verify() is setting an error without returning FALSE to make the
validation fail. When the parent is set, the device is a Infiniband
partition and it must have a p-key != -1.
Fixes: d595f7843e ('libnm: add libnm/libnm-core (part 1)')
(cherry picked from commit f4f1ecc7ea)
Most string properties should not accept empty strings. Add a generic
way to reject them during verify.
Add a new flag NMSettInfoProperty.direct_string_allow_empty.
Note that properties must opt-in to allow empty values. Since all
existing properties didn't have this check (but hopefully re-implemented
it in verify()), all existing properties get this flag set to TRUE.
The main point here it that new properties get the strict check by
default.
We should also review existing uses of direct_string_allow_empty,
whether the flag can be cleared. This can be done if verify() already
enforces a non-empty string, or if we accept to break behavior by
tightening up the check.
Historically, the NMSetting types were in public headers. Theoretically,
that allowed users to subtype our classes. However in practice that was
impossible because they lacked required access to internal functions to
fully create an NMSetting class outside of libnm. And it also was not
useful, because you simply cannot extend libnm by subtyping a libnm
class. And supporting such a use case would be hard and limit what we can
do in libnm.
Having GObject structs in public headers also require that we don't
change it's layout. The ABI of those structs must not change, if anybody
out there was actually subclassing our GObjects.
In libnm 1.34 (commit e46d484fae ('libnm: hide NMSetting types from
public headers')) we moved the structs from headers to internal.
This would have caused a compiler error if anybody was using those
struct definitions. However, we still didn't change the ABI/layout so
that we didn't break users who relied on it (for whatever reason).
It doesn't seem there were any affected user. We waited long enough.
Change internal ABI.
No longer use g_type_class_add_private(). Instead, embed the private
structs directly (_NM_GET_PRIVATE()) or indirectly
(_NM_GET_PRIVATE_PTR()) in the object.
The main benefit is for debugging in the debugger, where we can now
easily find the private data. Previously that was so cumbersome to be
effectively impossible.
It's also the fastest possible way, since NM_SETTING_*_GET_PRIVATE()
literally resolves to "&self->_priv" (plus static asserts and
nm_assert() for type checking).
_NM_GET_PRIVATE() also propagates constness and requires that the
argument is a compatible pointer type (at compile time).
Note that g_type_class_add_private() is also deprecated in glib 2.58 and
replaced by G_ADD_PRIVATE(). For one, we still don't rely on 2.58. Also,
G_ADD_PRIVATE() is a worse solution as it supports a usecase that we
don't care for (public structs in headers). _NM_GET_PRIVATE() is still
faster, works with older glib and most importantly: is better for
debugging as you can find the private data from an object pointer.
For NMSettingIPConfig this is rather awkward, because all direct
properties require a common "klass->private_offset". This was however
the case before this change. Nothing new here. And if you ever touch
this and do something wrong, many unit tests will fail. It's almost
impossible to get wrong, albeit it can be confusing to understand.
https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1773
Infiniband profiles can have a p-key set. Both in kernel API
("create_child" sysctl) and in NetworkManager API, that key can range
from 0x0001 to 0xFFFF (0x8000 excluded). NetworkManager does not support
renaming the interface, so kernel always assigns the interface name
"$PHYSDEV.$PKEY_ID" (with $PKEY_ID as 4 character hex digits).
Note that the highest bit in the p-key (0x8000) is the full-membership
flag. Internally, kernel only supports full-membership so when we create
for example "ib0.00c1" and "ib0.80c1" interfaces, their actually used
p-key is in both cases 0x80c1 and you can see it with `ip -d link`.
Nonetheless, kernel and NetworkManager allow to configure the p-key
without the highest bit set, and the result differs in the interface
name.
Note that initscripts' ifup-ib0 would always internally coerce the
PKEY_ID variable to have the high bit set ([1]). It also would require
that the `DEVICE=` variable is specified and matches the expected
interface name. So both these configurations are identical and valid:
DEVICE=ib0.80c1
PHYSDEV=ib0
PKEY_ID=0x80c1
and
DEVICE=ib0.80c1
PHYSDEV=ib0
PKEY_ID=0x00c1
Historically, NetworkManager would also implement the same restrictions
([2], [3], [4]). That meant, not all valid NetworkManager infiniband
profiles could be expressed as ifcfg file. For example, NetworkManager
allows to have "connection.interface-name" (`DEVICE=`) unset (which
ifup-ib and ifcfg reader did not allow). Also, NetworkManager would
allow configuring a "infiniband.p-key" without full membership flag, and
the reader would mangle that.
