- Tweak the benchmarks to work with b.N where appropriate. In many cases,
b.N was simply being ignored. This creates an implicit dependency in the
user passing a reasonable benchtime (less than or equal to the actual
runtime of the test, or using the X syntax) otherwise the test runs
forever.
- In cases where the above is impossible, explicitly set benchtime from
the test wrapper, to prevent the above behavior (tensorflow).
- Drop the *Reverse variants, which are simply hey benchmarks. We should
just add a hey benchmark. The platforms benchmarks already include a
native platform, and thus these benchmarks are incredibly confusing.
(In other words, BenchmarkNginxReverse has nothing to do with an nginx
benchmark for runsc.)
- Remove the redunant Harness object, which contains no state, in order
to slightly simplify the code.
- Make Block and Heap profiling actually work, but setting appropriate
runtime parameters (and plumbing them through the config).
- Split the profiling into two phases: start and stop, since some will
need to be started early, and others will need to happen at the end.
PiperOrigin-RevId: 349495377
This allows for a model of profiling when you can start collection, and
it will terminate when the sandbox terminates. Without this synchronous
call, it is effectively impossible to collect length blocking and mutex
profiles.
PiperOrigin-RevId: 349483418
Removes the period of time in which subseqeuent traffic to a Failed neighbor
immediately fails with ErrNoLinkAddress. A Failed neighbor is one in which
address resolution fails; or in other words, the neighbor's IP address cannot
be translated to a MAC address.
This means removing the Failed state for linkAddrCache and allowing transitiong
out of Failed into Incomplete for neighborCache. Previously, both caches would
transition entries to Failed after address resolution fails. In this state, any
subsequent traffic requested within an unreachable time would immediately fail
with ErrNoLinkAddress. This does not follow RFC 4861 section 7.3.3:
If address resolution fails, the entry SHOULD be deleted, so that subsequent
traffic to that neighbor invokes the next-hop determination procedure again.
Invoking next-hop determination at this point ensures that alternate default
routers are tried.
The API for getting a link address for a given address, whether through the link
address cache or the neighbor table, is updated to optionally take a callback
which will be called when address resolution completes. This allows `Route` to
handle completing link resolution internally, so callers of (*Route).Resolve
(e.g. endpoints) don’t have to keep track of when it completes and update the
Route accordingly.
This change also removes the wakers from LinkAddressCache, NeighborCache, and
Route in favor of the callbacks, and callers that previously used a waker can
now just pass a callback to (*Route).Resolve that will notify the waker on
resolution completion.
Fixes#4796
Startblock:
has LGTM from sbalana
and then
add reviewer ghanan
PiperOrigin-RevId: 348597478
...when performing source address selection for IPv6.
These are defined in RFC 6724 section 5 rule 6 (prefer matching label)
and rule 8 (use longest matching prefix).
This change also considers ULA of global scope instead of its own scope,
as per RFC 6724 section 3.1:
Also, note that ULAs are considered as global, not
site-local, scope but are handled via the prefix policy table as
discussed in Section 10.6.
Test: stack_test.TestIPv6SourceAddressSelectionScope
Startblock:
has LGTM from peterjohnston
and then
add reviewer brunodalbo
PiperOrigin-RevId: 348580996
sacked_out is required in RACK to check the number of duplicate
acknowledgements during updating the reorder window. If there is no reordering
and the value for sacked_out is greater than the classic threshold value 3,
then reorder window is set to zero.
It is calculated by counting the number of segments sacked in the ACK and is
reduced when a cumulative ACK is received which covers the SACK blocks. This
value is set to zero when the connection enters recovery.
PiperOrigin-RevId: 347872246
When the scaled receive window size > 65535 (max uint16), we advertise
the scaled value as 65535, but are not adjusting the saved receive
window value when doing so. This would keep our current window
calculation logic to be incorrect, as the saved receive window value
is different from what was advertised.
Fixes#4903
PiperOrigin-RevId: 347771340
RFC 2711 specifies that the router alert's length field is always 2
so we should make sure only 2 bytes are read from a router alert
option's data field.
Test: header.TestIPv6OptionsExtHdrIterErr
PiperOrigin-RevId: 347727876
syzkaller reported the closing of a nil channel. This is only possible when the
AIOContext was destroyed twice.
Some scenarios that could lead to this:
- It died and then some called aioCtx.Prepare() on it and then killed it again
which could cause the double destroy. The context could have been destroyed
in between the call to LookupAIOContext() and Prepare().
