On 09/08/2015 12:40 PM, Andrew Haley wrote:
On 09/08/2015 06:03 PM, David M. Lloyd wrote:
On 09/08/2015 08:11 AM, Andrew Haley wrote:
On 09/08/2015 12:37 PM, David M. Lloyd wrote:
Then you do the real unmap when
the buffer is GC'd (maybe via Cleaner). This is very akin to how file
descriptors are cleaned up, AFAICT.
Indeed it is, and it's no better than the status quo. This is an attempt
to do better.
Actually it *is* better, as it releases the mapping to the underlying
file, and does not require any resources (outside of the page table
itself and any associated kernel bookkeeping) to maintain a valid state
during the time window between user-driven unmap and automatic GC. The
underlying file can (for example) immediately relinquish reserved disk
storage or RAM at unmap time (the desired effect), rather than waiting
for GC to come around (status quo).
I believe that reclaiming the reserved address space is of much less
importance than freeing the backing resources (especially for 64-bit
systems), which is why I suggest this approach.
I think you may be assuming rather a lot about the way that the
underlying OS actually works. The bug report mentions race conditions
when remapping, and not every OS has nice UNIXy semantics. Hence my
proposal, which might be portable to other OSs.
It's at least been in POSIX since 2008: "This volume of POSIX.1‐2008
specifies that [..] new [MAP_FIXED] mappings replace any existing
mappings, following existing practice in this regard." Linux agrees:
"If the memory region specified by addr and len overlaps pages of any
existing mapping(s), then the overlapped part of the existing mapping(s)
will be discarded." BSD seems to agree as well, as does Solaris, and I
can't seem to find any public reference to known race conditions with
this kind of operation.
I think the only questionable platform at this point is Windows (to me
at least, given that I do not have much experience on it).
--
- DML
