RFR: 8017629: G1: UseSHM in combination with a G1HeapRegionSize > os::large_page_size() falls back to use small pages

[Stefan Karlsson]

Hi Thomas,

On 2016-04-11 14:39, Thomas Stüfe wrote:
> Hi Stefan,
>
> short question, why the mmap before the shmat? Why not shmat right 
> away at the requested address?

If we have a requested_address we do exactly what you propose.

   if (req_addr == NULL && alignment > os::large_page_size()) {
     return shmat_with_large_alignment(shmid, bytes, alignment);
   } else {
     return shmat_with_normal_alignment(shmid, req_addr);
   }

...

static char* shmat_with_normal_alignment(int shmid, char* req_addr) {
   char* addr = (char*)shmat(shmid, req_addr, 0);

   if ((intptr_t)addr == -1) {
     shm_warning_with_errno("Failed to attach shared memory.");
     return NULL;
   }

   return addr;
}


It's when you don't have a requested address that mmap is used to find a 
large enough virtual memory area.

>
> Also note that mmap- and shmat-allocated memory may have different 
> alignment requirements: mmap requires a page-aligned request address, 
> whereas shmat requires alignment to SHMLBA, which may be multiple 
> pages (e.g. for ARM: 
> http://lxr.free-electrons.com/source/arch/arm/include/asm/shmparam.h#L9). 
> So, for this shat-over-mmap trick to work, request address has to be 
> aligned to SHMLBA, not just page size.
>
> I see that you assert alignment of requ address to 
> os::large_page_size(), which I would assume is a multiple of SHMLBA, 
> but I am not sure of this.

I've added some defensive code and asserts to catch this if/when this 
assumption fails:

http://cr.openjdk.java.net/~stefank/8017629/webrev.02.delta/
http://cr.openjdk.java.net/~stefank/8017629/webrev.02

I need to verify that this works on other machines than my local Linux 
x64 machine.

Thanks,
StefanK

>
> Kind Regards, Thomas
>
>
>
> On Mon, Apr 11, 2016 at 1:03 PM, Stefan Karlsson 
> <stefan.karlsson at oracle.com <mailto:stefan.karlsson at oracle.com>> wrote:
>
>     Hi all,
>
>     Please review this patch to enable SHM large page allocations even
>     when the requested alignment is larger than os::large_page_size().
>
>     http://cr.openjdk.java.net/~stefank/8017629/webrev.01
>     <http://cr.openjdk.java.net/%7Estefank/8017629/webrev.01>
>     https://bugs.openjdk.java.net/browse/JDK-8017629
>
>     G1 is affected by this bug since it requires the heap to start at
>     an address that is aligned with the heap region size. The patch
>     fixes this by changing the UseSHM large pages allocation code.
>     First, virtual memory with correct alignment is pre-reserved and
>     then the large pages are attached to this memory area.
>
>     Tested with vm.gc.testlist and ExecuteInternaVMTests
>
>     Thanks,
>     StefanK
>
>