Re: [PATCH v1 06/11] mm: support GUP-triggered unsharing via FAULT_FLAG_UNSHARE (!hugetlb)

From
David Hildenbrand <hidden>
Date
2021-12-17 20:42:47
Also in
linux-kselftest, linux-mm, lkml
Reply to
Linus Torvalds, 2021-12-17
On 17.12.21 21:36, Linus Torvalds wrote:
On Fri, Dec 17, 2021 at 12:18 PM David Hildenbrand [off-list ref] wrote:
quoted
On 17.12.21 20:22, Linus Torvalds wrote:
quoted
On Fri, Dec 17, 2021 at 11:04 AM Linus Torvalds
[off-list ref] wrote:
quoted
 - get a "readonly" copy of a local private page using FAULT_FLAG_UNSHARE.

   This just increments the page count, because mapcount == 1.

 - fork()

 - unmap in the original

 - child now has "mapcount == 1" on a page again, but refcount is
elevated, and child HAS TO COW before writing.
Hi Linus,

This is just GUP before fork(), which is in general
problematic/incompatible with sharing.
Note that my example was not meant to be an example of a problem per
se, but purely as an example of how meaningless 'mapcount' is, and how
'mapcount==1' isn't really a very meaningful test.

So it wasn't mean to show "look, GUP before fork is problematic".  We
have that problem already solved at least for regular pages.

It was purely meant to show how "mapcount==1" isn't a meaningful thing
to test, and my worry about how you're adding that nonsensical test to
the new code.
quoted
Let's just take a look at what refcount does *wrong*. Let's use an
adjusted version of your example above, because it's a perfect fit:

1. mem = mmap(pagesize, MAP_PRIVATE)
-> refcount == 1

2. memset(mem, 0, pagesize); /* Page is mapped R/W */

3. fork() /* Page gets mapped R/O */
-> refcount > 1

4. child quits
-> refcount == 1

5. Take a R/O pin (RDMA, VFIO, ...)
-> refcount > 1

6. memset(mem, 0xff, pagesize);
-> Write fault -> COW
I do not believe this is actually a bug.
It's debatable if it's a BUG or not (I think it is one). It's for sure
inconsistent.
You asked for a R/O pin, and you got one.

Then somebody else modified that page, and you got exactly what you
asked for - a COW event. The original R/O pin has the original page
that it asked for, and can read it just fine.
Where in the code did I ask for a COW event? I asked for a R/O pin, not
any kind of memory protection.

-- 
Thanks,

David / dhildenb
Thread (137 messages) 137 messages, 11 authors, last reply 2022-01-04
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help