Our current reference leak checker uses finalizers to verify whether an object
has reached zero references before it is garbage collected. There are multiple
problems with this mechanism, so a rewrite is in order.
With finalizers, there is no way to guarantee that a finalizer will run before
the program exits. When an unreachable object with a finalizer is garbage
collected, its finalizer will be added to a queue and run asynchronously. The
best we can do is run garbage collection upon sandbox exit to make sure that
all finalizers are enqueued.
Furthermore, if there is a chain of finalized objects, e.g. A points to B
points to C, garbage collection needs to run multiple times before all of the
finalizers are enqueued. The first GC run will register the finalizer for A but
not free it. It takes another GC run to free A, at which point B's finalizer
can be registered. As a result, we need to run GC as many times as the length
of the longest such chain to have a somewhat reliable leak checker.
Finally, a cyclical chain of structs pointing to one another will never be
garbage collected if a finalizer is set. This is a well-known issue with Go
finalizers (https://github.com/golang/go/issues/7358). Using leak checking on
filesystem objects that produce cycles will not work and even result in memory
leaks.
The new leak checker stores reference counted objects in a global map when
leak check is enabled and removes them once they are destroyed. At sandbox
exit, any remaining objects in the map are considered as leaked. This provides
a deterministic way of detecting leaks without relying on the complexities of
finalizers and garbage collection.
This approach has several benefits over the former, including:
- Always detects leaks of objects that should be destroyed very close to
sandbox exit. The old checker very rarely detected these leaks, because it
relied on garbage collection to be run in a short window of time.
- Panics if we forgot to enable leak check on a ref-counted object (we will try
to remove it from the map when it is destroyed, but it will never have been
added).
- Can store extra logging information in the map values without adding to the
size of the ref count struct itself. With the size of just an int64, the ref
count object remains compact, meaning frequent operations like IncRef/DecRef
are more cache-efficient.
- Can aggregate leak results in a single report after the sandbox exits.
Instead of having warnings littered in the log, which were
non-deterministically triggered by garbage collection, we can print all
warning messages at once. Note that this could also be a limitation--the
sandbox must exit properly for leaks to be detected.
Some basic benchmarking indicates that this change does not significantly
affect performance when leak checking is enabled, which is understandable
since registering/unregistering is only done once for each filesystem object.
Updates #1486.
PiperOrigin-RevId: 338685972
Using the newer bazel rules necessitates a transition from proto1 to
proto2. In order to resolve the incompatibility between proto2 and
gogoproto, the cri runtimeoptions proto must be vendored.
Further, some of the semantics of bazel caching changed during the
transition. It is now necessary to:
- Ensure that :gopath depends only on pure library targets, as the
propagation of go_binary build attributes (pure, static) will
affected the generated files (though content remains the same,
there are conflicts with respect to the gopath).
- Update bazel.mk to include the possibility of binaries in the
bazel-out directory, as it will now put runsc and others there.
This required some refinements to the mechanism of extracting
paths, since some the existing regex resulted in false positives.
- Change nogo rules to prevent escape generation on binary targets.
For some reason, the newer version of bazel attempted to run the
nogo analysis on the binary targets, which fails due to the fact
that objdump does not work on the final binary. This must be due
to a change in the semantics of aspects in bazel3.
PiperOrigin-RevId: 337958324
OCI configuration includes support for specifying seccomp filters. In runc,
these filter configurations are converted into seccomp BPF programs and loaded
into the kernel via libseccomp. runsc needs to be a static binary so, for
runsc, we cannot rely on a C library and need to implement the functionality
in Go.
The generator added here implements basic support for taking OCI seccomp
configuration and converting it into a seccomp BPF program with the same
behavior as a program generated by libseccomp.
- New conditional operations were added to pkg/seccomp to support operations
available in OCI.
- AllowAny and AllowValue were renamed to MatchAny and EqualTo to better reflect
that syscalls matching the conditionals result in the provided action not
simply SCMP_RET_ALLOW.
- BuildProgram in pkg/seccomp no longer panics if provided an empty list of
rules. It now builds a program with the architecture sanity check only.
- ProgramBuilder now allows adding labels that are unused. However, backwards
jumps are still not permitted.
