Kernel.cpuClockTicker increments kernel.cpuClock, which tasks use as a clock to
track their CPU usage. This improves latency in the syscall path by avoid
expensive monotonic clock calls on every syscall entry/exit.
However, this timer fires every 10ms. Thus, when all tasks are idle (i.e.,
blocked or stopped), this forces a sentry wakeup every 10ms, when we may
otherwise be able to sleep until the next app-relevant event. These wakeups
cause the sentry to utilize approximately 2% CPU when the application is
otherwise idle.
Updates to clock are not strictly necessary when the app is idle, as there are
no readers of cpuClock. This commit reduces idle CPU by disabling the timer
when tasks are completely idle, and computing its effects at the next wakeup.
Rather than disabling the timer as soon as the app goes idle, we wait until the
next tick, which provides a window for short sleeps to sleep and wakeup without
doing the (relatively) expensive work of disabling and enabling the timer.
PiperOrigin-RevId: 272265822
'docker exec' was getting CAP_NET_RAW even when --net-raw=false
because it was not filtered out from when copying container's
capabilities.
PiperOrigin-RevId: 272260451
Linux changed this behavior in 16e72e9b30986ee15f17fbb68189ca842c32af58
(v4.11). Previously, extra pages were always mapped RW. Now, those pages will
be executable if the segment specified PF_X. They still must be writeable.
PiperOrigin-RevId: 272256280
The test is checking the wrong poll_fd for POLLHUP. The only
reason it passed till now was because it was also checking
for POLLIN which was always true on the other fd from the
previous poll!
PiperOrigin-RevId: 270780401
Previously, when we set hostname:
$ strace hostname abc
...
sethostname("abc", 3) = -1 ENAMETOOLONG (File name too long)
...
According to man 2 sethostname:
"The len argument specifies the number of bytes in name. (Thus, name
does not require a terminating null byte.)"
We wrongly use the CopyStringIn() to check terminating zero byte in
the implementation of sethostname syscall.
To fix this, we use CopyInBytes() instead.
Fixes: #861
Reported-by: chenglang.hy <chenglang.hy@antfin.com>
Signed-off-by: Jianfeng Tan <henry.tjf@antfin.com>
Adresses a deadlock with the rolled back change:
b6a5b950d2
Creating a session from an orphaned process group was causing a lock to be
acquired twice by a single goroutine. This behavior is addressed, and a test
(OrphanRegression) has been added to pty.cc.
Implemented the following ioctls:
- TIOCSCTTY - set controlling TTY
- TIOCNOTTY - remove controlling tty, maybe signal some other processes
- TIOCGPGRP - get foreground process group. Also enables tcgetpgrp().
- TIOCSPGRP - set foreground process group. Also enabled tcsetpgrp().
Next steps are to actually turn terminal-generated control characters (e.g. C^c)
into signals to the proper process groups, and to send SIGTTOU and SIGTTIN when
appropriate.
PiperOrigin-RevId: 270088599
Note that the exact semantics for these signalfds are slightly different from
Linux. These signalfds are bound to the process at creation time. Reads, polls,
etc. are all associated with signals directed at that task. In Linux, all
signalfd operations are associated with current, regardless of where the
signalfd originated.
In practice, this should not be an issue given how signalfds are used. In order
to fix this however, we will need to plumb the context through all the event
APIs. This gets complicated really quickly, because the waiter APIs are all
netstack-specific, and not generally exposed to the context. Probably not
worthwhile fixing immediately.
PiperOrigin-RevId: 269901749
- Fix ARG syntax in Dockerfiles.
- Fix curl commands in Dockerfiles.
- Fix some paths in proctor binaries.
- Check error from Walk in search helper.
PiperOrigin-RevId: 269641686
* Use multi-stage builds in Dockerfiles.
* Combine all proctor binaries into a single binary.
* Change the TestRunner interface to reduce code duplication.
PiperOrigin-RevId: 269462101
absl flags are more modern and we can easily depend on them directly.
