On 5/17/20 3:51 AM, Max Gurtovoy wrote:
On 5/15/2020 9:59 PM, Sagi Grimberg wrote:
It's not only the fmr_pools. The FMR API operates on a page granularity,
so a sub-page registration, or a non-page-aligned one, ends up exposing
data outside the buffer. This is done silently, and is a significant
vulnerability for most upper layers.
You're right Tom, I forgot about that...
I guess I'd vote to remove it altogether then...
Santosh,
I don't really understand the mechanism you use from user/kernel
regarding the registration.
But I saw you use wait_event and wait for completing the
FRWR/INVALIDATION. For sure this is not optimal.
But can't really say what is the bottleneck in your case.
Can you share the numbers you get for FMR/FRWR benchmarks ?
The upstream code for FRWR is behind the optimal code we have
been using. The invalidation path is removed from datapath
with proxy QP within the PD.
Anyway, please bounce your patches for me to review to make sure
nothing breaks. From various responses, it seems to be aligned to
move ahead with the patch-set, so lets go with it.
Regards,
Santosh
