Currently, we send a single SIGKILL to the shim process
once and then we spin in a loop where we use kill(pid, 0)
to detect when the pid has disappeared completely.
Unfortunately, this has a race condition since pids can be reused causing us
to spin in an infinite loop when that happens.
This adds a timeout to this loop which logs a warning and exits the
infinite loop.
Signed-off-by: Ashray Jain <ashrayj@palantir.com>
Dependencies may be switching to use the new `%w` formatting
option to wrap errors; switching to use `errors.Is()` makes
sure that we are still able to unwrap the error and detect the
underlying cause.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Instead of having several dialer implementations, leave only one in
`pkg/dialer` and call it from `pkg/ttrpcutil`, `runtime/v(1|2)/shim`
which had their own
Closes#3471.
Signed-off-by: Kiril Vladimiroff <kiril@vladimiroff.org>
When there is timeout or cancel for create container, killShim will fail
because of canceled context. The shim will be dangling and unmanageable.
Need to use new context to do cleanup.
Signed-off-by: Wei Fu <fuweid89@gmail.com>
Reized the I/O buffers to align with the size of the kernel buffers with fifos
and move the close aspect of the console to key off of the stdin closing.
Fixes#3738
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
When containerd-shim does reaper, the most processes are not init
process. Since json.Decode consumes more CPU resource, we should check
killall option for init process only.
Signed-off-by: Wei Fu <fuweid89@gmail.com>
This changes the shim's OOM score from a static max killable of -999 to
be +1 of the containerd daemon's score. This should allow the shim's to
be killed first in an OOM condition but leave the daemon alone for a bit
to help cleanup and manage the containers during this situation.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
Closes#3286
This and a combination of a couple Docker changes are needed to fully
resolve the issue on the Docker side. However, this ensures that after
processes exit, we still leave some time for the I/O to fully flush
before closing. Without this timeout, the delete methods would block
forever.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
Allow third-party runtime implementations to reuse NewBinaryIO
in order to support pluggable shim logging binary protocol.
Signed-off-by: Maksym Pavlenko <makpav@amazon.com>
skip hidden directories in load task, and return soon if path not exist
in atomicDelete
carry of #3233Closes#3233
Signed-off-by: Ace-Tang <aceapril@126.com>
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
Currently when we restart containerd it will load all tasks with shim
logs whether the `shim_debug` is set or not.
Signed-off-by: Li Yuxuan <liyuxuan04@baidu.com>
Open shim v2 log with the flag `O_RDWR` will cause the `Read()` block
forever even if the pipe has been closed on the shim side. Then the
`io.Copy()` would never return and lead to a fd leak.
Fix typo when closing shim v1 log which causes the `stdouLog` leak.
Update `numPipes` function in test case to get the opened FIFO
correctly.
Signed-off-by: Li Yuxuan <liyuxuan04@baidu.com>
1. kill shim in cleanupAfterDeadShim avoid shim leak
2. refactor cleanupAfterDeadShim, get pid from bundle
path instead of make pid as a parameter
Signed-off-by: Ace-Tang <aceapril@126.com>
* Rootfs dir is created during container creation not during bundle
creation
* Add support for v2
* UnmountAll is a no-op when the path to unmount (i.e. the rootfs dir)
does not exist or is invalid
Co-authored-by: Danail Branekov <danailster@gmail.com>
Signed-off-by: Georgi Sabev <georgethebeatle@gmail.com>
Use sha256 hash to shorten the unix socket path to satisfy the
length limitation of abstract socket path
This commit also backports the feature storing address path to
a file from v2 to keep compatibility
Fixes#3032
Signed-off-by: Eric Lin <linxiulei@gmail.com>
Closes#603
This adds logging facilities at the shim level to provide minimal I/O
overhead and pluggable logging options. Log handling is done within the
shim so that all I/O, cpu, and memory can be charged to the container.
A sample logging driver setting up logging for a container the systemd
journal looks like this:
```go
package main
import (
"bufio"
"context"
"fmt"
"io"
"sync"
"github.com/containerd/containerd/runtime/v2/logging"
"github.com/coreos/go-systemd/journal"
)
func main() {
logging.Run(log)
}
func log(ctx context.Context, config *logging.Config, ready func() error) error {
// construct any log metadata for the container
vars := map[string]string{
"SYSLOG_IDENTIFIER": fmt.Sprintf("%s:%s", config.Namespace, config.ID),
}
var wg sync.WaitGroup
wg.Add(2)
// forward both stdout and stderr to the journal
go copy(&wg, config.Stdout, journal.PriInfo, vars)
go copy(&wg, config.Stderr, journal.PriErr, vars)
// signal that we are ready and setup for the container to be started
if err := ready(); err != nil {
return err
}
wg.Wait()
return nil
}
func copy(wg *sync.WaitGroup, r io.Reader, pri journal.Priority, vars map[string]string) {
defer wg.Done()
s := bufio.NewScanner(r)
for s.Scan() {
if s.Err() != nil {
return
}
journal.Send(s.Text(), pri, vars)
}
}
```
A `logging` package has been created to assist log developers create
logging plugins for containerd.
This uses a URI based approach for logging drivers that can be expanded
in the future.