This caused various problems to the point that when you configure an
infiniband.p-key with a non-full-membership key, the ifcfg-rh written by
NetworkManager was invalid. Either, you could leave
"connection.interface-name" unset, but then the reader would complain
about missing `DEVICE=`. Or, we could write `DEVICE=ib0.00c1;
PKEY_ID=0x00c1`, which was invalid as we expected `DEVICE=ib0.80c1`.
This was addressed by rhbz 2122703 ([5]). The fix was to
- not require a `DEVICE=` ([6]).
- don't mangle the `PKEY_ID=` in the reader ([7]).
which happened in 1.41.2 and 1.40.2 (rhel-8.8).
With this change, we could persist any valid infiniband profile to ifcfg
format. We also could read back any valid ifcfg file that NetworkManager
would have written in the past (note that it could not write valid ifcfg
files previously, if the p-key didn't have the full-membership key set).
The problem is, that users were used to edit ifcfg files by hand, and
users would have files with:
DEVICE=ib0.80c1
PHYSDEV=ib0
PKEY_ID=0x00c1
This files had worked before, but now failed to verify as we would
expect `DEVICE=ib0.00c1`. Also, there was a change in behavior that
PKEY_ID is now interpreted without the high bit set. This is reported as
rhbz 2209164 ([8]).
We will do several things to fix that:
1) we now normalize the "connection.interface-name" to be valid. It was
not useful to set it anyway, as it was redundant. Complaining about a
redundant setting, which makes little sense to configure, is not useful.
This is done by [9].
2) we now again treat PKEY_ID= as if it had 0x8000 flag set. This was done by
[10].
With step 1) and 2), we are able to read any existing ifcfg files out
there in the way we did before 1.41.2.
There is however one piece missing. When we now create a profile using
nmcli/libnm/D-Bus, which has a non-full-membership p-key, then the
profile gets mangled in the process.
If the user uses NetworkManager API to configure an interface and
chooses a non-full-membership p-key, then this should work the same as
with keyfile plugin (or on rhel-9, where keyfile is the default). Note
that before 1.41.2 it didn't work at all, when the user used ifcfg-rh
backend. Likely(?) there are no users who rely on creating such a profile
with nmcli/libnm/D-Bus and expect to automatically have the p-key
normalized. That didn't work before 1.41.2 and didn't behave that way
between 1.41.2 and now.
This patch fixes that by introducing a new key PKEY_ID_NM= for holding
the real p-key. Now ifcfg backend is consistent with handling infiniband
profiles, and old, hand-written ifcfg files still work as before.
There is of course change in behavior, that ifcfg files between 1.41.2
and now were interpreted differently. But that is bug 2209164 ([8]) and
what we fix here.
For now strong reasons, we keep writing the PKEY_ID to file too. It's
redundant, but that is what a human might expect there.
[1] 05333c3602/f/rdma.ifup-ib (_75)
[2] https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/blob/1.40.0/src/core/settings/plugins/ifcfg-rh/nms-ifcfg-rh-reader.c#L5386
[3] cb5606cf1c (a7a78fccb2c8c945fd09038656ae734c1b0349ab_3493_3532)
[4] cb5606cf1c (a7a78fccb2c8c945fd09038656ae734c1b0349ab_3493_3506)
[5] https://bugzilla.redhat.com/show_bug.cgi?id=2122703
[6] 4c32dd9d25
[7] a4fe16a426
[8] https://bugzilla.redhat.com/show_bug.cgi?id=2209164
[9] 4610fd67e6
[10] f8e5e07355
In nm_setting_infiniband_get_virtual_interface_name(), no longer try to
detect whether the cached value is still up to date. Instead, as we now
have a fix sized buffer for the name, just always generate the name on
every call. It's simpler.
This is the same what kernel does, when the parent name is so long
that it would result in a too long overall name.
We need that the result is still a valid interface name.