The repo now successfully builds with --incompatible_load_cc_rules_from_bzl.
PiperOrigin-RevId: 269387081
This also allows the tee(2) implementation to be enabled, since dup can now be
properly supported via WriteTo.
Note that this change necessitated some minor restructoring with the
fs.FileOperations splice methods. If the *fs.File is passed through directly,
then only public API methods are accessible, which will deadlock immediately
since the locking is already done by fs.Splice. Instead, we pass through an
abstract io.Reader or io.Writer, which elide locks and use the underlying
fs.FileOperations directly.
PiperOrigin-RevId: 268805207
A recent Kokoro change pointed to go_tests.cfg (in line with the
other configurations), which unfortunately broke the presubmits.
This change also enabled the KVM tests, which were still using a
remote execution strategy.
This fixes both of these issues and allows presubmits to pass.
One additional test was caught with this case, which seems to
have been broken. It's unclear why this was not being caught.
PiperOrigin-RevId: 268166291
This is done because the root container for CRI is the infrastructure (pause)
container and always gets a low oom_score_adj. We do this to ensure that only
the oom_score_adj of user containers is used to calculated the sandbox
oom_score_adj.
Implemented in runsc rather than the containerd shim as it's a bit cleaner to
implement here (in the shim it would require overwriting the oomScoreAdj and
re-writing out the config.json again). This processing is Kubernetes(CRI)
specific but we are currently only supporting CRI for multi-container support
anyway.
PiperOrigin-RevId: 267507706
TestNoDuplicates is racy as it tries to read the /proc file system
while the test is running. But it's possible that from the time a
directory entries are read and each entry processed something could
change and in some cases the entry being processed could have been
deleted. In such cases we should not fail the test but just
ignore the error and move on.
PiperOrigin-RevId: 267483094
- Most AIO tests call io_setup(nr_events = 128). sizeof(struct io_event)
(128*32 = 4096). However, the actual size of the mapping created by
io_setup() is determined by:
(from fs/aio.c:ioctx_alloc())
/*
* We keep track of the number of available ringbuffer slots, to prevent
* overflow (reqs_available), and we also use percpu counters for this.
*
* So since up to half the slots might be on other cpu's percpu counters
* and unavailable, double nr_events so userspace sees what they
* expected: additionally, we move req_batch slots to/from percpu
* counters at a time, so make sure that isn't 0:
*/
nr_events = max(nr_events, num_possible_cpus() * 4);
nr_events *= 2;
(from fs/aio.c:aio_setup_ring())
/* Compensate for the ring buffer's head/tail overlap entry */
nr_events += 2; /* 1 is required, 2 for good luck */
size = sizeof(struct aio_ring);
size += sizeof(struct io_event) * nr_events;
nr_pages = PFN_UP(size);
When we mremap() only the first page of a multi-page AIO ring buffer
mapping, fs/aio.c:aio_ring_mremap() updates struct kioctx::mmap_base -
but struct kioctx::mmap_size is untouched, so sys_io_destroy() =>
kill_ioctx() vm_unmaps() the mremapped page, plus some number of pages
after it. Just get the actual size of the mapping from /proc/self/maps.
- Delete test case MremapOver; while it is correct that Linux will not
complain if you overwrite the AIO ring buffer with another mapping, it
won't actually work in the sense that AIO events will not be written to
the new mapping, because Linux stores the struct pages of the ring
buffer in struct kioctx::ring_pages and writes to those through kmap()
rather than using userspace addresses.
- Don't munmap() after mremap(MREMAP_FIXED) returns EFAULT; see new
comment in factored-out test case MremapExpansion.
PiperOrigin-RevId: 267482903
The simple test script has gotten out of control. Shard this script into
different pieces and attempt to impose order on overall test structure. This
change helps lay some of the foundations for future improvements.
* The runsc/test directories are moved into just test/.
* The runsc/test/testutil package is split into logical pieces.
* The scripts/ directory contains new top-level targets.
* Each test is now responsible for building targets it requires.
* The install functionality is moved into `runsc` itself for simplicity.
