id2entry: when an error occurs attempting to insert fetched
entry into cache, check to see if entry was added by another
thread. If so, return the entry added by the other thread.
The concurrency tests now run without the dreaded "no such object"
error messages!
Removed external include/library paths from projects. External paths should
be set via Tools | Options | Directories. This allows each developer the
freedom to install external libraries where they desire.
Used libdb.lib instead of libdbs.lib to avoid thread conflicts.
Added hs_regex.lib to library input. We require some form of regex, this
library works (and is relatively easy for the user to install).
Removed a little lint which MCVC5 detected.
Need to sort out single-threaded vs multithreaded library generation.
remove race conditions on connection close.
BROKEN: various counters for dn=monitor.
Initial testing on FreeBSD (with and without pthreads) was successfull.
Have not yet tested preemptive threading environments.
Have not built against backends other than LDBM.
cache_set_state() and state args are no history.
Move cache_entry_cmp() and friends to slapd/entry.c to allow sharing
of functions between backends.. Renamed to entry_cmp().
The new e_private field can be used for any backend purpose. In LDBM,
it's used to point th the private entry info struct. The LDBM entry
info struct is only visible to the LDBM cache.
WARNING: This change has not be ported to other backends! Breakage is likely.
Integration into ./configure ( --enable-bdb2 ).
Adaptation of the test-suite ( 'make' -> ldbm; 'make bdb2-local' -> bdb2 ).
Minor fixes in slapd/libldbm.
BEWARE: the backend will be compiled.
the backend will NOT be invoked, yet.
the backend CANNOT be invoked, yet, because it is NOT yet
integrated into the new initialization/startup environment
of the slapd server.
Separates per backend type from per backend database initialization
and startup. Also supports per type / per backend shutdown.
New frontend startup/shutdown routines are also provided:
slap_init() slap_startup() slap_shutdown() slap_destroy()
New frontend->backend startup/shutdown is managed by:
backend_init() backend_startup() backend_shutdown backend_destroy
backend_init() now calls bi_init() to initial all function pointers
for the backend (excepting bi_init() which is now the only hardcoded
entry point). New entry points are detailed in slap.h struct
backend_info. backend_info is a per database type structure.
Besides the new startup/shutdown entry points, the new interface
also supports per backend type configuration options. One could have:
backend bdb2 (new Berkeley DB 2 backend)
bdb2_home /directory
database bdb2
...
*** This code is fairly experimental ***
*** Much cleanup and testing is still needed ***
see slap.h for details on struct backend_db and backend_info.
okay to read and write LDBM specific fields (state, refcnt,
LRU. The id field, though is read-only once set.
cache_find_entry_dn2id(), hence, does not require any entry locks.
cache_find_entry_id() must do a entry_rdwr_trylock() and back
off if busy.
Add new rdwr lock code with trylock() functionality.
Implement entry_rdwr_trylock().
Fix idl_nextid not to skip an ID if ID not in list.
Both of these should not occur if used correctly, but cheap insurance
against incorrect usage is welcomed.
instead of array with overlayed struct fields. (which is inherently
non-portable).
Change maxid check in insert to be greater than or equal instead
of equal.
Use SAFEMEMCPY.
Likely broke things for non-posix threadings....
Update -lldap_r implementation to:
remove attribute support
hide thread detachment
provide concurrency accessors
provide initialization function
fix gethostby{addr,name}_r codes (not coverred by HAVE_REENTRANT_FUNCTIONS)
Update servers/libraries to use ldap_pvt_thread_ calls.
Cleanup server codes (no #ifdef HAVE_PTHREAD_THIS or _THATs)!
Removed -llthread
back-ldbm was using a cache specific lock to protect non-reentrant
database routines from being reenterred. This is inadequate.
Also, reentrant database systems calls were serialized unnecessarily.
Non-reentrant database calls must have a big_mutex. Implemented
this within -lldbm itself. library requires ldbm_initialize()
be called before any other ldbm call to initialize the big_mutex
and to do any other db specific initialization (ie: such as
required for DB2).
The dbc_mutex, dbc_cv, & dbc_readers fileds of dbcache are history.
The "reentrant_database" (REENTRANT_DATABASE) define is also axed.
behind REENTRANT_DATABASE. rename reentrant_database -> REENTRANT_DATABASE.
When REENTRANT_DATABASE is defined, the simple lock is replaced
with a simple reader/writer lock. This functionality should be removed
from back-ldbm to libldbm.
SLAPD_NEXTID_CHUNK. Code protects NEXTID file to ensure
its equal to or greater than nextid. Updated on close
to actual nextid. next_id_save() could be called periodically
if desired. Default chunk size is 32. Define to 1 to disable
chunking.
Update slap_op to maintain dn and ndn (derived from conn->c_dn).
Update ldbm_back_bind to return actual bound dn (including rootdn)
for use in slapd_conn. Other backends use client dn.
Modify other codes to use ndn (normalized uppercase dn) most everywhere.
Aliasing, Suffixing and modrdn could use more work.
Applied suffixing to compare and modrdn.
creation codes to provide this field. Update cache_entrydn_cmp
to strcasecmp() the e_ndn instead of e_dn. Note: strcasecmp()
is still used as e_ndn isn't in uppercase. Maybe it should
be. Did not update other codes to use e_ndn. Hence, there
are lots of dn_normalize() calls that could be eliminated.
(The case determination of e_ndn should be made first).
parent's "children" attribute. Write lock parent to prevent
multiple clients making conflicting operations concurrently.
If parent doesn't exist (ie: is backend root), acquire a writer
lock (a simple mutex) on the "root."
Use -DSLAPD_CHILD_MODIFICATION_WITH_ENTRY_ACL to use the child's
"entry" acl for modrdn/delete operations.
- slapd's pid/args file names are based on the servers binary file names,
providing for multiple servers beeing run on one host.
- slapd supports the -l command line parameter for selection of a
syslog LOCAL user (-lLOCAL0 .. -lLOCAL7)
- db_appinit() is called during first ldbm_open() in DB 2.x to initialize
DB debugging features (good to find bugs in the DB code :-)
- a patch for a non-initialized variable in DB's 2.x db_open is provided.
Resolved deadlock by passing target entry to be_group and using
this if dn same as bdn. It might actually be safer to check
entry ids instead of dns.
Resolved bogus add to cache after failed acl check by deferring
cache add until after parent/acl checks have successful been
completed.
Eliminated race condition caused by concurrent adds
of same dn by adding 'li_add_mutex' around the critical section
of code (most of ldbm_back_add).
This code is preliminary and still needs significant testing.
