consider applying MADV_DONTNEED to snapshot data section after load

[mraleph]

Once snapshot is unpacked into isolate memory it is likely that it will not be needed again (especially if lightweight isolates are enabled). For environments with strict memory constraints we should consider findings ways to let OS release this memory and decrease RSS.

/cc @mkustermann @rmacnak-google

[mkustermann]

If the memory is mmap()ed in RO mode, the OS always has the freedom to remove RAM behind the virtual pages - and page it in again on-demand. Though I agree, we could make it easier for the OS.

[rmacnak-google]

I've tried this out, and unfortunately madvise(DONTNEED) has the wrong semantics for the VM to safely use it. If the memory isn't file-backed, it will be brought back in as zeros, but we need such memory to be left resident / ignored by madvise instead. (Parts of the clustered snapshot are not read-once because of external typed data pointing into it, the embedder reusing the snapshot for other isolates, or free reading malloc metadata.) The VM isn't always passed a snapshot that is file-backed, so any madvise(DONTNEED) needs to be done in embedders when the embedder can guarantee file-backed snapshots.

[mkustermann]

If we expose the right symbols, could we let the embedder(s) do the madvise calls for the region of RO which contains the serialized snapshot only (not the RO mapped heap objects)?