IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
The history of this file is a mess with lots of bugs. Most of the bugs
I believe are based on the cache of database contents we maintain in
struct dbwrap_watch_rec. This patch removes that cache and does all
modifications directly in the backend database.
This means we have to mess with the database format in a few more
places, but I think the format is simple enough that this does not
really hurt.
I tried for a few days to split this up into small pieces that are
easier to understand, but every time I separated out individual chunks
I found difficult to track down bugs that are all resolved in the
final code presented here. It's more lines of code, but I hope it's more
robust.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
With the wakeup messages changed, we don't need to store what used to
be the old message in the struct dbwrap_watched_watch_state
anymore. We still need to store the key that is watched for removal of
our watch entry in the request destructor
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
The instance ID per process globally identifies a watch instance
across databases. It's not necessary to send the database ID and the
watched key across, the uin64_t instance ID is sufficient.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
This can lead to very confusing bugs, and the code right now does not
deal with it well.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
After dbwrap_record_storev()/delete(), dbwrap_record_get_value()
information is stale. Assert on the attempt to re-fetch data after it
became stale. This can't protect against someone copying the result
from dbwrap_record_get_value() somewhere else, but it's better than
nothing.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
I want to reduce dbwrap_record_get_value(). It makes the caller believe it can
make a copy of the TDB_DATA returned and that the value remains constant. It's
not, as you can always do a dbwrap_record_store().
This patch removes one requirement for getting the value out of a
db_record via dbwrap_record_get_value(). You can still make a copy, but from an
API perspective to me it's more obvious that "value" as a parameter to the
callback has a limited lifetime.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
"wrec->num_watchers" changes in dbwrap_watch_rec_del_watcher(). In
32d6cc84c I forgot to update the copy of that variable.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
Autobuild-User(master): Jeremy Allison <jra@samba.org>
Autobuild-Date(master): Fri Oct 18 22:28:07 UTC 2019 on sn-devel-184
This fixes the following flaky test:
UNEXPECTED(failure): samba3.raw.oplock.batch26(nt4_dc)
REASON: Exception: Exception: (../../source4/torture/raw/oplock.c:3718): wrong value for break_info.count got 0x2 - should be 0x1
You can reproduce it with two small msleeps, which means it's a race
condition:
diff --git a/source3/smbd/open.c b/source3/smbd/open.c
index 20b5a3e294c..126c7fc021d 100644
--- a/source3/smbd/open.c
+++ b/source3/smbd/open.c
@@ -1917,6 +1917,14 @@ NTSTATUS send_break_message(struct messaging_context *msg_ctx,
DATA_BLOB blob;
NTSTATUS status;
+ {
+ static bool sent = false;
+ if (sent) {
+ smb_msleep(500);
+ }
+ sent = true;
+ }
+
if (DEBUGLVL(10)) {
struct server_id_buf buf;
DBG_DEBUG("Sending break message to %s\n",
diff --git a/source3/smbd/oplock.c b/source3/smbd/oplock.c
index b3da84b1269..d9c4dbb9487 100644
--- a/source3/smbd/oplock.c
+++ b/source3/smbd/oplock.c
@@ -858,6 +858,8 @@ static void process_oplock_break_message(struct messaging_context *msg_ctx,
uint16_t break_to;
bool break_needed = true;
+ smb_msleep(100);
+
msg = talloc(talloc_tos(), struct oplock_break_message);
if (msg == NULL) {
DBG_WARNING("talloc failed\n");
15a8af075a introduced a bug where we immediately wake up ourselves
after doing a watch_send, leading to two inter-smbd oplock break
messages for this case. In theory, this should not matter, as in the
oplock break handler in the destination smbd we check
(fsp->sent_oplock_break != NO_BREAK_SENT)
so that the break does not get sent twice. However, with the above two
sleeps the oplock holding client could send out its oplock downgrade
while the second inter-smbd break messages was on its way.
The real fix would be to note in the share mode array that the
inter-smbd message has already been sent, but as other users of
dbwrap_watched_watch_send might also be affected by this bug, this fix
should be sufficient to get rid of this flaky test.
