On 02/08/2012 01:28 PM, Dave Hansen wrote:

On 02/08/2012 09:53 AM, Nitin Gupta wrote:

quoted

vmap() is not just slower but also does memory allocations at various
places. Under memory pressure, this may cause failure in reading a
stored object just because we failed to map it. Also, it allocates VA
region each time its called which is a real big waste when we can simply
pre-allocate 2 * PAGE_SIZE'ed VA regions (per-cpu).

Yeah, vmap() is a bit heavy-handed.  I'm just loathe to go mucking
around in the low-level pagetables too much.  Just seems like there'll
be a ton of pitfalls, like arch-specific TLB flushing, and it _seems_
like one of the existing kernel mechanisms should work.

I guess if you've exhaustively explored all of the existing kernel
mapping mechanisms and found none of them to work, and none of them to
be in any way suitably adaptable to your use, you should go ahead and
roll your own.  I guess you do at least use alloc_vm_area().  What made
map_vm_area() unsuitable for your needs?  If you're remapping, you
should at least be guaranteed not to have to allocate pte pages.

map_vm_area() needs 'struct vm_struct' parameter but for mapping kernel
allocated pages within kernel, what should we pass here?  I think we can
instead use map_kernel_range_noflush() -- surprisingly
unmap_kernel_range_noflush() is exported but this one is not.

Nitin

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>