* The existing kokoro run_tests.sh file now just calls all (can be split).
After this change is merged, I will create multiple distinct workflows for
Kokoro, one for each of the scripts currently targeted by `run_tests.sh` today,
which should dramatically reduce the time-to-run for the Kokoro tests, and
provides a better foundation for further improvements to the infrastructure.
PiperOrigin-RevId: 267081397
Ioctl was returning just the buffer size from epsocket.endpoint
and it was not considering data from epsocket.SocketOperations
that was read from the endpoint, but not yet sent to the caller.
PiperOrigin-RevId: 266485461
When abstract unix domain socket paths are displayed in
/proc/net/unix, Linux historically emitted null bytes as padding at
the end of the path. Newer versions of Linux (v4.9,
e7947ea770d0de434d38a0f823e660d3fd4bebb5) display these as '@'
characters.
Update proc_net_unix test to handle both version of the padding.
PiperOrigin-RevId: 266230200
Using "go run ..." in the ENTRYPOINT causes the go compiler to run each time
the container is started. We can just compile the binary once as part of the
image.
PiperOrigin-RevId: 266212462
The flake had the call to futex_unlock_pi() returning EINVAL with the
FUTEX_OWNER_DIED set. In this case, userspace has to clean up stale
state. So instead of calling FUTEX_UNLOCK_PI outright, we'll use the
advised atomic compare_exchange as advised in the man page.
PiperOrigin-RevId: 265163920
In cl/264434674 and cl/264498919, we stop running test cases
in parallel to not overload test hosts. But now tests requires
more time to run, so we need to increase a default number of
shards or a default test timeout. Let's start with increasing
the number of shards and see how it will works.
PiperOrigin-RevId: 264917055
For SOCK_STREAM type unix socket, we shall return ECONNRESET if peer is
closed with data not read.
We explictly set a flag when closing one end, to differentiate from
just shutdown (where zero shall be returned).
Fixes: #735
Signed-off-by: Jianfeng Tan <henry.tjf@antfin.com>
Previously, recvmsg() on a unix stream socket with its peer closed will
never return, with goroutine call trace like this:
...
2 in gvisor.dev/gvisor/pkg/sentry/kernel.(*Task).block
at pkg/sentry/kernel/task_block.go:124
3 in gvisor.dev/gvisor/pkg/sentry/kernel.(*Task).BlockWithDeadline
at pkg/sentry/kernel/task_block.go:69
4 in gvisor.dev/gvisor/pkg/sentry/socket/unix.(*SocketOperations).RecvMsg
at pkg/sentry/socket/unix/unix.go:612
5 in gvisor.dev/gvisor/pkg/sentry/syscalls/linux.recvFrom
at pkg/sentry/syscalls/linux/sys_socket.go:885
6 in gvisor.dev/gvisor/pkg/sentry/syscalls/linux.RecvFrom
at pkg/sentry/syscalls/linux/sys_socket.go:910
...
The issue is caused by that ErrClosedForReceive returned by
unix/transport.queue is turned into nil in
unix.(*EndpointReader).ReadToBlocks():
err.ToError()
As a result, in unix.(*SocketOperations).RecvMsg():
n == 0 and err == nil
We shall differentiate it from another case - no data to read where
ErrWouldBlock shall be returned; and return 0 immediately.
Fixes: #734
Reported-by: chenglang.hy <chenglang.hy@antfin.com>
Signed-off-by: Jianfeng Tan <henry.tjf@antfin.com>
This fixes the issue of not being able to bind to either a multicast or
broadcast address as well as to send and receive data from it. The way to solve
this is to treat these addresses similar to the ANY address and register their
transport endpoint ID with the global stack's demuxer rather than the NIC's.
That way there is no need to require an endpoint with that multicast or
broadcast address. The stack's demuxer is in fact the only correct one to use,
because neither broadcast- nor multicast-bound sockets care which NIC a
packet was received on (for multicast a join is still needed to receive packets
on a NIC).
