(RFR)(S)(10): 8176768: hotspot ignores PTHREAD_STACK_MIN when creating new threads
Thomas Stüfe
thomas.stuefe at gmail.com
Tue Mar 21 17:46:31 UTC 2017
Hi Chris,
On Tue, Mar 21, 2017 at 6:43 PM, Chris Plummer <chris.plummer at oracle.com>
wrote:
> Hi Thomas,
>
> On 3/21/17 1:08 AM, Thomas Stüfe wrote:
>
> Hi Chris,
>
> looks fine to me.
>
> small nit: os_linux_cpp:
>
>
> + // In that cast just subract the page size to get the maximum possible
> stack size.
>
> subract->substract
>
> Ok. I had the same typo in os_posix.cpp.
>
>
> I would have preferred SIZE_MAX instead of UINT_MAX, but both is fine to
> me.
>
> I'll use SIZE_MAX. I didn't see it in limits.h so I didn't know of it's
> existence. Turns out it is in stdint.h.
>
>
> --
>
> As for the size_overflow problem, I wonder whether we are overthinking
> this. Maybe a more pragmatic approach would have been to just define a
> reasonable upper maximum for stack sizes which could be well below
> UINT_MAX. Somehow I cannot see users wanting to start threads with more
> than ~4GB of thread stack size.
>
> I had debated doing this. I decided not to artificially limit the stack
> size for a couple of reasons. (1) I figured we should make a best effort to
> use the size the user requested. (2) There is no obvious large size to
> choose other than the max size. If we choose a smaller size it might still
> fail or might be smaller than the largest size that will succeed.
>
>
Makes sense.
> Thanks again for the review.
>
>
Sure, thanks for taking my input.
Kind regards, Thomas
Chris
>
>
> Kind Regards, Thomas
>
>
>
> On Tue, Mar 21, 2017 at 12:29 AM, Chris Plummer <chris.plummer at oracle.com>
> wrote:
>
>> On 3/17/17 11:37 PM, Chris Plummer wrote:
>>
>>> On 3/17/17 8:17 PM, Chris Plummer wrote:
>>>
>>>> On 3/17/17 7:01 PM, David Holmes wrote:
>>>>
>>>>> On 18/03/2017 9:11 AM, Chris Plummer wrote:
>>>>>
>>>>>> Looks like this will need some more work since the added asserts are
>>>>>> triggering on mac os x (which is the only place we'd currently expect
>>>>>> them to assert).
>>>>>>
>>>>>> The problem is that the code I found that rounds up to the page size
>>>>>> is
>>>>>> only applied to java threads created by the VM for which the java user
>>>>>> specified no stack size. The VM and Compiler thread sizes are not
>>>>>> rounded. The failure I saw was with
>>>>>> runtime/CommandLine/OptionsValidation/TestOptionsWithRanges.java
>>>>>> when is
>>>>>> specified -XX:CompilerThreadStackSize=9007199254740991. I hit the
>>>>>> assert
>>>>>> with an EINVAL. The size is not aligned, but it could also be
>>>>>> complaining because it is too big. I haven't tried aligning it yet to
>>>>>> see.
>>>>>>
>>>>>> On Linux we do the following:
>>>>>>
>>>>>> stack_size = align_size_up(stack_size +
>>>>>> os::Linux::default_guard_size(thr_type), vm_page_size());
>>>>>>
>>>>>> We don't do this on BSD. I think the same on BSD would solve this
>>>>>> problem. I'm not sure about adding the guard size. I'll need to see if
>>>>>> mac os x has the same pthread bug as linux does.
>>>>>>
>>>>>
>>>>> At this stage I would only deal with alignment issues. IIRC the guard
>>>>> issue only affected Linux.
>>>>>
>>>> Yes, that's what I eventually concluded. I put the fix in
>>>> os::Posix::get_initial_stack_size() in os_posix.cpp, and only did the
>>>> page aligning, not add the guard page. That way all Posix ports are fixed
>>>> in one place. It seems to be working.
>>>>
>>>>>
>>>>> BTW, did you know java users can specify the size of the a new thread's
>>>>>> stack:
>>>>>>
>>>>>
>>>>> Yes I mentioned that in another reply - wondering whether we suitably
>>>>> check and aligned such requests.
>>>>>
>>>> No we don't. Below I mentioned I was able to trigger the assert with a
>>>> 257k stack size. I guess I wasn't clear that I did that from Java. I have a
>>>> new test to add that will be testing this, plus the 9007199254740991 stack
>>>> size (which fails to create the thread with an OOME, but that's
>>>> acceptable). The fix I mention above in os::Posix::get_initial_stack_size()
>>>> takes care of this issue also.
