On 3/14/2022 7:40 PM, Laurence Oberman wrote:
On Mon, 2022-03-14 at 11:55 -0400, David Jeffery wrote:
On Mon, Mar 14, 2022 at 10:52 AM Max Gurtovoy <mgurtovoy@xxxxxxxxxx>
wrote:
On 3/14/2022 3:57 PM, David Jeffery wrote:
On Sun, Mar 13, 2022 at 5:59 AM Max Gurtovoy <
mgurtovoy@xxxxxxxxxx> wrote:
Hi David,
thanks for the report.
Please check how we fixed that in NVMf in Sagi's commit:
nvmet-rdma: fix possible bogus dereference under heavy load
(commit:
8407879c4e0d77)
Maybe this can be done in isert and will solve this problem in
a simpler
way.
is it necessary to change max_cmd_sn ?
Hello,
Sure, there are alternative methods which could fix this
immediate
issue. e.g. We could make the command structs for scsi commands
get
allocated from a mempool. Is there a particular reason you don't
want
to do anything to modify max_cmd_sn behavior?
according to the description the command was parsed successful and
sent
to the initiator.
Yes.
Why do we need to change the window ? it's just a race of putting
the
context back to the pool.
And this race is rare.
Sure, it's going to be rare. Systems using isert targets with
infiniband are going to be naturally rare. It's part of why I left
the
max_cmd_sn behavior untouched for non-isert iscsit since they seem to
be fine as is. But it's easily and regularly triggered by some
systems
which use isert, so worth fixing.
I didn't do something like this as it seems to me to go against
the
intent of the design. It makes the iscsi window mostly
meaningless in
some conditions and complicates any allocation path since it now
must
gracefully and sanely handle an iscsi_cmd/isert_cmd not existing.
I
assume special commands like task-management, logouts, and pings
would
need a separate allocation source to keep from being dropped
under
memory load.
it won't be dropped. It would be allocated dynamically and freed
(instead of putting it back to the pool).
If it waits indefinitely for an allocation it ends up with a
variation
of the original problem under memory pressure. If it waits for
allocation on isert receive, then receive stalls under memory
pressure
and won't process the completions which would have released the other
iscsi_cmd structs just needing final acknowledgement.
If your system is under such memory pressure can you can't allocate few
bytes for isert response, the silent drop
of the command is your smallest problem. You need to keep the system
from crashing. And we do that in my suggestion.
David Jeffery
Folks this is a pending issue stopping a customer from making progress.
They run Oracle and very high workloads on EDR 100 so David fixed this
fosusing on the needs of the isert target changes etc.
Are you able to give us technical reasons why David's patch is not
suitable and why we he would have to start from scratch.
You shouldn't start from scratch. You did all the investigation and the
debugging already.
Coding a solution is the small part after you understand the root cause.
We literally spent weeks on this and built another special lab for
fully testing EDR 100.
This issue was pending in a BZ for some time and Mellnox had eyes on it
then but this latest suggestion was never put forward in that BZ to us.
Mellanox maintainers saw this issue few days before you sent it
upstream. I suggested sending it upstream and have a discussion here
since it has nothing to do with Mellanox adapters and Mellanox SW stack
MLNX_OFED.
Our job as maintainers and reviewers in the community is to see the big
picture and suggest solutions that not always same as posted in the
mailing list.
Sincerely
Laurence