On 01/05/2017 12:22 PM, Dave Hansen wrote:
On 01/04/2017 04:26 PM, Khalid Aziz wrote:
...
No, we do not have space to stuff PAGE_SIZE/64 version tags in swap pte.
There is enough space for just one tag per page. DaveM had suggested
doing this since the usual case is for a task to set one tag per page
even though MMU does not require it. I have implemented this as first
pass to start a discussion and get feedback on whether rest of the
swapping implementation and other changes look right, hence the patch is
"RFC". If this all looks good, I can expand swapping support in a
subsequent patch or iteration of this patch to allocate space in
mm_context_t possibly to store per cacheline tags. I am open to any
other ideas on storing this larger number of version tags.
FWIW, This is the kind of thing that would be really useful to point out
to reviewers instead of requiring them to ferret it out of the code. It
has huge implications for how applications use this feature.
Hi Dave,
Thanks for taking the time to review this. I appreciate your patience. I
will add more details.
As for where to store the tags... It's potentially a *lot* of data, so
I think it'll be a pain any way you do it.
If you, instead, can live with doing things on a PAGE_SIZE granularity
like pkeys does, you could just store it in the VMA and have the kernel
tag the data at the same time it zeroes the pages.
It is very tempting to restrict tags to PAGE_SIZE granularity since it
makes code noticeably simpler and that is indeed going to be the
majority of cases. Sooner or later somebody would want to use multiple
tags per page though. There can be 128 4-bit tags per 8K page which
requires 64 bytes of tag storage for each page. This can add up. What I
am considering doing is store the tag in swp pte if I find only one tag
on the page. A VMA can cover multiple pages and we have unused bits in
swp pte. It makes more sense to store the tags in swp pte. If I find
more than one tag on the page, I can allocate memory, attach it to a
data structure in mm_context_t and store the tags there. I will need to
use an rb tree or some other way to keep the data sorted to make it
quick to retrieve the tags for one of the millions of pages a task might
have. As I said, it gets complex trying to store tags per cacheline as
opposed to per page :)
--
Khalid
--
To unsubscribe from this list: send the line "unsubscribe linux-arch" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