>>>>
>>> Rounding up triggers a new assert, this time on 32-bit x86 and arm.
>>>
>>> I should have clarified it's 9007199254740991 "K", which is
>>> 9223372036854774784. Unfortunately on 32bit systems that is asserting with
>>> EINVAL. I think we need to do a better job of dealing with 32-bit size_t
>>> values:
>>>
>>> jlong java_lang_Thread::stackSize(oop java_thread) {
>>> if (_stackSize_offset > 0) {
>>> return java_thread->long_field(_stackSize_offset);
>>> } else {
>>> return 0;
>>> }
>>> }
>>>
>>> jlong size =
>>> java_lang_Thread::stackSize(JNIHandles::resolve_non_null(jthread));
>>> // Allocate the C++ Thread structure and create the native
>>> thread. The
>>> // stack size retrieved from java is signed, but the constructor
>>> takes
>>> // size_t (an unsigned type), so avoid passing negative values
>>> which would
>>> // result in really large stacks.
>>> size_t sz = size > 0 ? (size_t) size : 0;
>>> native_thread = new JavaThread(&thread_entry, sz);
>>>
>>> 9223372036854774784 is 0x7ffffffffffffc00 (close to 64 bit MAX_INT),
>>> which is 0xfffffc00 when cast to a size_t on a 32-bit system (close to
>>> 32-bit MAX_UINT). Round it up to the 4k page size and you get 0, which I
>>> guess pthread_attr_setstacksize() doesn't like. So I think more processing
>>> of the size is needed here. Maybe in os::create_thread() we should check
>>> for 0 after rounding up, and subtract the os page size if it is 0. However,
>>> I think we should also avoid truncating on 32-bit to what is basically some
>>> random number. Maybe if "size" (a jlong) is greater than UINT_MAX, then
>>> just set "sz" (a size_t) it to UINT_MAX.
>>>
>>> Ok, I think I have this all worked out now. I've added fixes for
>> unaligned stack sizes, 32-bit truncating of stack size, and the "aligning
>> up to 0" problem. I also added a test. Here's the latest webrev:
>>
>> http://cr.openjdk.java.net/~cjplummer/8176768/webrev.02/webrev.hotspot
>>
>> Here's what's changed since webrev.01:
>>
>> os_posix.cpp: In os::Posix::get_initial_stack_size(), first round up the
>> stack size to be paged aligned. This fixes issues on Mac OS X (other
>> platforms seem to be immune to this). Then check if the size is zero after
>> rounding up to the page size. Subtract the page size in this case to
>> produce the maximum stack size allowed. Surprisingly I got no complaint
>> from gcc for subtracting from an unsigned value that is known to be 0.
>>
>> os_linux.cpp: In os::create_thread(), I also check here to make sure the
>> size is not 0 after adding the guard page and aligning up, and subtract the
>> os page size if it is 0.
>>
>> jvm.c: In JVM_StartThread(), on 32-bit platforms if the size is greater
>> than UINT_MAX, then I set the size to UINT_MAX. Note it will later be
>> rounded up to 0 in os::Posix::get_initial_stack_size(), which will
>> result in subtracting the os page size to get the actual maximum allowed
>> stack size.
>>
>> TooSmallStackSize.java: added test case for unaligned stack sizes.
>>
>> TestThreadStackSizes.java: New test. Creates new threads with every size
>> up to 320k in 1k increments. Then creates threads with a few other sizes
>> that can be problematic.
>>
>> thanks,
>>
>> Chris
>>
>> Chris
>>>
>>>
>>>
>>>>> public Thread(ThreadGroup group, Runnable target, String name,
>>>>>> long stackSize) {
>>>>>> init(group, target, name, stackSize);
>>>>>> }
>>>>>>
>>>>>> Fortunately we already force the stackSize to be at least
>>>>>> _java_thread_min_stack_allowed. However, we don't do any OS page
>>>>>> rounding on Mac OS X as noted above, and I was able to trigger the
>>>>>> assert by creating a thread with size 257k.
>>>>>>
>>>>>
>>>>> Note this means that OSX stack logic is broken because it will
>>>>> currently be silently failing due to EINVAL!
>>>>>
>>>> Yes, that is correct.
>>>>
>>>> Chris
>>>>
>>>>>
>>>>> I'll get another webrev out once I've made the needed fixes. I also
>>>>>> have
>>>>>> a new test I'd like to add.
>>>>>>
>>>>>
>>>>> Ok.