NetworkManager does not support changing the interface name for
infiniband interfaces. Consequently, we verify that
"connection.interface-name" is either unset or set to the expected
"$parent.$p_key". Anything else wouldn't work anyway and is rejected as
invalid configuration. That brings problems however.
Rejecting invalid configuration seems fine at first:
$ nmcli --offline connection add type infiniband infiniband.parent ib0 infiniband.p-key 0x8010 connection.interface-name xxx
Error: Error writing connection: connection.interface-name: interface name of software infiniband device must be 'ib0.8010' or unset (instead it is 'xxx')
However, when we modify the p-key, we also get an error message:
$ nmcli --offline connection add type infiniband infiniband.parent ib0 infiniband.p-key 0x8010 connection.interface-name ib0.8010 |
nmcli --offline connection modify infiniband.p-key 5
Error: Error writing connection: connection.interface-name: interface name of software infiniband device must be 'ib0.0005' or unset (instead it is 'ib0.8010')
It's worse, because ifcfg-rh reader will mangle the PKEY_ID with |=0x8000 to set
the full membership flag. That means, if you add a profile like
$ nmcli --offline connection add type infiniband infiniband.parent ib0 infiniband.p-key 0x0010 connection.interface-name ib0.0010
it gets written to ifcfg-rh file. Then upon reload it's invalid (as the
interface name mismatches).
There are multiple solutions for this. For example, ifcfg-rh reader could also
mangle the connection.interface-name, so that the overall result is valid. Or
we could just not validate at all, and accept any bogus interface-name.
With this patch instead we will just normalize the invalid configuration to
make it right.
$ nmcli --offline connection add type infiniband infiniband.parent ib0 infiniband.p-key 0x8010 connection.interface-name ib0.8010 |
nmcli --offline connection modify infiniband.p-key 5
...
The downside is that this happens silently, so a user doesn't
notice that configuration is ignored:
$ nmcli --offline connection add type infiniband infiniband.parent ib0 infiniband.p-key 0x8010 connection.interface-name foo
...
interface-name=ib0.8010
This approach still seems preferable, because setting
"connection.interface-name" for infiniband profiles makes little sense,
so what we care here is to avoid problems.
A virtual infiniband profile (with p-key>=0) can also contain a
"connection.interface-name". But it is required to match the
f"{parent}.{p-key}" format.
However, such a profile can also set "mac_address" instead of "parent".
In that case, the validation code was crashing.
nmcli connection add type infiniband \
infiniband.p-key 6 \
infiniband.mac-address 52:54:00:86:f4:eb:aa:aa:aa:aa:52:54:00:86:f4:eb:aa:aa:aa:aa \
connection.interface-name aaaa
The crash was introduced by commit 99d898cf1f ('libnm: rework caching
of virtual-iface-name for infiniband setting'). Previously, it would not
have crashed, because we just called
g_strdup_printf("%s.%04x", priv->parent, priv->p_key)
with a NULL string. It would still not have validated the connection
and passing NULL as string to printf is wrong. But in practice, it
would have worked mostly fine for users.
Fixes: 99d898cf1f ('libnm: rework caching of virtual-iface-name for infiniband setting')
(cherry picked from commit fd5945b408)
We cache the virtual-iface-name. The caching is also part of the API as
nm_setting_infiniband_get_virtual_interface_name() returns a const
string.
As the value is computed and based on the parent and the p-key, we must
clear the cache when the parent or p-key changes (or detect that it's
invalid).
Previously, we were simply clearing the value in the set_property() function,
which is the only setter of these two properties. If we make these
properties "direct properties", then they will be directly set via
from_dbus_fcn() which bypasses the GObject setter. Which is a problem
for the cache invalidation.
We could either not make those properties direct properties. The problem
is that direct properties are nice, and they will in the future
implement further optimizations for them. Also, they are the default
implementation, and it seems clearer to build something on top of that,
instead of deviating from the default.
Instead, let the caching detect when the value needs to be regenerated.
We use clang-format for automatic formatting of our source files.
Since clang-format is actively maintained software, the actual
formatting depends on the used version of clang-format. That is
unfortunate and painful, but really unavoidable unless clang-format
would be strictly bug-compatible.
