Re: [dpdk-dev] [PATCH v2 2/6] ethdev: add port ownership

[Matan Azrad <matan@mellanox.com>]

Hi Konstantin
From: Ananyev, Konstantin, Sent: Wednesday, January 17, 2018 1:24 PM
> Hi Matan,
> 
> > Hi Konstantin
> >
> > From: Ananyev, Konstantin, Tuesday, January 16, 2018 9:11 PM
> > > Hi Matan,
> > >
> > > >
> > > > Hi Konstantin
> > > > From: Ananyev, Konstantin, Monday, January 15, 2018 8:44 PM
> > > > > Hi Matan,
> > > > > > Hi Konstantin
> > > > > > From: Ananyev, Konstantin, Monday, January 15, 2018 1:45 PM
> > > > > > > Hi Matan,
> > > > > > > > Hi Konstantin
> > > > > > > > From: Ananyev, Konstantin, Friday, January 12, 2018 2:02
> > > > > > > > AM
> > > > > > > > > Hi Matan,
> > > > > > > > > > Hi Konstantin
> > > > > > > > > > From: Ananyev, Konstantin, Thursday, January 11, 2018
> > > > > > > > > > 2:40 PM
> > > > > > > > > > > Hi Matan,
> > > > > > > > > > > > Hi Konstantin
> > > > > > > > > > > > From: Ananyev, Konstantin, Wednesday, January 10,
> > > > > > > > > > > > 2018
> > > > > > > > > > > > 3:36 PM
> > > > > > > > > > > > > Hi Matan,
> > > >  <snip>
> > > > > > > > > > > > > It is good to see that now scanning/updating
> > > > > > > > > > > > > rte_eth_dev_data[] is lock protected, but it
> > > > > > > > > > > > > might be not very plausible to protect both
> > > > > > > > > > > > > data[] and next_owner_id using the
> > > > > > > same lock.
> > > > > > > > > > > >
> > > > > > > > > > > > I guess you mean to the owner structure in
> > > > > > > rte_eth_dev_data[port_id].
> > > > > > > > > > > > The next_owner_id is read by ownership APIs(for
> > > > > > > > > > > > owner validation), so it
> > > > > > > > > > > makes sense to use the same lock.
> > > > > > > > > > > > Actually, why not?
> > > > > > > > > > >
> > > > > > > > > > > Well to me next_owner_id and rte_eth_dev_data[] are
> > > > > > > > > > > not directly
> > > > > > > > > related.
> > > > > > > > > > > You may create new owner_id but it doesn't mean you
> > > > > > > > > > > would update rte_eth_dev_data[] immediately.
> > > > > > > > > > > And visa-versa - you might just want to update
> > > > > > > > > > > rte_eth_dev_data[].name or .owner_id.
> > > > > > > > > > > It is not very good coding practice to use same lock
> > > > > > > > > > > for non-related data structures.
> > > > > > > > > > >
> > > > > > > > > > I see the relation like next:
> > > > > > > > > > Since the ownership mechanism synchronization is in
> > > > > > > > > > ethdev responsibility, we must protect against user
> > > > > > > > > > mistakes as much as we can by
> > > > > > > > > using the same lock.
> > > > > > > > > > So, if user try to set by invalid owner (exactly the
> > > > > > > > > > ID which currently is
> > > > > > > > > allocated) we can protect on it.
> > > > > > > > >
> > > > > > > > > Hmm, not sure why you can't do same checking with
> > > > > > > > > different lock or atomic variable?
> > > > > > > > >
> > > > > > > > The set ownership API is protected by ownership lock and
> > > > > > > > checks the owner ID validity By reading the next owner ID.
> > > > > > > > So, the owner ID allocation and set API should use the
> > > > > > > > same atomic
> > > > > > > mechanism.
> > > > > > >
> > > > > > > Sure but all you are doing for checking validity, is  check
> > > > > > > that owner_id > 0 &&& owner_id < next_ownwe_id, right?
> > > > > > > As you don't allow owner_id overlap (16/3248 bits) you can
> > > > > > > safely do same check with just atomic_get(&next_owner_id).
> > > > > > >
> > > > > > It will not protect it, scenario:
> > > > > > - current next_id is X.
> > > > > > - call set ownership of port A with owner id X by thread 0(by
> > > > > > user
> > > mistake).
> > > > > > - context switch
> > > > > > - allocate new id by thread 1 and get X and change next_id to
> > > > > > X+1
> > > > > atomically.
> > > > > > -  context switch
> > > > > > - Thread 0 validate X by atomic_read and succeed to take
> ownership.
> > > > > > - The system loosed the port(or will be managed by two
> > > > > > entities) -
> > > crash.
> > > > >
> > > > >
> > > > > Ok, and how using lock will protect you with such scenario?
> > > >
> > > > The owner set API validation by thread 0 should fail because the
> > > > owner
> > > validation is included in the protected section.
> > >
> > > Then your validation function would fail even if you'll use atomic
> > > ops instead of lock.
> > No.
> > With atomic this specific scenario will cause the validation to pass.
> 
> Can you explain to me how?
> 
> rte_eth_is_valid_owner_id(uint16_t owner_id) {
>               int32_t cur_owner_id = RTE_MIN(rte_atomic32_get(next_owner_id),
> UINT16_MAX);
> 
> 	if (owner_id == RTE_ETH_DEV_NO_OWNER || owner >
> cur_owner_id) {
> 		RTE_LOG(ERR, EAL, "Invalid owner_id=%d.\n", owner_id);
> 		return 0;
> 	}
> 	return 1;
> }
> 
> Let say your next_owne_id==X, and you invoke
> rte_eth_is_valid_owner_id(owner_id=X+1)  - it would fail.

