Before this change, the join count and the state for IGMP/MLD was held
across different types which required multiple locks to be held when
accessing a multicast group's state.
Bug #4682, #4861Fixes#4916
PiperOrigin-RevId: 345019091
Fixing the sendto deadlock exposed yet another deadlock where a lock inversion
occurs on the handleControlPacket path where e.mu and demuxer.epsByNIC.mu are
acquired in reverse order from say when RegisterTransportEndpoint is called
in endpoint.Connect().
This fix sidesteps the issue by just making endpoint.state an atomic and gets rid
of the need to acquire e.mu in e.HandleControlPacket.
PiperOrigin-RevId: 344939895
Currently, if containerd is installed locally via tools/installers/containerd,
then it will not necessarily be used if containerd is installed in the system
path. This means that the existing containerd tests are all likely broken.
Also, use libbtrfs-dev instead of btrfs-tools, which is not actually required.
PiperOrigin-RevId: 344879109
Ports the following options:
- TCP_NODELAY
- TCP_CORK
- TCP_QUICKACK
Also deletes the {Get/Set}SockOptBool interface methods from all implementations
PiperOrigin-RevId: 344378824
We will use SocketOptions for all kinds of options, not just SOL_SOCKET options
because (1) it is consistent with Linux which defines all option variables on
the top level socket struct, (2) avoid code complexity. Appropriate checks
have been added for matching option level to the endpoint type.
Ported the following options to this new utility:
- IP_MULTICAST_LOOP
- IP_RECVTOS
- IPV6_RECVTCLASS
- IP_PKTINFO
- IP_HDRINCL
- IPV6_V6ONLY
Changes in behavior (these are consistent with what Linux does AFAICT):
- Now IP_MULTICAST_LOOP can be set for TCP (earlier it was a noop) but does not
affect the endpoint itself.
- We can now getsockopt IP_HDRINCL (earlier we would get an error).
- Now we return ErrUnknownProtocolOption if SOL_IP or SOL_IPV6 options are used
on unix sockets.
- Now we return ErrUnknownProtocolOption if SOL_IPV6 options are used on non
AF_INET6 endpoints.
This change additionally makes the following modifications:
- Add State() uint32 to commonEndpoint because both tcpip.Endpoint and
transport.Endpoint interfaces have it. It proves to be quite useful.
- Gets rid of SocketOptionsHandler.IsListening(). It was an anomaly as it was
not a handler. It is now implemented on netstack itself.
- Gets rid of tcp.endpoint.EndpointInfo and directly embeds
stack.TransportEndpointInfo. There was an unnecessary level of embedding
which served no purpose.
- Removes some checks dual_stack_test.go that used the errors from
GetSockOptBool(tcpip.V6OnlyOption) to confirm some state. This is not
consistent with the new design and also seemed to be testing the
implementation instead of behavior.
PiperOrigin-RevId: 344354051
...as defined by RFC 2710. Querier (router)-side MLDv1 is not yet
supported.
The core state machine is shared with IGMPv2.
This is guarded behind a flag (ipv6.Options.MLDEnabled).
Tests: ip_test.TestMGP*
Bug #4861
PiperOrigin-RevId: 344344095
Multiple goroutines may use the same stack.Route concurrently so
the stack.Route should make sure that any functions called on it
are thread-safe.
Fixes#4073
PiperOrigin-RevId: 344320491
To create DUTs in parallel, we need to create goroutines to do the setup. The
old code base has a lot of t.Fatal(f) usage in those setup functions which is
not great for this change: "FailNow must be called from the goroutine running
the test or benchmark function, not from other goroutines created during the
test" (https://golang.org/pkg/testing/#T.FailNow).
- Cleanup all t.Fatal(f) usage in DUT.Prepare()
- use goroutines to create DUTs in parallel
PiperOrigin-RevId: 344275809
Because the code handles a bad header as "payload" right up to the last moment
we need to make sure payload handling does not remove the error information.
Fixes#4909
PiperOrigin-RevId: 344141690
Add a NIC-specific neighbor table statistic so we can determine how many
packets have been queued to Failed neighbors, indicating an unhealthy local
network. This change assists us to debug in-field issues where subsequent
traffic to a neighbor fails.
Fixes#4819
PiperOrigin-RevId: 344131119
1. setsockopt(SO_RCVTIMEO, 0) == never timeout
2. float64(time.Microsecond/time.Second) == 0
3. packetimpact tests use a lot of 1s timeouts
This becomes a more significant problem because of a recent change that binds
the sniffer only on the specific testNet interface so now the traffic on the
ctrlNet cannot wake up the blocking call anymore.
PiperOrigin-RevId: 344123465
The IGMPv2 core state machine can be shared with MLDv1 since they are
almost identical, ignoring specific addresses, constants and packets.
