On Jan 15, 2019, at 9:00 PM, Linus Torvalds [off-list ref] wrote:

quoted

On Wed, Jan 16, 2019 at 12:42 PM Josh Snyder [off-list ref] wrote:

For Netflix, losing accurate information from the mincore syscall would
lengthen database cluster maintenance operations from days to months.  We
rely on cross-process mincore to migrate the contents of a page cache from
machine to machine, and across reboots.

Ok, this is the kind of feedback we need, and means I guess we can't
just use the mapping existence for mincore.

The two other ways that we considered were:

(a) owner of the file gets to know cache information for that file.

(b) having the fd opened *writably* gets you cache residency information.

Sadly, taking a look at happycache, you open the file read-only, so
(b) doesn't work.

Judging just from the source code, I can't tell how the user ownership
works. Any input on that?

And if you're not the owner of the file, do you have another
suggestion for that "Yes, I have the right to see what's in-core for
this file". Because the problem is literally that if it's some random
read-only system file, the kernel shouldn't leak access patterns to
it..

Something like CAP_DAC_READ_SEARCH might not be crazy.