Fixes#510
PiperOrigin-RevId: 331938697
VFS1 and VFS2 host FDs have different dupping behavior,
making error prone to code for both. Change the contract
so that FDs are released as they are used, so the caller
can simple defer a block that closes all remaining files.
This also addresses handling of partial failures.
With this fix, more VFS2 tests can be enabled.
Updates #1487
PiperOrigin-RevId: 330112266
Major differences from existing overlay filesystems:
- Linux allows lower layers in an overlay to require revalidation, but not the
upper layer. VFS1 allows the upper layer in an overlay to require
revalidation, but not the lower layer. VFS2 does not allow any layers to
require revalidation. (Now that vfs.MkdirOptions.ForSyntheticMountpoint
exists, no uses of overlay in VFS1 are believed to require upper layer
revalidation; in particular, the requirement that the upper layer support the
creation of "trusted." extended attributes for whiteouts effectively required
the upper filesystem to be tmpfs in most cases.)
- Like VFS1, but unlike Linux, VFS2 overlay does not attempt to make mutations
of the upper layer atomic using a working directory and features like
RENAME_WHITEOUT. (This may change in the future, since not having a working
directory makes error recovery for some operations, e.g. rmdir, particularly
painful.)
- Like Linux, but unlike VFS1, VFS2 represents whiteouts using character
devices with rdev == 0; the equivalent of the whiteout attribute on
directories is xattr trusted.overlay.opaque = "y"; and there is no equivalent
to the whiteout attribute on non-directories since non-directories are never
merged with lower layers.
- Device and inode numbers work as follows:
- In Linux, modulo the xino feature and a special case for when all layers
are the same filesystem:
- Directories use the overlay filesystem's device number and an
ephemeral inode number assigned by the overlay.
- Non-directories that have been copied up use the device and inode
number assigned by the upper filesystem.
- Non-directories that have not been copied up use a per-(overlay,
layer)-pair device number and the inode number assigned by the lower
filesystem.
- In VFS1, device and inode numbers always come from the lower layer unless
"whited out"; this has the adverse effect of requiring interaction with
the lower filesystem even for non-directory files that exist on the upper
layer.
- In VFS2, device and inode numbers are assigned as in Linux, except that
xino and the samefs special case are not supported.
- Like Linux, but unlike VFS1, VFS2 does not attempt to maintain memory mapping
coherence across copy-up. (This may have to change in the future, as users
may be dependent on this property.)
- Like Linux, but unlike VFS1, VFS2 uses the overlayfs mounter's credentials
when interacting with the overlay's layers, rather than the caller's.
- Like Linux, but unlike VFS1, VFS2 permits multiple lower layers in an
overlay.
- Like Linux, but unlike VFS1, VFS2's overlay filesystem is
application-mountable.
Updates #1199
PiperOrigin-RevId: 316019067
We can register any number of tables with any number of architectures, and
need not limit the definitions to the architecture in question. This allows
runsc to generate documentation for all architectures simultaneously.
Similarly, this simplifies the VFSv2 patching process.
PiperOrigin-RevId: 310224827
This change includes:
- Modifications to loader_test.go to get TestCreateMountNamespace to
pass with VFS2.
- Changes necessary to get TestHelloWorld in image tests to pass with
VFS2. This means runsc can run the hello-world container with docker
on VSF2.
Note: Containers that use sockets will not run with these changes.
See "//test/image/...". Any tests here with sockets currently fail
(which is all of them but HelloWorld).
PiperOrigin-RevId: 308363072
This change adds a layer of abstraction around the internal Docker APIs,
and eliminates all direct dependencies on Dockerfiles in the infrastructure.
A subsequent change will automated the generation of local images (with
efficient caching). Note that this change drops the use of bazel container
rules, as that experiment does not seem to be viable.
PiperOrigin-RevId: 308095430
Included:
- loader_test.go RunTest and TestStartSignal VFS2
- container_test.go TestAppExitStatus on VFS2
- experimental flag added to runsc to turn on VFS2
Note: shared mounts are not yet supported.
PiperOrigin-RevId: 307070753
TCP/IP will work with netstack networking. hostinet doesn't work, and sockets
will have the same behavior as it is now.
