Without this change the following situation was possible:
1. Provider is orphaned from within class' access() method on last write
close - orphan provider event is send.
2. GEOM detects last write close on a provider and sends new provider event.
3. g_orphan_register() is called, and calls all orphan methods of attached
consumers.
4. New provider event is executed on orphaned provider, all classes can
taste already orphaned provider, and some may attach consumers to it.
Those consumers will never go away, because the g_orphan_register()
was already called.
We end up with a zombie provider.
With this change, at step 3, we will cancel new provider event.
How to repeat this problem:
# mdconfig -a -t malloc -s 10m
# geli init -i 0 md0
# geli attach md0
# newfs -L test /dev/md0.eli
# mount /dev/ufs/test /mnt/tmp
# geli detach -l md0.eli
# umount /mnt/tmp
# glabel status
Name Status Components
ufs/test N/A N/A
Reviewed by: phk
Approved by: re (kensmith)
This way GEOM classes can safely detach from provider when an orphan
event is received. This fixes 'detach with active requests' panic for
gstripe/gconcat under load.
PR: kern/102766
Submitted by: mjacob
OK'ed by: phk
MFC after: 1 week
waiting for geom events to happen:
Instead of maintaining a count of outstanding events, simply look if
the queue is empty. Make sure to not remove events from the queue
until they are executed in order to not open a new race.
Much work by: pjd
Tested by: kris
MT6: yes, should be.
This flag gets set whenever the thread posts an event on the GEOM
event queue, and if the flag is set when the thread is prepared
to return to userland from the kernel, g_waitidle() will be called
to make sure that the posted events have completed.
This can replace an insufficient number of g_waitidle() calls in
various other places, and has the advantage of being failsafe: Any
system call which does a VOP_OPEN()/VOP_CLOSE will now correctly
wait for any geom events it posted as part of spoils or tastes.
Assert that topology and Giant is not held in g_waitidle().
When we orphan/wither a provider, an attached geom+consumer could
end up being withered as a result and it may be in front of us in
the normal object scanning order so we need to do multi-pass. On
the other hand, there may be withering stuff we can't get rid off
(yet), so we need to keep track of both the existence of withering
stuff and if there is more we can do at this time.
Retire g_sanity() and corresponding debugflag (0x8)
Retire g_{stall,release}_events().
Under #ifdef DIAGNOSTIC:
Make g_valid_obj() an official function and have it return an an
non-zero integer which indicates the kind of object when found.
Implement G_VALID_{CLASS,GEOM,CONSUMER,PROVIDER}() macros based
on g_valid_obj().
Sprinkle calls to these macros liberally over the infrastructure.
Always check that we do not free a live object.
test is built to test GEOM as running in the kernel.
This commit is basically "unifdef -D_KERNEL" to remove the mainly #include
related code to support the userland-harness.
event posting functions varargs to fill these.
Attribute g_call_me() to appropriate g_geom's where necessary.
Add a flag argument to g_call_me() methods which will be used to signal
cancellation of events in the future.
This commit should be a no-op.
to be performed in the event-thread.
To do this, we need to lock the eventlist with g_eventlock (nee g_doorlock),
since g_call_me() being called from the UP/DOWN paths will not be able to
aquire g_topology_lock.
This also means that for now these events are not referenced on any
particular consumer/provider/geom.
For UP/DOWN path use, this will not become a problem since the access()
function will make sure we drain any bio's before we dismantle.
Sponsored by: DARPA & NAI Labs.