I also took the liberty of refactoring udp_test.go to consolidate a lot of
duplicate code and make it easier to create repetitive tests that test the same
feature for a variety of packet and socket types. For this purpose I created a
"flowType" that represents two things: 1) the type of packet being sent or
received and 2) the type of socket used for the test. E.g., a "multicastV4in6"
flow represents a V4-mapped multicast packet run through a V6-dual socket.
This allows writing significantly simpler tests. A nice example is testTTL().
PiperOrigin-RevId: 264766909
test/syscalls/linux/proc_net_tcp.cc:252: Failure
Value of: connect(client->get(), &addr, addrlen)
Expected: not -1 (success)
Actual: -1 (of type int), with errno PosixError(errno=4 Interrupted system call)
PiperOrigin-RevId: 264743815
We need real sharding, and will let Bazel handle the
parallelization. That is coming soon. Until then, remove
this call to t.Parallel() so that we can run the tests without
eating all CPU.
PiperOrigin-RevId: 264498919
The test is long running (175128 ms or so) which causes timeouts.
The test simply makes sure that private futexes can acquire
locks concurrently. Dropping current threads and increasing the
number of locks each thread tests the same concurrency concerns
but drops execution time to ~1411 ms.
PiperOrigin-RevId: 264476144
bazel runs a few instances of syscall_test_runner in parallel
and then syscall_test_runner runs test cases in parallel. It might
be a reason why we see that test hosts are overloaded and sandboxes
start slowly. It should be better to control how many tests are
running in parallel from one place, so let's try to disable this
feature in syscall_test_runner.
PiperOrigin-RevId: 264434674
SendMsg before this change would copy all the data over into a
new slice even if the underlying socket could only accept a
small amount of data. This is really inefficient with non-blocking
sockets and under high throughput where large writes could get
ErrWouldBlock or if there was say a timeout associated with the sendmsg()
syscall.
With this change we delay copying bytes in till they are needed and only
copy what can be potentially sent/held in the socket buffer. Reducing
the need to repeatedly copy data over.
Also a minor fix to change state FIN-WAIT-1 when shutdown(..., SHUT_WR) is called
instead of when we transmit the actual FIN. Otherwise the socket could remain in
CONNECTED state even though the user has called shutdown() on the socket.
Updates #627
PiperOrigin-RevId: 263430505
The new version has a change in behavior when using a custom platform:
* Old behavior: rules that don't require a toolchain used host_platform, no
matter what execution platforms are specified.
* New behavior: rules that don't require a toolchain use standard platform
resolution that starts with execution platforms.
As part of this change, we cannot use the "extra_exectution_platforms" flag
provided by the default bazelrc. I got rid of the default bazelrc file, and
made our custom .bazelrc as minimal as possible.
PiperOrigin-RevId: 263176802
Now if a process sends an unsupported netlink requests,
an error is returned from the send system call.
The linux kernel works differently in this case. It returns errors in the
nlmsgerr netlink message.
Reported-by: syzbot+571d99510c6f935202da@syzkaller.appspotmail.com
PiperOrigin-RevId: 262690453
Using the path_test.go file built by the Golang
devs as a base, tests have been created to verify
the functionality of common.Search().
A mock file system is created and fake test files
are generated to see if they get picked up by
common.Search().
Also included in this CL is a bug fix for
proctor-nodejs that was discovered using this test.
proctor-nodejs used to allow multiple "-" in its
test name filter. The regex has been updated to
prevent this.
PiperOrigin-RevId: 262647263
This can happen because endpoint.Close() closes the accept channel first and
then drains/resets any accepted but not delivered connections. But there can be
connections that are connected but not delivered to the channel as the channel
was full. But closing the channel can cause these writes to fail with a write to
a closed channel.
The correct solution is to abort any connections in SYN-RCVD state and
drain/abort all completed connections before closing the accept channel.
PiperOrigin-RevId: 261951132
After the refactoring of the proctor binaries, the Dockerfiles for each
language must be altered to copy the common folder into their image.