So the version that we must use is from the current Fedora release, which
is also tested by our gitlab-ci. Previously, we were using Fedora 34 with
clang-tools-extra-12.0.1-1.fc34.x86_64.
As Fedora 35 comes along, we need to update our formatting as Fedora 35
comes with version "13.0.0~rc1-1.fc35".
An alternative would be to freeze on version 12, but that has different
problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it
would be cumbersome for our developers which are on Fedora 35 to use a
clang that they cannot easily install).
The (differently painful) solution is to reformat from time to time, as we
switch to a new Fedora (and thus clang) version.
Usually we would expect that such a reformatting brings minor changes.
But this time, the changes are huge. That is mentioned in the release
notes [1] as
Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353)
[1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
Certain properties need to release memory when destroying the NMSetting.
For "direct" properties, we have all the information we need to do that
generically in the NMSetting base class. In practice, this only concerns
string properties.
See _finalize_direct() in "nm-setting.c".
However, if the NMSetting base class takes care of freeing the strings,
then the subclasses must not also unref the variable (to avoid double free).
Previously, subclasses had to opt-in for the base class to indicate that
they are fine with that.
Now, let the base class always handle it. We only need to make sure that
classes that implement direct string properties don't also try to free
the values during destruction.
A MAC address is a relatively common "type". The GObject property is of type string,
but the D-Bus type is a bytestring ("ay"). We will need a special NMSettInfoPropertType.
Note that like most implementations, the from-dbus implementation still is based
on GObject setters. This will change in the future.
Also note that the previous compare function was
_nm_setting_property_compare_fcn_default(). That is, it used to convert
the property to GVariant and compare those. The conversion to GVariant
in that case normalizes the string (e.g. it is case insensitive). Also,
only properties could be compared which were also convertible to D-Bus
(which is probably fine, because there is no guarantee the profiles that
don't verify can be compared).
The code now uses the direct comparison of the strings. That mostly
preserves the case-insensitivity of the previous comparison, because
the property setters for mac addresses all use
_nm_utils_hwaddr_canonical_or_invalid() to normalize the strings.
This is subtle, but still correct. Note that this will improve later,
by ensuring that the property setters for mac addresses automatically
perform the right normalization.
NMSetting instances either have no private data, they use
g_type_add_class_private(), or they embed the private data in the
NMSetting struct.
In all cases, we can find the private data at a fixed offset. Track that
offset in the NMSettInfoSetting meta data.
This will be useful, because properties really are stored in simple
fields, like a boolean property can be stored in a "bool" field. We will
extend the property meta data to track the offset of this property
field, but we also need to know where the offset starts.
When subclassing a GObject type, the class and object structs
must be available and defined in the header.
For libnm, and in particular for NMSetting classes, we don't want
users to subclass NMSetting. It also doesn't work, because libnm
has internal code that is necessary to hook up the NMSetting class.
You cannot define your own type and make it work together with
libnm.
Having the structs in public headers limits what we can do with them.
For example, we could embed the private data directly in the structures
and avoid the additional indirection.
This is an API break, but for something that most likely nobody cares
about. Or better, nobody should care about. API is not what is
accidentally defined in a header, API was the library provides to
meaningfully use. Subclassing these types is not meaningful and was
only accidentally possible so far.
Only hide the structs for now. More cleanup is possible later. We shall
however aim to keep the padding and struct layout to not also break ABI.
(cherry picked from commit e46d484fae)
"libnm-core/" is rather complicated. It provides a static library that
is linked into libnm.so and NetworkManager. It also contains public
headers (like "nm-setting.h") which are part of public libnm API.
Then we have helper libraries ("libnm-core/nm-libnm-core-*/") which
only rely on public API of libnm-core, but are themself static
libraries that can be used by anybody who uses libnm-core. And
"libnm-core/nm-libnm-core-intern" is used by libnm-core itself.
Move "libnm-core/" to "src/". But also split it in different
directories so that they have a clearer purpose.
The goal is to have a flat directory hierarchy. The "src/libnm-core*/"
directories correspond to the different modules (static libraries and set
of headers that we have). We have different kinds of such modules because
of how we combine various code together. The directory layout now reflects
this.
2021-02-18 19:46:51 +01:00
Renamed from libnm-core/nm-setting-infiniband.c (Browse further)
