The abbreviations "ns" and "ms" seem not very clear to me. Spell them
out to nsec/msec. Also, in parts we already used the longer abbreviations,
so it wasn't consistent.
- systemd-networkd and initscripts both support it.
- it seems suggested to configure routes with scope "link" on AWS.
- the scope is only supported for IPv4 routes. Kernel ignores the
attribute for IPv6 routes.
- we don't support the aliases like "link" or "global". Instead
only the numeric value is supported. This is different from
systemd-networkd, which accepts names like "global" and "link",
but no numerical values. I think restricting ourself only to
the aliases unnecessarily limits what is possible on netlink.
The alternative would be to allow aliases and numbers both,
but that causes multiple ways to define something and has
thus downsides. So, only numeric values.
- when setting rtm_scope to RT_SCOPE_NOWHERE (0, the default), kernel
will coerce that to RT_SCOPE_LINK. This ambiguity of nowhere vs. link
is a problem, but we don't do anything about it.
- The other problem is, that when deleting a route with scope RT_SCOPE_NOWHERE,
this acts as a wild care and removes the first route that matches (given the
other route attributes). That means, NetworkManager has no meaningful
way to delete a route with scope zero, there is always the danger that
we might delete the wrong route. But this is nothing new to this
patch. The problem existed already previously, except that
NetworkManager could only add routes with scope nowhere (i.e. link).
Track whether IP addresses were added by NM or externally. In this way
it becomes possible in a later commit to add prefix route only for
addresses added by NM.
IPv6 router advertisement messages contain the following parameters
(RFC 4861):
- Reachable time: 32-bit unsigned integer. The time, in
milliseconds, that a node assumes a neighbor is reachable after
having received a reachability confirmation. Used by the Neighbor
Unreachability Detection algorithm. A value of zero means
unspecified (by this router).
- Retrans Timer: 32-bit unsigned integer. The time, in milliseconds,
between retransmitted Neighbor Solicitation messages. Used by
address resolution and the Neighbor Unreachability Detection
algorithm. A value of zero means unspecified (by this router).
Currently NM ignores them; however, since it leaves accept_ra=1, the
kernel parses RAs and applies those parameters for us [1].
In the next commit kernel handling of RAs will be disabled, so let NM
set those neighbor-related parameters.
[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/net/ipv6/ndisc.c?h=v5.2#n1353
Drop nm_platform_link_get_address_as_bytes() and introduce
nmp_link_address_get_as_bytes() so that it becomes possible to obtain
also the broadcast address without an additional lookup of the link.
We no longer add these. If you use Emacs, configure it yourself.
Also, due to our "smart-tab" usage the editor anyway does a subpar
job handling our tabs. However, on the upside every user can choose
whatever tab-width he/she prefers. If "smart-tabs" are used properly
(like we do), every tab-width will work.
No manual changes, just ran commands:
F=($(git grep -l -e '-\*-'))
sed '1 { /\/\* *-\*- *[mM]ode.*\*\/$/d }' -i "${F[@]}"
sed '1,4 { /^\(#\|--\|dnl\) *-\*- [mM]ode/d }' -i "${F[@]}"
Check remaining lines with:
git grep -e '-\*-'
The ultimate purpose of this is to cleanup our files and eventually use
SPDX license identifiers. For that, first get rid of the boilerplate lines.
While at it, rename the "addr" field to "l_address". The term "addr" is
used over and over. Instead we should use distinct names that make it
easier to navigate the code.
clang (3.4.2-9.el7) on CentOS 7.6 fails related to nm_hash_update_vals().
I am not even quoting the error message, it's totally non-understandable.
nm_hash_update_vals() uses typeof(), and in some obscure cases, clang dislikes
when the argument itself is some complex macro. I didn't fully understand why,
but this works around it.
I would prefer to fix nm_hash_update_vals() to not have this limitation.
But I don't know how.
There is probably no downside to have this an inline function instead of
a macro.
When changing the number of VFs the kernel can block for very long
time in the write() to sysfs, especially if autoprobe-drivers is
enabled. Turn the nm_platform_link_set_sriov_params() into an
asynchronous function.
In general, all fields of public NMPlatform* structs must be
plain/simple. Meaning: copying the struct must be possible without
caring about cloning/duplicating memory.
