5_db_promote(). Get rid of FKLOAD since it isn't needed any more;
we can just call ulog_init_header() manually.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Don't use "extern" for kdb_log.h prototypes.
* Avoid passing structures by value.
* Avoid the need to cast the result of the INDEX macro, and use char *
instead of unsigned long for pointer arithmetic.
* Reorganize kdb_log.c so static helpers are at the top and don't use
the "ulog_" prefix.
* Get rid of ulog_finish_update_slave since it's more concise to
open-code it in ulog_replay.
* Get rid of ulog_delete_update. In krb5_db_delete_principal, just
call ulog_add_update with kdb_deleted set in upd.
* Modernize coding style of kproplog.c. Use k5memdup0 instead of
snprintf in print_str to convert a byte range to a C string.
Remove an unnecesary textdomain call; libkrb5 takes care of calling
bindtextdomain in the library initializer.
* Modernize coding style of kpropd.c and kprop.c.
No functional changes.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
If a process terminates in the middle of a logged database update, it
will release its lock and leave the ulog in the KDB_UNSTABLE state.
kadmind should notice this when it calls ulog_get_entries, but right
now it only checks for the KDB_CORRUPT state (which is never set any
more) and does not recover. Check for any state other than KDB_STABLE
and recover by resetting the ulog and forcing full resyncs to slaves.
ticket: 7853 (new)
target_version: 1.12.2
tags: pullup
|
| |
|
|
|
|
|
|
|
|
|
| |
When we wrap the serial number, reinitialize the ulog so that kdb_num
starts counting from 1, instead of leaving the ulog in a weird state
where kdb_num == ulogentries but kdb_last_sno == 1 and the serial
numbers in the circular array aren't necessarily contiguous.
ticket: 7839 (new)
target_version: 1.12.2
tags: pullup
|
| |
|
|
|
|
| |
Add k5-int.h static functions to duplicate byte ranges, optionally
with a trailing zero byte, and set an error code like k5alloc does.
Use them where they would shorten existing code.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The db2 DB is not power-fail safe. There's no point trying to
replay an incompletely committed entry from the ulog at kadmind
startup time. For that matter, even if the db2 DB was power-fail
safe there'd be no point replaying an uncommitted entry from the
ulog as the libkadm5srv app (nor any client of it, as in the case of
kadmind) will not have received any notice of success -- it'd be
wrong to complete that operation later when the user thought it'd
failed.
[ghudson@mit.edu: merge with master, adjust comment]
ticket: 7552 (new)
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If the master iprop log is reinitialized to serial number 0, slaves
will need to take a full dump--but after that happens, we need to know
whether the slave has taken that full dump, we we don't offering full
dumps indefinitely.
So, record a timestamp in kdb_last_time when we reinitialize the log
header, and compare the slave timestamp to kdb_last_time whenever it
has the current serial number, even if it's 0. Test this by
performing a propagation with sno 0 in t_iprop.py and detecting
whether kpropd gets a second UPDATE_FULL_RESYNC_NEEDED response from
kadmind.
ticket: 7550 (new)
|
| |
|
|
|
| |
ulog_get_entries had an unreachable branch which was removed during
de-indentation.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If a kdb5_util load gets killed between rename()ing the new KDB file
into place and resetting the iprop ulog then the ulog can reflect the
pre-load state, which will almost certainly be incorrect.
This matters because we want to impose a timeout on full resyncs in
kpropd when iprop dictates that a full resync is needed, and the
simplest timeout scheme involves signaling the kdb5_util load process.
But also, we want no such races in general.
The fix is simple: re-initialize the ulog before renaming the new KDB
file into place, then proceed as usual. If the ulog is not properly
updated at the end of the load it will at least always result in
subsequent iprop get updates operations always indicating that a full
resync is required.
ticket: 7399
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently kadmind allows slaves to poll for updates as often as they
like, but not within 10s of the last update. This means that iprop will
appear to fail to synchronize the KDC at any site whose master KDC
processes at least one write transaction every 10 seconds consistently.
The original intention must have been to throttle iprop clients (slave
KDCs) that poll too often. But UPDATE_BUSY as implemented is not that,
and implementing a throttle would be difficult (requires keeping state
in a table) and mostly useless (admins can manage their poll timers just
fine without a throttle in kadmind).
ticket: 7369
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Use a global dump (the default dump file) for full syncs for iprop.
When a slave asks for a fullsync we kprop the existing global dump to it
if that is good enough, else we dump the DB and send the new global
dump.
Before this change kadmind would run kdb5_util dump -i... each time a
slave asked for a full dump. This was done in a sub-process,
thankfully, but it was still a waste of time and storage (e.g., if one
has a huge KDB).
Also, long dump times might cause a slave to give up (the timeout for
this is now configurable). But since iprop dumps bear a serial number
and timestamp and since slaves will resync from that point forward, it
doesn't matter if the dump we send a slave is fresh as long as it is
fresh enough (i.e., that its sno and timestamp are in the ulog).
Also:
- Rename dumps into place instead of unlink, create, write (but we
still keep the dump ok files as lock files and as a method of
signaling to kprop that the dump is complete).
ticket: 7371
|
| |
|
|
|
|
| |
Unsigned long might not be 32-bits, in which case the overflow check
fails for a 32-bit kdb_sno_t. Cast from -1 explicitly as a more
robust check.
|
| |
|
|
|
|
| |
ticket: 6918
git-svn-id: svn://anonsvn.mit.edu/krb5/trunk@24961 dc483132-0cff-0310-8789-dd5450dbe970
|
| |
|
|
| |
git-svn-id: svn://anonsvn.mit.edu/krb5/trunk@24703 dc483132-0cff-0310-8789-dd5450dbe970
|
| |
|
|
|
|
|
|
|
|
|
| |
free_principal, delete_principal, and get_policy. Make get_principal
allocate the DB entry container. Fold krb5_db_get_principal_ext into
krb5_db_get_principal.
ticket: 6749
status: open
git-svn-id: svn://anonsvn.mit.edu/krb5/trunk@24175 dc483132-0cff-0310-8789-dd5450dbe970
|
| |
|
|
|
|
| |
make reindent
git-svn-id: svn://anonsvn.mit.edu/krb5/trunk@23100 dc483132-0cff-0310-8789-dd5450dbe970
|
| |
|
|
|
|
|
|
| |
Fix a number of warnning suggesting parenthesis.
Fix a signed/unsigned warning.
Update dependencies.
git-svn-id: svn://anonsvn.mit.edu/krb5/trunk@21627 dc483132-0cff-0310-8789-dd5450dbe970
|
| |
|
|
| |
git-svn-id: svn://anonsvn.mit.edu/krb5/trunk@20624 dc483132-0cff-0310-8789-dd5450dbe970
|
| |
|
|
| |
git-svn-id: svn://anonsvn.mit.edu/krb5/trunk@20619 dc483132-0cff-0310-8789-dd5450dbe970
|
|
|
git-svn-id: svn://anonsvn.mit.edu/krb5/trunk@20465 dc483132-0cff-0310-8789-dd5450dbe970
|