Bug #4682, #4861
PiperOrigin-RevId: 344102615
Summary of the approach: the test runner will set up a few DUTs according to
a flag and pass all the test networks to the testbench. The testbench will only
reside in a single container. The testbench will put all the test networks into
a buffered channel which served as a semaphore and now the user can freely use
t.Parallel() in (sub)tests and the true parallelism will be determined by how
many DUTs are configured. Creating DUTs on demand is not supported yet, the
test author should determine the number of DUTs to be used statically.
Specifically in this change:
- Don't export any global variables about the test network in testbench.
- Sniffer only binds on the local interface because it will be possible to have
multiple interfaces to multiple DUTs in a single testbench container.
- Migrate existing tests to stop using global variables.
PiperOrigin-RevId: 343965962
This change also simplifies and documents the build_cmd pipeline, and
reduces general noise for debugging Makefile issues.
It also drops the mapping for /etc/docker/daemon.json, which if it
does not exist initially will create this as a directory (causing lots
of confusion and breaks).
PiperOrigin-RevId: 343932456
This function does not exist in Go 1.13. We need to add an adaptor
to build against Go 1.13, which is the default Ubuntu version.
PiperOrigin-RevId: 343929132
Not all files are always accessible by the process itself. This
was specifically seen with map_files, but there's no rule that
every entry must be accessible by the process itself.
PiperOrigin-RevId: 343919117
If no vild syndrome(data abort outside memslots) was reported by kvm, let userspace to do the
ext_dabt injection to bail out this issue.
Signed-off-by: Robin Luk <lubin.lu@antgroup.com>
1. Add getD/getDentry methods to avoid long casting line in each test
2. Factor all calls to vfs.OpenAt/UnlinkAt/RenameAt on lower filesystem
to their own method (for both lower file and lower Merkle file) so
the tests are more readable
3. Add descriptive test names for delete/remove tests
PiperOrigin-RevId: 343540202
Added headers, stats, checksum parsing capabilities from RFC 2236 describing
IGMPv2.
IGMPv2 state machine is implemented for each condition, sending and receiving
IGMP Membership Reports and Leave Group messages with backwards compatibility
with IGMPv1 routers.
Test:
* Implemented igmp header parser and checksum calculator in header/igmp_test.go
* ipv4/igmp_test.go tests incoming and outgoing IGMP messages and pathways.
* Added unit test coverage for IGMPv2 RFC behavior + IGMPv1 backwards
compatibility in ipv4/igmp_test.go.
Fixes#4682
PiperOrigin-RevId: 343408809
We would like to track locks ordering to detect ordering violations. Detecting
violations is much simpler if mutexes must be unlocked by the same goroutine
that locked them.
Thus, as a first step to tracking lock ordering, add this lock/unlock
requirement to gVisor's sync.Mutex. This is more strict than the Go standard
library's sync.Mutex, but initial testing indicates only a single lock that is
used across goroutines. The new sync.CrossGoroutineMutex relaxes the
requirement (but will not provide lock order checking).
Due to the additional overhead, enforcement is only enabled with the
"checklocks" build tag. Build with this tag using:
bazel build --define=gotags=checklocks ...
From my spot-checking, this has no changed inlining properties when disabled.
Updates #4804
PiperOrigin-RevId: 343370200
Group addressable endpoints can simply check if it has joined the
multicast group without maintaining address endpoints. This also
helps remove the dependency on AddressableEndpoint from
GroupAddressableEndpoint.
Now that group addresses are not tracked with address endpoints, we can
avoid accidentally obtaining a route with a multicast local address.
PiperOrigin-RevId: 343336912
This patch adds support for tlbi-vale1ls/tlbi-aside1ls.
And make the code consistent with the flush strategy of the x86 platform.
Signed-off-by: Robin Luk <lubin.lu@antgroup.com>
If a kernfs user does not cache dentries, then cacheLocked will destroy the
dentry. The current DecRef implementation will be racy in this case as the
following can happen:
- Goroutine 1 calls DecRef and decreases ref count from 1 to 0.
- Goroutine 2 acquires d.fs.mu for reading and calls IncRef and increasing the
ref count from 0 to 1.
- Goroutine 2 releases d.fs.mu and calls DecRef again decreasing ref count from
1 to 0.
- Goroutine 1 now acquires d.fs.mu and calls cacheLocked which destroys the
dentry.
- Goroutine 2 now acquires d.fs.mu and calls cacheLocked to find that the dentry
is already destroyed!
Earlier we would panic in this case, we could instead just return instead of
adding complexity to handle this race. This is similar to what the gofer client
does.
We do not want to lock d.fs.mu in the case that the filesystem caches dentries
(common case as procfs and sysfs do this) to prevent congestion due to lock
contention.
PiperOrigin-RevId: 343229496
Ghanan Gowripalan
Dean Deng
Zeling Feng
Bhasker Hariharan
Toshi Kikuchi
Jamie Liu
Adin Scannell
Ayush Ranjan
Robin Luk
Sam Balana
Julian Elischer
Chong Cai
gVisor bot
Ryan Heacock
Rahat Mahmood
Fabricio Voznika
Michael Pratt
Bruno Dal Bo
Ting-Yu Wang