5/4/2019 11:22 PM, Leon Romanovsky wrote:
On Thu, May 02, 2019 at 10:59:58AM -0700, Santosh Shilimkar wrote:
On 5/1/2019 11:18 PM, Leon Romanovsky wrote:
On Wed, May 01, 2019 at 10:54:50AM -0700, Santosh Shilimkar wrote:
On 5/1/2019 12:45 AM, Leon Romanovsky wrote:
On Mon, Apr 29, 2019 at 04:37:20PM -0700, Santosh Shilimkar wrote:
RDS doesn't support RDMA on memory apertures that require On Demand
Paging (ODP), such as FS DAX memory. A sysctl is added to indicate
whether RDMA requiring ODP is supported.
Reviewed-by: H??kon Bugge <haakon.bu...@oracle.com>
Reviewed-tested-by: Zhu Yanjun <yanjun....@oracle.com>
Signed-off-by: Hans Westgaard Ry <hans.westgaard...@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilim...@oracle.com>
---
net/rds/ib.h | 1 +
net/rds/ib_sysctl.c | 8 ++++++++
2 files changed, 9 insertions(+)
This sysctl is not needed at all
Its needed for application to check the support of the ODP support
feature which in progress. Failing the RDS_GET_MR was just one path
and we also support inline MR registration along with message request.
Basically application runs on different kernel versions and to be
portable, it will check if underneath RDS support ODP and then only
use RDMA. If not it will fallback to buffer copy mode. Hope
it clarifies.
Using ODP sysctl to determine if to use RDMA or not, looks like very
problematic approach. How old applications will work in such case
without knowledge of such sysctl?
How new applications will distinguish between ODP is not supported, but
RDMA works?
Actually this is not ODP sysctl but really whether RDS supports
RDMA on fs_dax memory or not. I had different name for sysctl but
in internal review it got changed.
Ignoring the name of the sysctl, here is the application logic.
- If fs_dax sysctl path doesn't exist, no RDMA on FS DAX memory(this
will cover all the older kernels, which doesn't have this patch)
- If fs_dax sysctl path exist and its value is 0, no RDMA on FS
DAX. This will cover kernels which this patch but don't have
actual support for ODP based registration.
- If fs_dax sysctl path exist and its value is 1, RDMA can be
issued on FS DAX memory. This sysctl will be updated to value 1
once the support gets added.
Hope it clarifies better now.
Santosh,
Thanks for explanation, I have one more question,
If I'm author of hostile application and write code to disregard that
new sysctl, will any of combinations of kernel/application cause to
kernel panic? If not, we don't really need to expose this information,
if yes, this sysctl is not enough.
It Won't panic. Thats why the other patch also makes the call fail when
tried to register FS DAX memory with RDS.
