RFR 8021820: Number of opened files used in select() is limited to 1024 [macosx]
Alan.Bateman at oracle.com
Wed Jul 31 13:45:47 UTC 2013
On 31/07/2013 05:18, Aleksej Efimov wrote:
> Can I have a review for the following problem:
> The MACOSX JDK (more precisely - the java.net classes) uses the
> select() system call to wait for different events on sockets fds. And
> the default behaviour for select() on Darwin is to fail when fdset
> contains the fd with id greater than FDSET_SIZE(=1024). Test case in
> webrev illustrates this behavior.
> There is at least one solution for it: use -D_DARWIN_UNLIMITED_SELECT
> compilation flag for all macosx sources: this won't affect other parts
> of JDK because they are not using select().
> Currently, I have added this compilation flag to
> common/autoconf/generated-configure.sh and
> common/autoconf/generated-configure.sh. I wonder, if there is a better
> place where I can put this flag?
> The webrev: http://cr.openjdk.java.net/~aefimov/8021820/webrev.00/
> BUG: http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=8021820
Thanks for looking into this one. The build changes look okay to me but
it's probably best that someone on build-dev agree to those. Michael
McMahon can probably explain why the net code is using select for timed
read/accept (I have a vague recollection of there being an issue with
poll due to the way that it is implemented on kqueue with the result
that it had to be changed to use select).
I think the test needs re-work. It looks to me that the it attempts to
connect to an Oracle internal site so that's not going to work
everywhere. In general we don't want the tests to be dependent on hosts
that may or may not exist (we had tests that used to this in the past
but they caused a lot of grief). It also looks like the test doesn't
close the 1023 files that it opens at the start and so I assume this
test will always fail on Windows when jtreg tries to clean-up.
More information about the core-libs-dev