Explanation:
The scenario with locks:
next_owner_id = X.
Thread 0 call to set API(with invalid owner Y=X) and take lock.
Context switch.
Thread 1 call to owner_new and stuck in the lock.
Context switch.
Thread 0 does owner id validation and failed(Y>=X) - unlock the lock and return failure to the user.
Context switch.
Thread 1 take the lock and update X to X+1, then, unlock the lock.
Everything is OK!

The same scenario with atomics:
next_owner_id = X.
Thread 0 call to set API(with invalid owner Y=X) and take lock.
Context switch.
Thread 1 call to owner_new and change X to X+1(atomically).
Context switch.
Thread 0 does owner id validation and success(Y<(atomic)X+1) - unlock the lock and return success to the  user.
Problem!
 
> > With lock no next_id changes can be done while the thread is in the set
> API.
> >
> > > But in fact your code is not protected for that scenario - doesn't
> > > matter will you'll use lock or atomic ops.
> > > Let's considerer your current code with the following scenario:
> > >
> > > next_owner_id  == 1
> > > 1) Process 0:
> > >      rte_eth_dev_owner_new(&owner_id);
> > >      now owner_id == 1 and next_owner_id == 2
> > > 2) Process 1 (by mistake):
> > >     rte_eth_dev_owner_set(port_id=1, owner->id=1); It will complete
> > > successfully, as owner_id ==1 is considered as valid.
> > > 3) Process 0:
> > >       rte_eth_dev_owner_set(port_id=1, owner->id=1); It will also
> > > complete with success, as owner->id is valid is equal to current port
> owner_id.
> > > So you finished with 2 processes assuming that they do own
> > > exclusively then same port.
> > >
> > > Honestly in that situation  locking around nest_owner_id wouldn't
> > > give you any advantages over atomic ops.
> > >
> >
> > This is a different scenario that we can't protect on it with atomic or locks.
> > But for the first scenario I described I think we can.
> > Please read it again, I described it step by step.
> >
> > > >
> > > > > I don't think you can protect yourself against such scenario
> > > > > with or without locking.
> > > > > Unless you'll make it harder for the mis-behaving thread to
> > > > > guess valid owner_id, or add some extra logic here.
> > > > >
> > > > > >
> > > > > >
> > > > > > > > The set(and others) ownership APIs already uses the
> > > > > > > > ownership lock so I
> > > > > > > think it makes sense to use the same lock also in ID allocation.
> > > > > > > >
> > > > > > > > > > > > > In fact, for next_owner_id, you don't need a
> > > > > > > > > > > > > lock - just rte_atomic_t should be enough.
> > > > > > > > > > > >
> > > > > > > > > > > > I don't think so, it is problematic in
> > > > > > > > > > > > next_owner_id wraparound and may
> > > > > > > > > > > complicate the code in other places which read it.
> > > > > > > > > > >
> > > > > > > > > > > IMO it is not that complicated, something like that
> > > > > > > > > > > should work I
> > > > > think.
> > > > > > > > > > >
> > > > > > > > > > > /* init to 0 at startup*/ rte_atomic32_t *owner_id;
> > > > > > > > > > >
> > > > > > > > > > > int new_owner_id(void) {
> > > > > > > > > > >     int32_t x;
> > > > > > > > > > >     x = rte_atomic32_add_return(&owner_id, 1);
> > > > > > > > > > >     if (x > UINT16_MAX) {
> > > > > > > > > > >        rte_atomic32_dec(&owner_id);
> > > > > > > > > > >        return -EOVERWLOW;
> > > > > > > > > > >     } else
> > > > > > > > > > >         return x;
> > > > > > > > > > > }
> > > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > > > > Why not just to keep it simple and using the same lock?
> > > > > > > > > > >
> > > > > > > > > > > Lock is also fine, I just think it better be a separate
> > > > > > > > > > > one
> > > > > > > > > > > - that would protext just next_owner_id.
> > > > > > > > > > > Though if you are going to use uuid here - all that
> > > > > > > > > > > probably not relevant any more.
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > I agree about the uuid but still think the same lock
> > > > > > > > > > should be used for
> > > > > > > both.
> > > > > > > > >
> > > > > > > > > But with uuid you don't need next_owner_id at all, right?
> > > > > > > > > So lock will only be used for rte_eth_dev_data[] fields
> anyway.
> > > > > > > > >
> > > > > > > > Sorry, I meant uint64_t, not uuid.
> > > > > > >
> > > > > > > Ah ok, my thought uuid_t is better as with it you don't need to
> > > > > > > support your own code to allocate new owner_id, but rely on
> > > > > > > system libs
> > > > > instead.
> > > > > > > But wouldn't insist here.
> > > > > > >
> > > > > > > >
> > > > > > > > > > > > > Another alternative would be to use 2 locks - one
> > > > > > > > > > > > > for next_owner_id second for actual data[] protection.
> > > > > > > > > > > > >
> > > > > > > > > > > > > Another thing - you'll probably need to grab/release
> > > > > > > > > > > > > a lock inside
> > > > > > > > > > > > > rte_eth_dev_allocated() too.
> > > > > > > > > > > > > It is a public function used by drivers, so need to
> > > > > > > > > > > > > be protected
> > > > > too.
> > > > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > > Yes, I thought about it, but decided not to use lock in
> next:
> > > > > > > > > > > > rte_eth_dev_allocated
> > > > > > > > > > > > rte_eth_dev_count
> > > > > > > > > > > > rte_eth_dev_get_name_by_port
> > > > > rte_eth_dev_get_port_by_name
> > > > > > > > > > > > maybe more...
> > > > > > > > > > >
> > > > > > > > > > > As I can see in patch #3 you protect by lock access to
> > > > > > > > > > > rte_eth_dev_data[].name (which seems like a good
> thing).
> > > > > > > > > > > So I think any other public function that access
> > > > > > > > > > > rte_eth_dev_data[].name should be protected by the
> same
> > > lock.
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > I don't think so, I can understand to use the ownership
> > > > > > > > > > lock here(as in port
> > > > > > > > > creation) but I don't think it is necessary too.
> > > > > > > > > > What are we exactly protecting here?
> > > > > > > > > > Don't you think it is just timing?(ask in the next moment
> > > > > > > > > > and you may get another answer) I don't see optional crash.
> > > > > > > > >
> > > > > > > > > Not sure what you mean here by timing...
> > > > > > > > > As I understand rte_eth_dev_data[].name unique identifies
> > > > > > > > > device and is used by  port allocation/release/find functions.
> > > > > > > > > As you stated above:
> > > > > > > > > "1. The port allocation and port release synchronization
> > > > > > > > > will be managed by ethdev."
> > > > > > > > > To me it means that ethdev layer has to make sure that all
> > > > > > > > > accesses to rte_eth_dev_data[].name are atomic.
> > > > > > > > > Otherwise what would prevent the situation when one
> process
> > > > > > > > > does
> > > > > > > > > rte_eth_dev_allocate()->snprintf(rte_eth_dev_data[x].name,
> > > > > > > > > ...) while second one does