Unfortunately, dbwrap_watch.c is now pretty complex and needs some
serious refactoring to become understandable again. But that's
something for another day, sorry.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
Before 15a8af075a we did not have a separately allocated watchers
array and dbwrap_watched_save() could play (too) smart tricks with
dbwrap_record_storev(). Now that we always have watchers talloc'ed, we
can remove those smart tricks from dbwrap_watched_save() in the next
commit.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
Samba so far on a dbwrap_watch record either watches or stores a
record, but never both from the same db_record instance acquired from
a dbwrap_fetch_locked(). In one of the next commits this will change,
we will watch a record and at the same time store data into it. This
patch enables a watch_send() followed by a storev() by properly
keeping the watchers around.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
Without this, to notify watchers you need to actually store data. This
might be a waste of resources. locking.tdb waiters might actually wait
for leases.tdb or brlock.tdb changes, and locking.tdb records can be
large.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
In the next step I want to make dbwrap_watched_wakeup() publically
available under that canonical name.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
Point out in the API that "backend" talloc_moves into the watched
database.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
Autobuild-User(master): Jeremy Allison <jra@samba.org>
Autobuild-Date(master): Fri Aug 17 21:29:15 CEST 2018 on sn-devel-144
The initial idea was to have some "atomicity" in this API. Every
caller interested in a record would have to do something with
it once it changes. However, only one caller really used this
feature, and that is easily changed to not use it. So
remove the complexity.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
The idea is to leave the "watchers" array unparsed until it's needed. This
avoids a few talloc calls and unnecessary parsing.
Also, it deletes quite a few lines of code and .text bytes.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
Convert all implementors of dbwrap_store to a storev-style call
by using the dbwrap_merge_dbufs call
For dbwrap_tdb, this matches tdb_storev.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
If locking.tdb contains invalid records, "get_file_infos" called from directory
enumeration crashes in Samba 4.4. The reason is that if "dbwrap_watched_parse"
returns -1 due to record corruption, dbwrap_watched_parse_record returns
NT_STATUS_OK without having called the parse function. Before 66cba9939b
this led to "lck->data" to be uninitialized data, so smbd 4.4 would crash in
this case. After 66cba9939b we implicitly initialize "state.lck" to NULL,
so we don't have this particular problem anymore
Apply the fix in master too, returning NT_STATUS_OK from parse_record without
having called the parser could lead to bugs in other cases too.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
The existing one with a separate dbwrap_watchers.tdb turns out to
create a performance penalty in a clustered environment. Non-clustered,
dbwrap_parse_record on non-existent records is very cheap, but in a
cluster environment this is very noticable.
This implementation puts the watcher information into the records itself. For
large records, this might be another performance penalty, because we have to
assemble the final record together with talloc and memcpy, but this might be
fixed later with a tdb_storev call.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
It's highly unlinkely that this will ever kick in, because our current tdb keys
are rather small, but offset calculations without overflow checks are bad.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
Typicall, when we watch a record, we wait for a process to give up some
resource. Be it an oplock, a share mode or the g_lock. If everything goes well,
the blocker sends us a message. If the blocker dies hard, we want to also be
informed immediately.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
This happens on every store to locking.tdb for example, so we should
make it cheap.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
Autobuild-User(master): Ralph Böhme <slow@samba.org>
Autobuild-Date(master): Tue Sep 22 07:50:58 CEST 2015 on sn-devel-104
It took a bit for me to figure out what the rec_key parameter to
dbwrap_record_watchers_key does. I think it's simpler to parse the watcher key
in dbwrap_record_watch_recv to retrieve the watched record key than to store it
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Ralph Boehme <slow@samba.org>
Replace all callers with direct calls to server_id_str_buf without
talloc_tos()
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
In order to receive the fd array the caller needs to use
messaging_filtered_read_send/recv(). For all higher level
methods we silently close/ignore the fd array.
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Michael Adam <obnox@samba.org>
This does not really save any code lines, but IMHO the code is simpler
this way. Also, in case we have lots of watchers this will be slightly
cheaper, because we don't have to re-establish a tevent_req.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
We have messaging_send_buf, so there's no point to use an explicit
DATA_BLOB here.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Christof Schmitt <cs@samba.org>
Autobuild-User(master): Jeremy Allison <jra@samba.org>
Autobuild-Date(master): Tue Feb 25 23:27:21 CET 2014 on sn-devel-104
This is in preparation to support handing flags to backends,
in particular activating read only record support for ctdb
databases. For a start, this does nothing but adding the
parameter, and all databases use DBWRAP_FLAG_NONE.
Signed-off-by: Michael Adam <obnox@samba.org>
Reviewed-by: Stefan Metzmacher <metze@samba.org>