Before the userspace is able to create device, the default loopback device can
be used to test.
/proc/net and /sys/net will still be connected to the root network stack; this
is the same behavior now.
Issue #1833
PiperOrigin-RevId: 296309389
FD table now holds both VFS1 and VFS2 types and uses the correct
one based on what's set.
Parts of this CL are just initial changes (e.g. sys_read.go,
runsc/main.go) to serve as a template for the remaining changes.
Updates #1487
Updates #1623
PiperOrigin-RevId: 292023223
Because the abi will depend on the core types for marshalling (usermem,
context, safemem, safecopy), these need to be flattened from the sentry
directory. These packages contain no sentry-specific details.
PiperOrigin-RevId: 291811289
* Rename syncutil to sync.
* Add aliases to sync types.
* Replace existing usage of standard library sync package.
This will make it easier to swap out synchronization primitives. For example,
this will allow us to use primitives from github.com/sasha-s/go-deadlock to
check for lock ordering violations.
Updates #1472
PiperOrigin-RevId: 289033387
This patch also include a minor change to replace syscall.Dup2
with syscall.Dup3 which was missed in a previous commit(ref a25a976).
Signed-off-by: Haibo Xu <haibo.xu@arm.com>
Change-Id: I00beb9cc492e44c762ebaa3750201c63c1f7c2f3
NETLINK_KOBJECT_UEVENT sockets send udev-style messages for device events.
gVisor doesn't have any device events, so our sockets don't need to do anything
once created.
systemd's device manager needs to be able to create one of these sockets. It
also wants to install a BPF filter on the socket. Since we'll never send any
messages, the filter would never be invoked, thus we just fake it out.
Fixes#1117
Updates #1119
PiperOrigin-RevId: 278405893
Each gofer now has a goroutine that polls on the FDs used
to communicate with the sandbox. The respective gofer is
destroyed if any of the FDs is closed.
Closes#601
PiperOrigin-RevId: 261383725
This renames FDMap to FDTable and drops the kernel.FD type, which had an entire
package to itself and didn't serve much use (it was freely cast between types,
and served as more of an annoyance than providing any protection.)
Based on BenchmarkFDLookupAndDecRef-12, we can expect 5-10 ns per lookup
operation, and 10-15 ns per concurrent lookup operation of savings.
This also fixes two tangential usage issues with the FDMap. Namely, non-atomic
use of NewFDFrom and associated calls to Remove (that are both racy and fail to
drop the reference on the underlying file.)
PiperOrigin-RevId: 256285890
runsc will now set the HOME environment variable as required by POSIX. The
user's home directory is retrieved from the /etc/passwd file located on the
container's file system during boot.
PiperOrigin-RevId: 253120627
Parse annotations containing 'gvisor.dev/spec/mount' that gives
hints about how mounts are shared between containers inside a
pod. This information can be used to better inform how to mount
these volumes inside gVisor. For example, a volume that is shared
between containers inside a pod can be bind mounted inside the
sandbox, instead of being two independent mounts.
For now, this information is used to allow the same tmpfs mounts
to be shared between containers which wasn't possible before.
PiperOrigin-RevId: 252704037
This is in preparation for improved page cache reclaim, which requires
greater integration between the page cache and page allocator.
PiperOrigin-RevId: 238444706
Change-Id: Id24141b3678d96c7d7dc24baddd9be555bffafe4
This help troubleshoot cases where the container is killed and the
app logs don't show the reason.
PiperOrigin-RevId: 236982883
Change-Id: I361892856a146cea5b04abaa3aedbf805e123724
Nothing reads them and they can simply get stale.
Generated with:
$ sed -i "s/licenses(\(.*\)).*/licenses(\1)/" **/BUILD
PiperOrigin-RevId: 231818945
Change-Id: Ibc3f9838546b7e94f13f217060d31f4ada9d4bf0
Andrei Vagin
Kevin Krakauer
Dean Deng
Adin Scannell
Fabricio Voznika
Ian Lewis
Bhasker Hariharan
gVisor bot
Ridwan Sharif
Nicolas Lacasse
Jamie Liu
Zach Koopmans
Ian Gudger
Haibo Xu
Michael Pratt
Brian Geffon