Re: RFR(S): 8073042: jcmd hangs until another jcmd is executed (which, in turn, also hangs) (on Windows)

[Dean Long]

Hi Markus.  OK, thanks for the explanation.

dl

On 3/3/2015 1:15 AM, Markus Gronlund wrote:

Hi Dean,

Thanks for your input.

I think a monitor would be less ideal here:

A monitor would introduce blocking in enqueuing operations.

The "posters" are (in this case) remotely injected threads (by another 
process), and the processes injecting those threads are in WaitForSingleObject() on those 
thread handles. If a monitor is used, it needs to be held by the AttachListener over 
executing (arbitrary) operations. Granted, each operation today is being executed in 
serial fashion (but it does not need to be), but I believe we would not like to have each 
client also block on the monitor as part of enqueuing their operation, just in order to 
issue the notify.

With a Windows Semaphore, clients can act asynchronously in regard to enqueuing. 
As long as the semaphore's current count is > 0, there is no blocking on the 
part of the AttachListener thread, WaitForSingleObject() return immediately 
decrementing the current count.

Thanks
Markus


-----Original Message-----
From: Dean Long
Sent: den 3 mars 2015 02:32
To: Markus Gronlund; serviceability-dev@openjdk.java.net; 
hotspot-runtime-...@openjdk.java.net
Subject: Re: RFR(S): 8073042: jcmd hangs until another jcmd is executed (which, 
in turn, also hangs) (on Windows)

Couldn't you instead treat it like a monitor?  Then it's OK if only the first 
notify wakes up the blocked thread, as long as that thread only blocks when the 
queue is empty.  In other words, when it wakes up, it should process all the 
items in the queue before blocking again.

dl

On 3/2/2015 4:34 AM, Markus Gronlund wrote:

Greetings,

Kindly asking for reviews for the following changeset:

Bug: https://bugs.openjdk.java.net/browse/JDK-8073042

Webrev: http://cr.openjdk.java.net/~mgronlun/8073042/webrev01/

Description:

The signaling mechanism used to communicate about attaching operations under 
Windows currently only allows for a single outstanding item to be visible. This 
leads to issues, such as the one described in this bug, where clients assume 
their operations are under processing (they have been enqueued after all), but 
the AttachListener thread does not see, and hence do not process, these 
operations.

Analysis:

The _wakeup semaphore only allows for a single outstanding operation:

CreateSemaphore(NULL, 0, 1, NULL);

When a thread has enqueued an operation, it will notify the AttachListener 
thread through the semaphore, by:

::ReleaseSemaphore(wakeup(), 1, NULL); // this increases the semaphore
count by 1

This will signal the semaphore and "wakeup" the AttachListener thread which 
(most likely) is in WaitForSingleObject() for the semaphore to become signaled. When the 
semaphore is signaled and AttachListener returns from WaitForSingleObject(), the 
semaphore's count is decremented, and the semaphore again becomes non-signaled (goes from 
current count 1 (which is also maximum count) to zero).


Problem: multiple client threads will enqueue their operations if they manage 
to get the list mutex() - if they do, they will insert their op into the queue, 
and call:

::ReleaseSemaphore(wakeup(), 1, NULL);

This means we could have two (or more) client threads having posted operations, 
before the AttachListener thread becomes scheduled.

Since the semaphore created has a maximum count == 1, any subsequent calls to 
::ReleaseSemaphore(wakeup(), 1, NULL);, taking the the current count to > 
maximum count, will have _no_effect - the current count of the semaphore remains 
at maximum count level AND ::ReleaseSemaphore(wakeup(), 1, NULL); returns FALSE - 
but currently there is no checking the return value here...
This means the client thread has managed to enqueue its op (and will move into 
ConnectPipe()), but the AttachListener will never see more than 1 enqueued op 
at any one time (hence it does not know it is expected to process another 
operation and signal the associated pipe).

This is how operations manage to get enqueued, but not processed until another 
thread eventually signals the semaphore by posting another op.

We must allow the semaphore to stay signaled when multiple ops are enqueued - 
and since we only allow preallocate_count number of operations to be enqueued, 
we can ensure the semaphore manages this upper limit as its maximum count.

Tests:

Jcmd/dcmd test

Attach tests

Thanks

Markus