>>>>>
>>>>> Thanks,
>>>>> David
>>>>>
>>>>> Chris
>>>>>>
>>>>>> On 3/16/17 9:27 PM, Chris Plummer wrote:
>>>>>>
>>>>>>> Ok, time for a new webrev:
>>>>>>>
>>>>>>> http://cr.openjdk.java.net/~cjplummer/8176768/webrev.01/webr
>>>>>>> ev.hotspot
>>>>>>>
>>>>>>> The only thing that has changed since the first webrev are the
>>>>>>> asserts
>>>>>>> added to os_linux.cpp and os_bsd.cpp. And to summarize what we
>>>>>>> discuss
>>>>>>> already:
>>>>>>>
>>>>>>> - The assert should never happen due to the stack size being too
>>>>>>> small since it will be at least PTHREAD_STACK_MIN.
>>>>>>> - The assert should never happen due to an unaligned stack size
>>>>>>> because we always align it to the page size.
>>>>>>> - Any assert would therefore be a VM bug and not due to user error.
>>>>>>> - No fixing the java launcher. If the user specifies a stack that is
>>>>>>> too small, hotspot will already detect this. If the user specifies a
>>>>>>> stack size that is large enough but not page aligned, then we just
>>>>>>> ignore any error (if the platform doth protest) and the user gets a
>>>>>>> main thread with a stack size set to whatever the OS default is.
>>>>>>>
>>>>>>> I still need to retest (I only ran TooSmallStackSize.java), but
>>>>>>> figure
>>>>>>> getting agreement on the changes first would be best before I bog
>>>>>>> down
>>>>>>> our testing resources.
>>>>>>>
>>>>>>> thanks,
>>>>>>>
>>>>>>> Chris
>>>>>>>
>>>>>>> On 3/15/17 10:03 PM, Chris Plummer wrote:
>>>>>>>
>>>>>>>> Hello,
>>>>>>>>
>>>>>>>> Please review the following:
>>>>>>>>
>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8176768
>>>>>>>> http://cr.openjdk.java.net/~cjplummer/8176768/webrev.00/webr
>>>>>>>> ev.hotspot
>>>>>>>>
>>>>>>>> While working on 8175342 I noticed our stack size on xgene was 8mb
>>>>>>>> even though I was specifying -Xss72k. It turns out the following
>>>>>>>> code
>>>>>>>> was failing:
>>>>>>>>
>>>>>>>> pthread_attr_setstacksize(&attr, stack_size);
>>>>>>>>
>>>>>>>> Although we computed a minimum stack size of 72k, so -Xss72k should
>>>>>>>> be fine, pthreads on this platform requires the stack be at least
>>>>>>>> 128k, so it failed the pthread_attr_setstacksize() call. The end
>>>>>>>> result is pthread_attr_setstacksize() had no impact on the thread's
>>>>>>>> stack size, and we ended up with the platform default of 8mb. The
>>>>>>>> fix
>>>>>>>> is to round up the following variables to PTHREAD_STACK_MIN after
>>>>>>>> computing their new values:
>>>>>>>>
>>>>>>>> _java_thread_min_stack_allowed
>>>>>>>> _compiler_thread_min_stack_allowed
>>>>>>>> _vm_internal_thread_min_stack_allowed
>>>>>>>>
>>>>>>>> For solaris, there was an issue using PTHREAD_STACK_MIN. You need to
>>>>>>>> #define _POSIX_C_SOURCE >= 199506L in order to get PTHREAD_STACK_MIN
>>>>>>>> #defined, and this needs to be done before including OS header
>>>>>>>> files.
>>>>>>>> I noticed that on solaris we were using thr_min_stack() elsewhere
>>>>>>>> instead of PTHREAD_STACK_MIN, so I decided to do the same with this
>>>>>>>> fix. Either way is ugly (the #define or using thr_min_stack()).
>>>>>>>>
>>>>>>>> And speaking of the existing use of thr_min_stack(), I deleted it.
>>>>>>>> It
>>>>>>>> was being applied before any adjustments to the stack sizes had been
>>>>>>>> made (rounding and adding red, yellow, and shadow zones). This mean
>>>>>>>> the stack ended up being larger than necessary. With the above fix
>>>>>>>> in
>>>>>>>> place, we are now applying thr_min_stack() after recomputing the
>>>>>>>> minimum stack sizes. If for any reason one of those stack sizes is
>>>>>>>> now too small, the correct fix is to adjust the initial stack sizes,
>>>>>>>> not apply thr_min_stack() to the initial stack sizes. However, it
>>>>>>>> looks like no adjustment is needed. I did something close to our
>>>>>>>> nightly testing on all affect platforms, and no new problems turned
>>>>>>>> up.
>>>>>>>>
>>>>>>>> thanks,
>>>>>>>>
>>>>>>>> Chris
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>
>>>
>>
>
>
More information about the hotspot-dev
mailing list