Previously, we would add exclusive routes via netlink message flags
NLM_F_CREATE | NLM_F_REPLACE for RTM_NEWROUTE. Similar to `ip route replace`.
Using that form of RTM_NEWROUTE message, we could only add a certain
route with a certain network/plen,metric triple once. That was already
hugely inconvenient, because
- when configuring routes, multiple (managed) interfaces may get
conflicting routes (multihoming). Only one of the routes can be actually
configured using `ip route replace`, so we need to track routes that are
currently shadowed.
- when configuring routes, we might replace externally configured
routes on unmanaged interfaces. We should not interfere with such
routes.
That was worked around by having NMRouteManager (and NMDefaultRouteManager).
NMRouteManager would keep a list of the routes which NetworkManager would like
to configure, even if momentarily being unable to do so due to conflicting routes.
This worked mostly well but was complicated. It involved bumping metrics to
avoid conflicts for device routes, as we might require them for gateway routes.
Drop that now. Instead, use the corresponding of `ip route append` to configure
routes. This allows NetworkManager to confiure (almost) all routes that we care.
Especially, it can configure all routes on a managed interface, without
replacing/interfering with routes on other interfaces. Hence, NMRouteManager
becomes obsolete.
It practice it is a bit more complicated because:
- when adding an IPv4 address, kernel will automatically create a device route
for the subnet. We should avoid that by using the IFA_F_NOPREFIXROUTE flag for
IPv4 addresses (still to-do). But as kernel may not support that flag for IPv4
addresses yet (and we don't require such a kernel yet), we still need functionality
similar to nm_route_manager_ip4_route_register_device_route_purge_list().
This functionality is now handled via nm_platform_ip4_dev_route_blacklist_set().
- trying to configure an IPv6 route with a source address will be rejected
by kernel as long as the address is tentative (see related bug rh#1457196).
Preferably, NMDevice would keep the list of routes which should be configured,
while kernel would have the list of what actually is configured. There is a
feed-back loop where both affect each other (for example, when externally deleting
a route, NMDevice must forget about it too). Previously, NMRouteManager would have
the task of remembering all routes which we currently want to configure, but cannot
due to conflicting routes.
We get rid of that, because now we configure non-exclusive routes. We however still
will need to remember IPv6 routes with a source address, that currently cannot be
configured yet. Hence, we will need to keep track of routes that
currently cannot be configured, but later may be.
That is still not done yet, as NMRouteManager didn't handle this
correctly either.
For completeness of the API. remove_obj() is basically a shortcut
of nm_dedup_multi_index_lookup_obj() combined with
nm_dedup_multi_index_remove_entry(). As such, it is useful to return
the actually deleted object. Note that the lookup needle @obj is not
necessarily the same instance as the one that will be removed, it's
only an instance that compares equal according to the index's equality
operator.
The return value shall indicate whether the add-call changed anything.
Reordering shall count as a change too.
On the other hand, clearing the dirty flag of the entry does not count
as a change.
Reasons:
- it adds an O(1) lookup index for accessing NMIPxConfig's addresses.
Hence, operations like merge/intersect have now runtime O(n) instead
of O(n^2).
Arguably, we expect low numbers of addresses in general. For low
numbers, the O(n^2) doesn't matter and quite likely in those cases
the previous implementation was just fine -- maybe even faster.
But the simple case works fine either way. It's important to scale
well in the exceptional case.
- the tracked objects can be shared between the various NMPI4Config,
NMIP6Config instances with NMPlatform and everybody else.
- the NMPObject can be treated generically, meaning it enables code to
handle both IPv4 and IPv6, or addresses and routes. See for example
_nm_ip_config_add_obj().
- I want core to evolve to somewhere where we don't keep copies of
NMPlatformIP4Address, et al. instances. Instead they shall all be
shared. I hope this will reduce memory consumption (although tracking a
reference consumes some memory too). Also, it shortcuts nmp_object_equal()
when comparing the same object. Calling nmp_object_equal() on the
identical objects would be a common case after the hash function
pre-evaluates equality.
And get rid of the unused obj_full_equality_allows_different_class.
It's hard to grasp how to implement different object types that can compare
despite having different klasses. The idea was, that stack allocated
objects (used as lookup needles), are some small lightweight objects,
that still compare equal to the full instance. But it's unused. Drop it.
by moving the core functionality to "nm-dedup-multi.c".
As the ref-counting mechanism now is part of "nm-dedup-multi.c",
this works better and is reusable outside of platform.
Implement the reference counting of NMPObject as part of
NMDedupMultiObj and get rid of NMDedupMultiBox.
With this change, the NMPObject is aware in which NMDedupMultiIndex
instance it is tracked.
- this saves an additional GSlice allocation for the NMDedupMultiBox.
- it is immediately known, whether an NMPObject is tracked by a
certain NMDedupMultiIndex or not. This saves an additional hash
lookup.
- previously, when all idx-types cease to reference an NMDedupMultiObj
instance, it was removed. Now, a tracked objects stays in the
NMDedupMultiIndex until it's last reference is deleted. This possibly
extends the lifetime of the object and we may reuse it better.
- it is no longer possible to add one object to more then one
NMDedupMultiIndex instance. As we anyway want to have only one
instance to deduplicate the objects, this is fine.
- the ref-counting implementation is now part of NMDedupMultiObj.
Previously, NMDedupMultiIndex could also track objects that were
not ref-counted. Hoever, the object anyway *must* implement the
NMDedupMultiObj API, so this flexibility is unneeded and was not
used.
- a downside is, that NMPObject grows by one pointer size, even if
it isn't tracked in the NMDedupMultiIndex. But we really want to
put all objects into the index for sharing and deduplication. So
this downside should be acceptable. Still, code like
nmp_object_stackinit*() needs to handle a larger object.
Add the NMDedupMultiIndex cache. It basically tracks
objects as doubly linked list. With the addition that
each object and the list head is indexed by a hash table.
Also, it supports tracking multiple distinct lists,
all indexed by the idx-type instance.
It also deduplicates the tracked objects and shares them.
- the objects that can be put into the cache must be immutable
and ref-counted. That is, the cache will deduplicate them
and share the reference. Also, as these objects are immutable
and ref-counted, it is safe that users outside the cache
own them too (as long as they keep them immutable and manage
their reference properly).
The deduplication uses obj_id_hash_func() and obj_id_equal_func().
These functions must cover *every* aspect of the objects when
comparing equality. For example nm_platform_ip4_route_cmp()
would be a function that qualifies as obj_id_equal_func().
The cache creates references to the objects as needed and
gives them back. This happens via obj_get_ref() and
obj_put_ref(). Note that obj_get_ref() is free to create
a new object, for example to convert a stack-allocated object
to a (ref-counted) heap allocated one.
The deduplication process creates NMDedupIndexBox instances
which are the ref-counted entity. In principle, the objects
themself don't need to be ref-counted as that is handled by
the boxing instance.
- The cache doesn't only do deduplication. It is a multi-index,
meaning, callers add objects using a index handle NMDedupMultiIdxType.
The NMDedupMultiIdxType instance is the access handle to lookup
the list and objects inside the cache. Note that the idx-type
instance may partition the objects in distinct lists.
For all operations there are cross-references and hash table lookups.
Hence, every operation of this data structure is O(1) and the memory
overhead for an index tracking an object is constant.
The cache preserves ordering (due to linked list) and exposes the list
as public API. This allows users to iterate the list without any
additional copying of elements.