> > > > > > > rte_eth_dev_allocated(rte_eth_dev_data[x].name, ...) ?
> > > > > > > > >
> > > > > > > > The second will get True or False and that is it.
> > > > > > >
> > > > > > > Under race condition - in the worst case it might crash, though
> > > > > > > for that you'll have to be really unlucky.
> > > > > > > Though in most cases as you said it would just not operate
> correctly.
> > > > > > > I think if we start to protect dev->name by lock we need to do
> > > > > > > it for all instances (both read and write).
> > > > > > >
> > > > > > Since under the ownership rules, the user must take ownership of a
> > > > > > port
> > > > > before using it, I still don't see a problem here.
> > > > >
> > > > > I am not talking about owner id or name here.
> > > > > I am talking about dev->name.
> > > > >
> > > > So? The user still should take ownership of a device before using it (by
> > > name or by port id).
> > > > It can just read it without owning it, but no managing it.
> > > >
> > > > > > Please, Can you describe specific crash scenario and explain how
> > > > > > could the
> > > > > locking fix it?
> > > > >
> > > > > Let say thread 0 doing rte_eth_dev_allocate()-
> > > > > >snprintf(rte_eth_dev_data[x].name, ...), thread 1 doing
> > > > > rte_pmd_ring_remove()->rte_eth_dev_allocated()->strcmp().
> > > > > And because of race condition - rte_eth_dev_allocated() will return
> > > > > rte_eth_dev * for the wrong device.
> > > > Which wrong device do you mean? I guess it is the device which
> currently is
> > > being created by thread 0.
> > > > > Then rte_pmd_ring_remove() will call rte_free() for related
> > > > > resources, while It can still be in use by someone else.
> > > > The rte_pmd_ring_remove caller(some DPDK entity) must take
> ownership
> > > > (or validate that he is the owner) of a port before doing it(free,
> release), so
> > > no issue here.
> > >
> > > Forget about ownership for a second.
> > > Suppose we have a process it created ring port for itself (without setting
> any
> > > ownership)  and used it for some time.
> > > Then it decided to remove it, so it calls rte_pmd_ring_remove() for it.
> > > At the same time second process decides to call rte_eth_dev_allocate()
> (let
> > > say for anither ring port).
> > > They could collide trying to read (process 0) and modify (process 1) same
> > > string rte_eth_dev_data[].name.
> > >
> > Do you mean that process 0 will compare successfully the process 1 new
> port name?
> 
> Yes.
> 
> > The state are in local process memory - so process 0 will not compare the
> process 1 port, from its point of view this port is in UNUSED
> > state.
> >
> 
> Ok, and why it can't be in attached state in process 0 too?

Someone in process 0 should attach it using protected attach_secondary somewhere in your scenario.


> Konstantin
> 
> > > Konstantin
> > >
> > > >
> > > >
> > > > Also I'm not sure I fully understand your scenario looks like moving
> > > > the device state setting in allocation to be after the name setting will be
> > > good.
> > > > What do you think?
> > > >
> > > > > Konstantin
> > > > >
> > > > > >
> > > > > > > > Maybe if it had been called just a moment after, It might get
> > > > > > > > different
> > > > > > > answer.
> > > > > > > > Because these APIs don't change ethdev structure(just read),
> > > > > > > > it can be
> > > > > OK.
> > > > > > > > But again, I can understand to use ownership lock also here.
> > > > > > > >
> > > > > > >
> > > > > > > Konstantin