RFR 8012326: Deadlock occurs when Charset.availableCharsets() is called by several threads at the same time (original) (raw)

Xueming Shen xueming.shen at oracle.com
Fri May 10 17:53:56 UTC 2013

Thanks for the review.

I have updated the Charset.java to use the init-on-depand holder idiom.

I don't think MSISO2022JP/ISO2022_JP_2 are really worth the cost of adding two more classes, so I leave them asis.

Agreed that the ExtendedCahrsets change can be separated, but since we're here already, It would be convenient to just do them together and the "cleaner' code then can be back-port together later when requested.

http://cr.openjdk.java.net/~sherman/8012326/webrev/

-Sherman

On 05/08/2013 12:49 PM, Mandy Chung wrote:

Hi Sherman,

I reviewed webrev and also checked out webrev.newECP. While it's good to simplify the synchronization needed in ExtendedCharsets, maybe better to separpate the ExtendedCharsets change from the fix for 8012326 that would require a more detailed review (I can do that next). Charset.java ExtendedCharsets provider only needs to be instantiated once. The ECP initialization can be simplified by using the initialization-on-demand holder idiom to replace the existing initialization logic using the extendedProviderLock and extendedProviderProbed. You could also simply have a holder class to define the lookupExtendedCharset (probably better to rename it to "charsetForName") and charsets() methods to return one or all available charsets from the extended provider respectively. I think that would be cleaner and easy to read. MSISO2022JP.java and ISO2022JP2.java Similiarly, it'd be cleaner to use a holder class to initialize the DEC02XX and ENC02XX variable. Mandy On 5/4/2013 12:05 PM, Xueming Shen wrote: Thanks Ulf!

There is another version with a new ExtendedCharsets.java at http://cr.openjdk.java.net/~sherman/8012326/webrev.newECP/ I merged the stuff in AbstractCharsetProvider into ExtendedCharsets.java. The standardcharset provider now uses the FastCharsetProvider, so there is no need to have an abstract class anymore, as long as we are not going to add a new or separate the charsets.jar. I kinda remember there was a plan to further divide the charsets.jar in the past though... I took the chance to "clean up" the synchronization mechanism in ExtendedCharsets. It appears there are two sync needs here. One is to protect the "cache" inside lookup(), which triggers the race condition if the lookup() gets invoked by multiple threads and the "cahce" map gets accessed/updated at the same time, this is reported and fixed by 4685305 [1], the original fix is to put the sync block in AbstractCharsetProvider.charsetForName(). We put in another sync block in iterator.next() for 6898310 [2], which is the trigger of this bug. In the new version, I "consolidated" them together into lookup() Another sync need is for the "init()", in which it may update the aliasMap, classMap and aliasNameMap via charset() and deleteCharset() if those special properties are defined. There are two sources for the charset()/ deleteCharset(), one is from the constructor, one is from the init(), given ExtendedCharsets is now singleton and get initialized at class init, these should be no race concern between these two sources, so no need to have any sync block inside charset() and deleteCharset(), the only thing need to defend is inside init(), and all three public methods invoke the init() at the beginning of the method body. It appears I will still have to keep the same logic in Charset to access the ExtendedCharset, as it is need to be "probed", just in case it is not "installed"... Yes, there is also room to improve in FastCharsetProvider...I know I need pick some time on it. -Sherman On 5/4/13 10:09 AM, Ulf Zibis wrote: Hi Sherman,

looks good to me. Maybe you like to compare with webrevs from: https://bugs.openjdk.java.net/showbug.cgi?id=100092 https://bugs.openjdk.java.net/showbug.cgi?id=100095 -Ulf Am 03.05.2013 06:29, schrieb Xueming Shen: Hi,

Please help review the proposed fix for 8012326. The direct trigger is the fix we put in #6898310 [1], in which we added the synchronization to prevent a possible race condition as described in $4685305. However it appears the added synchronization triggers another race condition as showed in this stack trace [4] when run the test case [2][3]. The root cause here is (1) The ExtendedCharsetProvider is intended to be used as a singleton (as the probeExtendedProvider + lookupExtendedCharset mechanism in Charset.java), however this is only true for the Charset.forName()->lookup()...scenario. Multiple instances of ExtendedCharsetProvider is being created in Charset.availableCharset() every time it is invoked, via providers()/ServiceLoader.load(). (2) ISO2022JP2 and MSISO2022JP are provided via ExtendedCharsetProvider, however both of them have two static variable DEC02{12|08}/ENC02{12|08} that need to be initialized during the "class initialization", which will indirectly trigger the invocation of ExtendedCharsetProvider.alisesFor(...) (3) static method ExtendedCharsets.aliaseFor() has a hacky implementation that goes to AbstractCharsetProvider.alise() for lookup, which needs to obtain a lock on its ExtendedCharesetProvider.instance. This will potential cause race condition if the "instance" it tries to synchronize on is not its "own" instance. This is possible because the constructor of ExtendedCharsetProvider overwrites static "instance" variable with "this". Unfortunately as demonstrated in [4], this appears to be what is happening. The Thread-1/#9 is trying to synchronize on someone else's ExtendedCharsetProvider instance <0xa4824730> (its own instance should be <0xa4835328>, which it holds the monitor in the iterator.next()), it is blocked because Thread-0 already holds the monitor of <0xa4824730> (in its iteratior.next()), but Thread-0 is blocked at Class.forName0(), which I think is because Thread-1 is inside it already...two theads are deadlocked. A "quick fix/solution" is to remove the direct trigger in ISO2022JP2 and MSISO2022JP to lazily-initialize those static variables as showed in the webrev. However while this probably will solve the race condition, the multiple instances of ExtendedCharsetProvider is really not desired. And given the fact we have already hardwired ExtendedCharsetProvider (as well as the StandardCharset, for performance reason) for charset lookup/forName(), the availableCharsets() should also utilize the "singleton" ExtendedCharsetProvider instanc (extendedCharsetProvider) as well, instead of going to the ServiceLoader every time for a new instance. The suggested change is to use Charset. extendedCharsetProvide via probeExtendedProvider() for extended charset iteration (availableCharset()). Meanwhile, since the ExtendedCharsetProvider/ charsets.jar should always be in the boot classpath (if installed, and we are looking up via Class.forName("sun.nio.cs.ext.ExtededCharsetProvider")), there is really no need for it to be looked up/loaded via the spi mechanism (in fact it's redundant to load it again after we have lookup/iterate the hardwired "extendedCharsetProvider". So I removed the spi description from the charsts.jar as well. An alternative is to really implement the singleton pattern in ECP, however given the existing mechanism we have in Charset.java and the "hacky" init() implementation we need in ECP, I go with the current approach. http://cr.openjdk.java.net/~sherman/8012326/webrev Thanks, Sherman [1] http://cr.openjdk.java.net/~sherman/6898310/webrev/ [2] http://cr.openjdk.java.net/~sherman/8012326/runtest.bat [3] http://cr.openjdk.java.net/~sherman/8012326/Test.java [4] http://cr.openjdk.java.net/~sherman/8012326/stacktrace

More information about the core-libs-dev mailing list