Some comments on Rohit's large page patch (looking at Andrew's version).
I agree with keeping the actual large page handling separate, per arch.
I agree that it's sensible to focus upon _large_ pages (e.g. 4MB) here;
grouping several pages together as superpages (e.g. 64KB), for non-x86
TLB or other reasons, better handled automatically in another project.
I agree that large pages be kept right away from swap (VM_RESERVE).
But I disagree with new interfaces distinct from known mmap/shm/tmpfs
usage, I think they will cause rather than save trouble. It's using
do_mmap_pgoff anyway, why not mmap itself? Much prefer MAP_LARGEPAGE,
SHM_LARGEPAGE - or might /dev/ZERO and TMPFS help, each on large pages?
munmap, mprotect, mremap patches are deficient: they just check whether
the first vma is VM_LARGEPAGE, but munmap and mprotect (and mremap's
do_munmap in MREMAP_MAYMOVE|MREMAP_FIXED case) may span several vmas.
So, must decide what to do when a VM_LARGEPAGE falls within do_munmap
span: pre-scan would waste time, we rely on unmap_region to unmap
at least length specified by user, we're not interested in splitting
VM_LARGEPAGE areas, so I suggest when a VM_LARGEPAGE area falls partly
or wholly within do_munmap span, it be wholly unmapped. In which case,
no need for sys_free_large_pages and sys_unshare_large_pages.
sys_get_large_pages, if retained as a separate syscall,
would be easier to understand if named sys_mmap_large_pages?
sys_share_large_pages: I'm having a lot of difficulty with this one,
and its set_lp_shm_seg. Share? but it says MAP_PRIVATE (whereas
sys_get_large_pages forces MAP_SHARED). Key? we got that from a
prior shmget? so already there's a tmpfs inode for this, and now
we allocate some other inode? No, I think it would be better off
integrated a little more within tmpfs (perhaps no SHM_LARGEPAGE
at all, just ordinary files in a TMPFS? Rohit mentioned wanting
ability to execute, straightforward from TMPFS file).
change_large_page_mem_size: wouldn't it be better as
set_large_page_mem_size, instead of by increments/decrements?
Whitespace offences, would benefit from a pass through Lindent.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to firstname.lastname@example.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Wed Aug 07 2002 - 22:00:33 EST