Supported URI scheme's are:
* binary
* fifo
* file
You can pass the log url via ctr on the command line:
```bash
> ctr run --rm --runtime io.containerd.runc.v2 --log-uri binary://shim-journald docker.io/library/redis:alpine redis
```
```bash
> journalctl -f -t default:redis
-- Logs begin at Tue 2018-12-11 16:29:51 EST. --
Mar 08 16:08:22 deathstar default:redis[120760]: 1:C 08 Mar 2019 21:08:22.703 # Warning: no config file specified, using the default config. In order to specify a config file use redis-server /path/to/redis.conf
Mar 08 16:08:22 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:22.704 # You requested maxclients of 10000 requiring at least 10032 max file descriptors.
Mar 08 16:08:22 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:22.704 # Server can't set maximum open files to 10032 because of OS error: Operation not permitted.
Mar 08 16:08:22 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:22.704 # Current maximum open files is 1024. maxclients has been reduced to 992 to compensate for low ulimit. If you need higher maxclients increase 'ulimit -n'.
Mar 08 16:08:22 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:22.705 * Running mode=standalone, port=6379.
Mar 08 16:08:22 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:22.705 # WARNING: The TCP backlog setting of 511 cannot be enforced because /proc/sys/net/core/somaxconn is set to the lower value of 128.
Mar 08 16:08:22 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:22.705 # Server initialized
Mar 08 16:08:22 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:22.705 # WARNING overcommit_memory is set to 0! Background save may fail under low memory condition. To fix this issue add 'vm.overcommit_memory = 1' to /etc/sysctl.conf and then reboot or run the command 'sysctl vm.overcommit_memory=1' for this to take effect.
Mar 08 16:08:22 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:22.705 # WARNING you have Transparent Huge Pages (THP) support enabled in your kernel. This will create latency and memory usage issues with Redis. To fix this issue run the command 'echo never > /sys/kernel/mm/transparent_hugepage/enabled' as root, and add it to your /etc/rc.local in order to retain the setting after a reboot. Redis must be restarted after THP is disabled.
Mar 08 16:08:22 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:22.705 * Ready to accept connections
Mar 08 16:08:50 deathstar default:redis[120760]: 1:signal-handler (1552079330) Received SIGINT scheduling shutdown...
Mar 08 16:08:50 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:50.405 # User requested shutdown...
Mar 08 16:08:50 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:50.406 * Saving the final RDB snapshot before exiting.
Mar 08 16:08:50 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:50.452 * DB saved on disk
Mar 08 16:08:50 deathstar default:redis[120760]: 1:M 08 Mar 2019 21:08:50.453 # Redis is now ready to exit, bye bye...
```
The following client side Opts are added:
```go
// LogURI provides the raw logging URI
func LogURI(uri *url.URL) Creator { }
// BinaryIO forwards contianer STDOUT|STDERR directly to a logging binary
func BinaryIO(binary string, args map[string]string) Creator {}
```
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
The two new method Add/Delete can allow custom plugin to add or migrate
existing task into major Runtime plugin.
close: #2888
Signed-off-by: Wei Fu <fuweid89@gmail.com>
This makes bundle removal atomic by first renaming the bundle and
working directories to a hidden path before removing the underlying
directories.
Closes#2567Closes#2327
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
There is still a special case where the client side fails to open or
load causes things to be slow and the shim can lock up when this
happens. This adds a timeout to the context for this case to abort fifo
creation.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
support checkpoint without committing a checkpoint dir into a
checkpoint image and restore without untar image into checkpoint
directory. support for both v1 and v2 runtime
Signed-off-by: Ace-Tang <aceapril@126.com>
There were races with the way process states. This displayed in ways,
especially around pausing the container for atomic operations. Users
would get errors like, cannnot delete container in paused state and
such.
This can be eaisly reproduced with `docker` and the following command:
```bash
> (for i in `seq 1 25`; do id=$(docker create alpine usleep 50000);docker start $id;docker commit $id;docker wait $id;docker rm $id; done)
```
This two issues that this fixes are:
* locks must be held by the owning process, not the state operations.
* If a container ends up being paused but before the operation
completes, the process exists, make sure we resume the container before
setting the the process as exited.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
1. avoid dead lock during kill, fetch allProcesses before handle events
2. use argu's ctx instead of context.Backgroud() in openlog
Signed-off-by: Wei Fu <fuweid89@gmail.com>
Fixes#2709
This increases the buffer size for process exit subscribers. It also
implements a non-blocking send on the subscriber channel. It is better
to drop an exit even than it is to block a shim for one slow subscriber.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
func (e *execProcess) delete(ctx context.Context) error {
e.wg.Wait()
...
}
delete exec process will wait for io copy finish, if wait here,
other process can not get lock of shim service.
1. apply lock around s.transition() calls in the Delete methods.
2. put lock after wait io copy in exec Delete.
Signed-off-by: Ace-Tang <aceapril@126.com>
This makes sure that runc does not get any valid IO for the pipe. Some
builds and other containers will be stuck if they inspect stdin
expecially and its a pipe but not connected to any user input.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
- Still KillAll if the task uses the hosts pid namespace
- Test for both host pid namespace and normal cases
Co-authored-by: Oliver Stenbom <ostenbom@pivotal.io>
Co-authored-by: Georgi Sabev <georgethebeatle@gmail.com>
Signed-off-by: Oliver Stenbom <ostenbom@pivotal.io>
This was found testing other runtime shims that are faster than runc(no
containerization). This is a race that can cause the shim to block
forever. It's not an issue for out/err because we open both sides of
the pipe, but for stdin, it expects the client to have it opened.
Signed-off-by: Michael Crosby <crosbymichael@gmail.com>
