Jeremy Allison [Thu, 3 Sep 2009 14:40:48 +0000 (07:40 -0700)]
Hopefully last part of the fix for bug 6651 - smbd SIGSEGV when breaking oplocks.
This one is subtle. There is a race condition where a signal can be
queued for oplock break, and then the file can be closed by the client
before the signal can be processed. Currently if this occurs we panic
(we can't match an incoming signal fd with a fsp pointer). Simply log
the error (at debug level 10 right now, might be too much) and then
return without processing the break request. It looks like there is
another race condition with this fix, but here's why it won't happen.
If the signal was pending (caused by a kernel oplock break from a
local file open), and the client closed the file and then re-opened
another file which happened to use the same file descriptor as the
file just closed, then theoretically the oplock break requests could
be processed on the wrong fd. Here's why this should be very rare..
Processing a pending signal always take precedence over an incoming
network request, so as long as the client close request is non-chained
then the break signal should always be harmlessly processed *before*
the open can be called. If the open is chained onto the close, and
the fd on the new open is the same as the old closed fd, then it's
possible this race will occur. However, all that will happen is that
we'll lose the oplock on this file. A shame, but not a fatal event.
Jeremy.
Jeremy Allison [Thu, 3 Sep 2009 14:38:21 +0000 (07:38 -0700)]
Another part of the fix for bug 6651 - smbd SIGSEGV when breaking oplocks.
SA_INFO_QUEUE_COUNT *MUST* be a power of 2, in order for the ring buffer
wrap to work correctly at the 32 bit boundary. Thanks to Petr
Vandrovec <petr@vandrovec.name> for this.
First attempt to implement dcesrv_drsuapi_DsGetNCChanges
So far it returns the ctr6 responce without proper linked attributes
support and metadata. A couple of improvements are the filter in the search
uses '(uSNChanged>=N)', added extended dn search support, non-replicated attributes
are excluded from the result.
This replace the dsdb_syntax_FOOBAR_ldb_to_drsuapi function,
which was left as a TODO code. Implementation in both added functions
is completely identical and probably should differ in the future.
Andrew Tridgell [Thu, 3 Sep 2009 08:36:02 +0000 (18:36 +1000)]
another large change to the linked_attribute module
This one copes with deleted objects where linked attributes have been
set on the module. We hit this when we do the ldb wipe at the start of
a provision, which trigers linked attribute updates, but for objects
that have disappeared. We need to ensure that the linked attribute
updates only happen on the right object, and if the object gets
re-created (as happens with a provision) then it is not the right
object.
To cope with this we record the GUID of the object when the operation
that triggered the linked attribute update comes in, and then find the
DN by suing that GUID when we apply the change in the prepare commit
hook.
Andrew Tridgell [Thu, 3 Sep 2009 08:31:45 +0000 (18:31 +1000)]
show more reasonable object counts during a vampire
We now show the total number of objects we have processed, which gives
the user a better idea of how much has been done. A vampire on a large
domain can take an hour or more (which needs to be fixed btw, it is a
problem with the lack of scalability of the ltdb index code). Watching
the same msg for an hour makes you wonder if any progress is being
made!
Andrew Tridgell [Thu, 3 Sep 2009 08:29:58 +0000 (18:29 +1000)]
always use prepare_commit in ldb transaction commits if possible
The reason we need this is to make multi-tdb transactions safe, with
the partition module. The linked_attributes and repl_meta_data modules
now do extra processing when the transaction ends, and that processing
can fail. When it fails we need to cancel the transaction, which we
can only do if the hook is on the prepare commit instead of the end
transaction call. Otherwise the partition module cannot ensure that no
commit has been done on another partition.
Andrew Tridgell [Thu, 3 Sep 2009 02:55:57 +0000 (12:55 +1000)]
change repl_meta_data to process linked_attributes structures in end_transaction
When running at functional level 2 or above, the repl_meta_data module
can receive linked attribute structures from the repl replication
task. These attributes can come through DRS before the associated
objects have been created. To cope with this, we need to process
linked attributes in the end_transaction hook.
Before the async libsmb rewrites, we sent tid==0 on negprot. With the rewrite,
we send 0xffff. This *should* not matter, but this is one difference in the
sniffs I see.
Andrew Tridgell [Wed, 2 Sep 2009 08:18:33 +0000 (18:18 +1000)]
traverse the ac list in reverse order
items are added to the linked attribute list using DLIST_ADD(), which
means to commit them to the database in the same order they came from
the server we need to walk the list backwards when we traverse it
Andrew Tridgell [Wed, 2 Sep 2009 07:02:56 +0000 (17:02 +1000)]
s4:dsdb rewrite the linked_atrributes code to commit in the end_transaction hook
linked attribute changes can come in any order. This means it is
possible for a forward link to come over the wire in DRS before the
target even exists. To make this work this patch changed the linked
attributes module to gather up all the changes it needs to make in a
linked list, then execute the changes in the end_transaction hook for
the module.
During that commit phase we also fix up all the DNs that we got by
searching for their GUID, as the objects may have moved after the
linked attribute was sent, but before the end of the transaction
Andrew Tridgell [Wed, 2 Sep 2009 03:36:54 +0000 (13:36 +1000)]
change the dsdb_control_current_partition to not include internal variables
This structures was used in two ways. In one way it held variables
that are logically internal to the partition module, and in the other
way it was used to pass the partition DN down to other modules. This
change makes the structure contain just the dn which is being passed
down.
This change is part of the support for linked attributes. We will be
passing this control down from above the partition module to force
which partition a request acts upon. The partition module now only
adds this control if it isn't already there.
Volker Lendecke [Sat, 29 Aug 2009 07:41:32 +0000 (09:41 +0200)]
tevent: Fix a segfault upon the first signal
When the first signal arrives, tevent_common_signal_handler() crashed: "ev" is
initialized to NULL, so the first "write(ev->pipe_fds[1], &c, 1);" dereferences
NULL.
Rusty, Tridge, please check. Also, can you tell me a bit more about the
environment you tested this in? I'd be curious to see where this survived.
projects / thirdparty / samba.git / log
