Minor correction: EPERM -> EACCES for Solaris
Hard to see how to get a transient EACCES when opening a file ... though
as it is really a door I guess there could be additional complexity.
David
On 3/10/2018 7:54 AM, Chris Plummer wrote:
On 10/2/18 2:38 PM, David Holmes wrote:
Chris,
On 3/10/2018 6:57 AM, Chris Plummer wrote:
On 10/2/18 1:44 PM, gary.ad...@oracle.com wrote:
The general attach sequence ...
src/jdk.attach/solaris/classes/sun/tools/attach/VirtualMachineImpl.java
the attacher creates an attach_pid file in a directory where the
attachee is runnning
issues a signal to the attacheee
loops waiting for the java_pid file to be created
default timeout is 10 seconds
So getting a FileNotFoundException while in this loop is OK, but
IOException is not.
src/hotspot/os/solaris/attachListener_solaris.cpp
attachee creates the java_pid file
listens til the attacher opens the door
I'm don't think this is related, but JDK-8199811 made a fix in
attachListener_solaris.cpp to make it wait up to 10 seconds for
initialization to complete before failing the enqueue.
...
Not sure when a bare IOException is thrown rather than the
more specific FileNotFoundException.
Where is the IOException originating from? I wonder if the issue is
that the file is in the process of being created, but is not fully
created yet. Maybe it is there, but owner/group/permissions have not
been set yet, and this results in an IOException instead of
FileNotFoundException.
The exception is shown in the bug report:
[java.io.IOException: Permission denied
at jdk.attach/sun.tools.attach.VirtualMachineImpl.open(Native Method)
at
jdk.attach/sun.tools.attach.VirtualMachineImpl.openDoor(VirtualMachineImpl.java:215)
at
jdk.attach/sun.tools.attach.VirtualMachineImpl.<init>(VirtualMachineImpl.java:71)
at
jdk.attach/sun.tools.attach.AttachProviderImpl.attachVirtualMachine(AttachProviderImpl.java:58)
at
jdk.attach/com.sun.tools.attach.VirtualMachine.attach(VirtualMachine.java:207)
at jdk.jcmd/sun.tools.jcmd.JCmd.executeCommandForPid(JCmd.java:114)
at jdk.jcmd/sun.tools.jcmd.JCmd.main(JCmd.java:98)
And if you look at the native code the EPERM from open will cause
IOException to be thrown.
./jdk.attach/solaris/native/libattach/VirtualMachineImpl.c
JNIEXPORT jint JNICALL Java_sun_tools_attach_VirtualMachineImpl_open
(JNIEnv *env, jclass cls, jstring path)
{
jboolean isCopy;
const char* p = GetStringPlatformChars(env, path, &isCopy);
if (p == NULL) {
return 0;
} else {
int fd;
int err = 0;
fd = open(p, O_RDWR);
if (fd == -1) {
err = errno;
}
if (isCopy) {
JNU_ReleaseStringPlatformChars(env, path, p);
}
if (fd == -1) {
if (err == ENOENT) {
JNU_ThrowByName(env, "java/io/FileNotFoundException",
NULL);
} else {
char* msg = strdup(strerror(err));
JNU_ThrowIOException(env, msg);
if (msg != NULL) {
free(msg);
}
We should add the path to the exception message.
Thanks David. So if EPERM is the error and a retry 100ms later works, I
think that supports my hypothesis that the file is not quite fully
created. So Gary's fix is probably fine. The only other possible fix I
can think of that wouldn't require an explicit delay (or multiple
retries) is probably not worth the complexity. It would require that the
attachee create two files, and the attacher try to open the second file
first. When it either opens or returns EPERM, you know the first file
can safety be opened.
Chris
David
-----
Chris
On 10/2/18 4:11 PM, Chris Plummer wrote:
Can you summarize how the attach handshaking is suppose to work?
I'm just wondering why the attacher would ever be looking for the
file before the attachee has created it. It seems a proper
handshake would prevent this. Maybe there's some sort of visibility
issue where the attachee has indeed created the file, but it is not
immediately visible to the attacher process.
Chris
On 10/2/18 12:27 PM, gary.ad...@oracle.com wrote:
The problem reproduced pretty quickly.
I added a call to checkPermission and revealed the
"file not found" from the stat call when the IOException
was detected.
There has been some flakiness from the Solaris test machines today,
so I'll continue with the testing a bit longer.
On 10/2/18 3:12 PM, Chris Plummer wrote:
Without the fix was this issue easy enough to reproduce that you
can be sure this is resolving it?
Chris
On 10/2/18 8:16 AM, Gary Adams wrote:
Solaris debug builds are failing tests that use the attach
interface.
An IOException is reported when the java_pid file is not opened.
It appears that the attempt to attach is taking place too quickly.
This workaround will allow the open operation to be retried
after a short pause.
Webrev: http://cr.openjdk.java.net/~gadams/8210337/webrev/
Issue: https://bugs.openjdk.java.net/browse/JDK-8210337
Testing is in progress.