In other words, if there are fields which lifetime is limited,
then these fields cannot be inside the public part NMPlatform*.
That is why
- "NMPlatformLink.kind", "NMPlatformQdisc.kind", "NMPlatformTfilter.kind"
are set by platform code to an interned string (g_intern_string())
that has a static lifetime.
- the "ingress_qos_map" field is inside the ref-counted struct NMPObjectLnkVlan
and not NMPlatformLnkVlan. This field requires managing the lifetime
of the array and NMPlatformLnkVlan cannot provide that.
See also for example NMPClass.cmd_obj_copy() which can deep-copy an object.
But this is only suitable for fields in NMPObject*. The purpose of this
rule is that you always can safely copy a NMPlatform* struct without
worrying about the ownership and lifetime of the fields (the field's
lifetime is unlimited).
This rule and managing of resource lifetime is the main reason for the
NMPlatform*/NMPObject* split. NMPlatform* structs simply have no mechanism
for copying/releasing fields, that is why the NMPObject* counterpart exists
(which is ref-counted and has a copy and destructor function).
This is violated in tc_commit() for the "kind" strings. The lifetime
of these strings is tied to the setting instance.
We cannot intern the strings (because these are arbitrary strings
and interned strings are leaked indefinitely). We also cannot g_strdup()
the strings, because NMPlatform* is not supposed to own strings.
So, just add comments that warn about this ugliness.
The more correct solution would be to move the "kind" fields inside
NMPObjectQdisc and NMPObjectTfilter, but that is a lot of extra effort.
Arguably, the structure is used inside a union with another (larger)
struct, hence no memory is saved.
In fact, it may well be slower performance wise to access a boolean bitfield
than a gboolean (int).
Still, boolean fields in structures should be bool:1 bitfields for
consistency.
Kernel calls the netlink attribute TCA_FQ_CODEL_MEMORY_LIMIT. Likewise,
iproute2 calls this "memory_limit".
Rename because TC parameters are inherrently tied to the kernel
implementation and we should use the familiar name.
iproute2 uses the special value ~0u to indicate not to set
TCA_FQ_CODEL_CE_THRESHOLD in RTM_NEWQDISC. When not explicitly
setting the value, kernel treats the threshold as disabled.
However note that 0xFFFFFFFFu is not an invalid threshold (as far as
kernel is concerned). Thus, we should not use that as value to indicate
that the value is unset. Note that iproute2 uses the special value ~0u
only internally thereby making it impossible to set the threshold to
0xFFFFFFFFu). But kernel does not have this limitation.
Maybe the cleanest way would be to add another field to NMPlatformQDisc:
guint32 ce_threshold;
bool ce_threshold_set:1;
that indicates whether the threshold is enable or not.
But note that kernel does:
static void codel_params_init(struct codel_params *params)
{
...
params->ce_threshold = CODEL_DISABLED_THRESHOLD;
static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
struct netlink_ext_ack *extack)
{
...
if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
}
static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
{
...
if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
codel_time_to_us(q->cparams.ce_threshold)))
goto nla_put_failure;
This means, kernel internally uses the special value 0x83126E97u to indicate
that the threshold is disabled (WTF). That is because
(((guint64) 0x83126E97u) * NSEC_PER_USEC) >> CODEL_SHIFT == CODEL_DISABLED_THRESHOLD
So in kernel API this value is reserved (and has a special meaning
to indicate that the threshold is disabled). So, instead of adding a
ce_threshold_set flag, use the same value that kernel anyway uses.
The memory-limit is an unsigned integer. It is ugly (if not wrong) to compare unsigned
values with "-1". When comparing with the default value we must also use an u32 type.
Instead add a define NM_PLATFORM_FQ_CODEL_MEMORY_LIMIT_UNSET.
Note that like iproute2 we treat NM_PLATFORM_FQ_CODEL_MEMORY_LIMIT_UNSET
to indicate to not set TCA_FQ_CODEL_MEMORY_LIMIT in RTM_NEWQDISC. This
special value is entirely internal to NetworkManager (or iproute2) and
kernel will then choose a default memory limit (of 32MB). So setting
NM_PLATFORM_FQ_CODEL_MEMORY_LIMIT_UNSET means to leave it to kernel to
choose a value (which then chooses 32MB).