Additionally, Java has been changed to use the pre-built version of
JDK-11 from Ubuntu, instead of building it from the source. This allows
for a smaller image and faster test execution within the container.
PiperOrigin-RevId: 261805158
This change adds functionality for running more languages using
the runtimes test suite. It divides the languages into separate
test functions, which each call the helper testLang function in the
runtimes_test.go file. This allows them to be run individually
or as a group.
PiperOrigin-RevId: 261791935
(Don't worry, this is mostly tests.)
Implemented the following ioctls:
- TIOCSCTTY - set controlling TTY
- TIOCNOTTY - remove controlling tty, maybe signal some other processes
- TIOCGPGRP - get foreground process group. Also enables tcgetpgrp().
- TIOCSPGRP - set foreground process group. Also enabled tcsetpgrp().
Next steps are to actually turn terminal-generated control characters (e.g. C^c)
into signals to the proper process groups, and to send SIGTTOU and SIGTTIN when
appropriate.
PiperOrigin-RevId: 261387276
This change removes the filepath.Walk() function from
proctor- go, php, and nodejs. The filepath.Walk() is
now defined in common.go in Search(). Each proctor binary
passes root directory and testFilter arguments to Search().
proctor-python.go no longer uses filepath.Walk() to search
for tests. There is a built-in list test function within
python's language test suite so that is being used instead.
PiperOrigin-RevId: 261242897
proctor-go had a bug where it would incorrectly identify
a tool test as a disk test. Instead of searching for the
test on disk as the identification method, we now check if
the test name ends in ".go". If the test ends in ".go" it
is run as a disk test, otherwise the test is run as a tool test.
Python tests need to be run from within the directory they exist.
Functionality to split the test name from it's parent directory
has been added and a cmd.Dir argument has been set.
PiperOrigin-RevId: 261021693
Shared code among proctor-*.go files has been refactored
into common/common.go. The common package is imported in
each proctor binary and a struct is created to implement
the testRunner interface defined in common.go. This allows
for the proctor binaries to be updated without having to
copy/paste the same code across all files. There are no
usage or functionality changes.
PiperOrigin-RevId: 260967080
The checksum was not being reset before being re-calculated and sent out.
This caused the sent checksum to always be `0x0800`.
Fixes#605.
PiperOrigin-RevId: 260965059
This doesn't currently pass on gVisor.
While I'm here, fix a bug where connecting to the v6-mapped v4 address doesn't
work in gVisor.
PiperOrigin-RevId: 260923961
This fixes a bug introduced in cl/251934850 that caused
connect-accept-close-connect races to result in the second connect call
failiing when it should have succeeded.
PiperOrigin-RevId: 259584525
It is now correctly initialized to the top of the signal stack.
Previously it was initialized to the address of 'stack.ss_sp' on
the main thread stack.
PiperOrigin-RevId: 258248363
iptables also relies on IPPROTO_RAW in a way. It opens such a socket to
manipulate the kernel's tables, but it doesn't actually use any of the
functionality. Blegh.
PiperOrigin-RevId: 257903078
Adds support to set/get the TCP_MAXSEG value but does not
really change the segment sizes emitted by netstack or
alter the MSS advertised by the endpoint. This is currently
being added only to unblock iperf3 on gVisor. Plumbing
this correctly requires a bit more work which will come
in separate CLs.
PiperOrigin-RevId: 257859112
A userspace process (CPL=3) can access an i/o port if the bit corresponding to
the port is set to 0 in the I/O permission bitmap.
Configure the I/O permission bitmap address beyond the last valid byte in the
TSS so access to all i/o ports is blocked.
Signed-off-by: Liu Hua <sdu.liu@huawei.com>
Change-Id: I3df76980c3735491db768f7210e71703f86bb989
PiperOrigin-RevId: 257336518
The error set in the loop in createAt was being masked
by other errors declared with ":=". This allowed an
ErrResolveViaReadlink error to escape, which can cause
a sentry panic.
Added test case which repros without the fix.
PiperOrigin-RevId: 257061767