- aioManager was destroyed but it did not update the contexts map. So
Lookup could still return a dead AIOContext and then someone could call
Prepare on it and kill it again.
So added a check in aioCtx.Prepare() for the context being dead. This will
prevent a dead context from resurrecting.
Also refactored code to destroy the aioContext consistently. Earlier we were not
munmapping the aioContexts that were destroyed upon aioManager destruction.
Reported-by: [email protected]
PiperOrigin-RevId: 347704347
We want to make the recvmsg syscall to the host regardless of if the dst is
empty or not so that:
- Host can populate the control messages if necessary.
- Host can return sender address.
- Host can return appropriate errors.
Earlier because we were using the IOSequence.CopyOutFrom() API, the usermem
package does not even call the Reader function if the destination is empty (as
an optimization).
PiperOrigin-RevId: 347684566
packetEPs may get into a state that `len < cap`, casuing append() modifying the
original slice storage.
Reported-by: [email protected]
PiperOrigin-RevId: 347634351
There are surprisingly few syscall tests that run with hostinet. For example
running the following command only returns two results:
`bazel query test/syscalls:all | grep hostnet`
I think as a result, as our control messages evolved, hostinet was left
behind. Update it to support all control messages netstack supports.
This change also updates sentry's control message parsing logic to make it up to
date with all the control messages we support.
PiperOrigin-RevId: 347508892
SO_OOBINLINE option is set/get as boolean value, which is the same as linux.
As we currently do not support disabling this option, we always return it as
true.
PiperOrigin-RevId: 347413905
- Deep-copy pkt.Data and hold it instead of shallow-copy (vv.Clone).
This allows the pkt's backing array, which includes the header portion,
to be freed.
- Remove fragHeap. The fragments are now held in holes struct instead.
- Stop reserving the initial capacity of holes slice.
PiperOrigin-RevId: 347198744
Adds IPv6 extension header serializer and Hop by Hop options serializer.
Add RouterAlert option serializer and use it in MLD.
Fixed#4996
Startblock:
has LGTM from marinaciocea
and then
add reviewer ghanan
PiperOrigin-RevId: 347174537
This command takes instruction pointers from stdin and converts them into their
corresponding file names and line/column numbers in the runsc source code. The
inputs are not interpreted as actual addresses, but as synthetic values that are
exposed through /sys/kernel/debug/kcov. One can extract coverage information
from kcov and translate those values into locations in the source code by
running symbolize on the same runsc binary.
This will allow us to generate syzkaller coverage reports.
PiperOrigin-RevId: 347089624
We do not rely on error for getsockopt options(which have boolean values)
anymore. This will cause issue in sendmsg where we used to return error
for IPV6_V6Only option. Fix the panic by returning error (for sockets other
than TCP and UDP) if the address does not match the type(AF_INET/AF_INET6) of
the socket.
PiperOrigin-RevId: 347063838
tcpip.ControlMessages can not contain Linux specific structures which makes it
painful to convert back and forth from Linux to tcpip back to Linux when passing
around control messages in hostinet and raw sockets.
Now we convert to the Linux version of the control message as soon as we are
out of tcpip.
PiperOrigin-RevId: 347027065
fdbased endpoint was enabling fragment reassembly on the host AF_PACKET socket
to ensure that fragments are delivered inorder to the right dispatcher. But this
prevents fragments from being delivered to gvisor at all and makes testing of
gvisor's fragment reassembly code impossible.
The potential impact from this is minimal since IP Fragmentation is not really
that prevelant and in cases where we do get fragments we may deliver the
fragment out of order to the TCP layer as multiple network dispatchers may
process the fragments and deliver a reassembled fragment after the next packet
has been delivered to the TCP endpoint. While not desirable I believe the impact
from this is minimal due to low prevalence of fragmentation.
Also removed PktType and Hatype fields when binding the socket as these are not
used when binding. Its just confusing to have them specified.
See: https://man7.org/linux/man-pages/man7/packet.7.html
"Fields used for binding are
sll_family (should be AF_PACKET), sll_protocol, and sll_ifindex."
Fixes#5055
PiperOrigin-RevId: 346919439
This change also considers interfaces and network endpoints enabled up
up to the point all work to disable them are complete. This was needed
so that protocols can perform shutdown work while being disabled (e.g.
sending a packet which requires the endpoint to be enabled to obtain a
source address).
Bug #4682, #4861Fixes#4888
Startblock:
has LGTM from peterjohnston
and then
add reviewer brunodalbo
PiperOrigin-RevId: 346869702