See kernel's net/sched/sch_fq_codel.c:
static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
struct netlink_ext_ack *extack)
{
...
q->memory_limit = 32 << 20; /* 32 MBytes */
static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
struct netlink_ext_ack *extack)
...
if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
Note that not having zero as default value is problematic. In fields like
"NMPlatformIP4Route.table_coerced" and "NMPlatformRoutingRule.suppress_prefixlen_inverse"
we avoid this problem by storing a coerced value in the structure so that zero is still
the default. We don't do that here for memory-limit, so the caller must always explicitly
set the value.
Next we will need to detect more kernel features. First refactor the
handling of these to require less code changes and be more efficient.
A plain nm_platform_kernel_support_get() only reqiures to access an
array in the common case.
The other important change is that the function no longer requires a
NMPlatform instance. This allows us to check kernel support from
anywhere. The only thing is that we require kernel support to be
initialized before calling this function. That means, an NMPlatform
instance must have detected support before.
(cherry picked from commit ee269b318e)
In some cases it is convenient to specify ranges of bridge vlans, as
already supported by iproute2 and natively by kernel. With this commit
it becomes possible to add a range in this way:
nmcli connection modify eth0-slave +bridge-port.vlans "100-200 untagged"
vlan ranges can't be PVIDs because only one PVID vlan can exist.
https://bugzilla.redhat.com/show_bug.cgi?id=1652910
(cherry picked from commit 7093515777)
Add and implement NMPlatformRoutingRule types and let the platform cache
handle rules.
Rules are special in two ways:
- they don't have an ifindex. That makes them different from all other
currently existing NMPlatform* types, which have an "ifindex" field and
"implement" NMPlatformObjWithIfindex.
- they have an address family, but contrary to addresses and routes, there
is only one NMPlatformRoutingRule object to handle both address
families.
Both of these points require some special considerations.
Kernel treats routing-rules quite similar to routes. That is, kernel
allows to add different rules/routes, as long as they differ in certain
fields. These "fields" make up the identity of the rules/routes. But
in practice, it's not defined which fields contribute to the identity
of these objects. That makes using the netlink API very hard. For
example, when kernel gains support for a new attribute which
NetworkManager does not know yet, then users can add two rules/routes
that look the same to NetworkManager. That can easily result in cache
inconsistencies.
Another problem is, that older kernel versions may not yet support all
fields, which NetworkManager (and newer kernels) considers for identity.
The older kernel will not simply reject netlink messages with these unknown
keys, instead it will proceed adding the route/rule without it. That means,
the added route/rule will have a different identity than what NetworkManager
intended to add.
The function is unused. It would require redesign to work with
future changes, and since it's unused, just drop it.
The long reasoning is:
Currently, a refresh-all is tied to an NMPObjectType. However, with
NMPObjectRoutingRule (for policy-routing-rules) that will no longer
be the case.
That is because NMPObjectRoutingRule will be one object type for
AF_INET and AF_INET6. Contrary to IPv4 addresses and routes, where
there are two sets of NMPObject types.
The reason is, that it's preferable to treat IPv4 and IPv6 objects
similarly, that is: as the same type with an address family property.
That also follows netlink, which uses RTM_GET* messages for both
address families, and the address family is expressed inside the
message.
But then an API like nm_platform_refresh_all() makes little sense,
it would require at least an addr_family argument. But since the
API is unused, just drop it.
Until now, all implemented NMPObject types have an ifindex field (from
links, addresses, routes, qdisc to tfilter).
The NMPObject structure contains a union of all available types, that
makes it easier to down-case from an NMPObject pointer to the actual
content.
The "object" field of NMPObject of type NMPlatformObject is the lowest
common denominator.
We will add NMPlatformRoutingRules (for policy routing rules). That type
won't have an ifindex field.
Hence, drop the "ifindex" field from NMPlatformObject type. But also add
a new type NMPlatformObjWithIfindex, that can represent all types that
have an ifindex.
We will need more flags.
WireGuard internal tools solve this by embedding the change flags inside
the structure that corresponds to NMPlatformLnkWireGuard. We don't do
that, NMPlatformLnkWireGuard is only for containing the information about
the link.
