Add limit for MemorySegment.allocateNative

[Maurizio Cimadamore]

Hi Robert,
you are right in that the foreign memory API does not respect the memory 
limit set by XX:MaxDirectMemorySize; and we're aware if that. The reason 
as to why we did not add a limit (or reuse existing ones) is that the 
allocation story is still in flux; right now memory allocation uses 
separate mallocs (as ByteBuffer.allocateDirect does). But it is possible 
that, in the not-so-distant future we will add other ways to allocate 
memory, which will yield significantly better allocation performances 
over malloc. Since the allocators we were looking at were based on the 
idea of reserving a fixed size of memory at VM startup (say 2G) and then 
commit said memory incrementally, as allocation requests are coming in, 
limiting the allocation size in a scheme such as this one is much much 
simpler, and requires no locking (if you run out of virtual address 
space, you simply fail to allocate). This is the main reason as to why 
we didn't tackle the memory limit problem (yet), but we'll get there - 
hopefully soon.

That said, if you happen to have a patch ready, it'd be useful to take a 
look and evaluate the impact of the size check on the segment allocation 
performances (IIRC that logic does some locking/CASing in order to 
ensure consistency across multiple threads).

Maurizio

On 21/06/2020 04:59, Robert Lu wrote:
> The jdk.incubator.foreign is really fast, but unlike
> ByteBuffer.allocateDirect , there is no such limit like
> XX:MaxDirectMemorySize .
>
> How could the user set max memory size used by jdk.incubator.foreign?
>
> Could I submit a patch to count jdk.incubator.foreign's memory to
> java.nio.Bits#RESERVED_MEMORY, and follow the limit of
> XX:MaxDirectMemorySize?
>
>
>