| Message ID | 20210407113444.150961-1-gi-oh.kim@ionos.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | New multipath policy for RTRS client | expand |
On Wed, Apr 07, 2021 at 01:34:40PM +0200, Gioh Kim wrote: > This patch set introduces new multipath policy 'min-latency'. > The latency is a time calculated by the heart-beat messages. Whenever > the client sends heart-beat message, it checks the time gap between > sending the heart-beat message and receiving the ACK. So this value > can be changed regularly. > If client has multi-path, it can send IO via a path having the least > latency. > > V3->V2: use sysfs_emit instead of scnprintf > V2->V1: use sysfs_emit instead of sprintf > > Gioh Kim (3): > RDMA/rtrs-clt: Add a minimum latency multipath policy > RDMA/rtrs-clt: new sysfs attribute to print the latency of each path > Documentation/ABI/rtrs-clt: Add descriptions for min-latency policy Applied to for-next, thanks Please be mindful about the subjects, rdma subject start with a capital letter after the tag > Md Haris Iqbal (1): > RDMA/rtrs-clt: Check state of the rtrs_clt_sess before reading its > stats This one was replaced by that other patch Jason
On Wed, Apr 14, 2021 at 12:47 AM Jason Gunthorpe <jgg@nvidia.com> wrote: > > On Wed, Apr 07, 2021 at 01:34:40PM +0200, Gioh Kim wrote: > > This patch set introduces new multipath policy 'min-latency'. > > The latency is a time calculated by the heart-beat messages. Whenever > > the client sends heart-beat message, it checks the time gap between > > sending the heart-beat message and receiving the ACK. So this value > > can be changed regularly. > > If client has multi-path, it can send IO via a path having the least > > latency. > > > > V3->V2: use sysfs_emit instead of scnprintf > > V2->V1: use sysfs_emit instead of sprintf > > > > Gioh Kim (3): > > RDMA/rtrs-clt: Add a minimum latency multipath policy > > RDMA/rtrs-clt: new sysfs attribute to print the latency of each path > > Documentation/ABI/rtrs-clt: Add descriptions for min-latency policy > > Applied to for-next, thanks > > Please be mindful about the subjects, rdma subject start with a > capital letter after the tag > > > Md Haris Iqbal (1): > > RDMA/rtrs-clt: Check state of the rtrs_clt_sess before reading its > > stats > > This one was replaced by that other patch Hi Jason, "RDMA/rtrs-clt: Check state of the rtrs_clt_sess before reading its stats" This patch is still necessary. Please apply that patch. As I wrote in the description of patch "RDMA/rtrs-clt: destroy sysfs after removing session from active list", each function still should check the session status because closing or error recovery paths can change the status. For example, the client sends the heart-beat and does not get the response, it changes the session status and stops IO processing. It is ok if the status is changed after checking status because the error recovery path does not free memory and only tries to reconnection. And closing the session changes the session status and flush all IO, and then free memory. Thank you for the review.
On Wed, Apr 14, 2021 at 06:59:59AM +0200, Gioh Kim wrote: > On Wed, Apr 14, 2021 at 12:47 AM Jason Gunthorpe <jgg@nvidia.com> wrote: > > > > On Wed, Apr 07, 2021 at 01:34:40PM +0200, Gioh Kim wrote: > > > This patch set introduces new multipath policy 'min-latency'. > > > The latency is a time calculated by the heart-beat messages. Whenever > > > the client sends heart-beat message, it checks the time gap between > > > sending the heart-beat message and receiving the ACK. So this value > > > can be changed regularly. > > > If client has multi-path, it can send IO via a path having the least > > > latency. > > > > > > V3->V2: use sysfs_emit instead of scnprintf > > > V2->V1: use sysfs_emit instead of sprintf > > > > > > Gioh Kim (3): > > > RDMA/rtrs-clt: Add a minimum latency multipath policy > > > RDMA/rtrs-clt: new sysfs attribute to print the latency of each path > > > Documentation/ABI/rtrs-clt: Add descriptions for min-latency policy > > > > Applied to for-next, thanks > > > > Please be mindful about the subjects, rdma subject start with a > > capital letter after the tag > > > > > Md Haris Iqbal (1): > > > RDMA/rtrs-clt: Check state of the rtrs_clt_sess before reading its > > > stats > > > > This one was replaced by that other patch > > Hi Jason, > > "RDMA/rtrs-clt: Check state of the rtrs_clt_sess before reading its stats" > This patch is still necessary. Please apply that patch. > > As I wrote in the description of patch > "RDMA/rtrs-clt: destroy sysfs after removing session from active list", > each function still should check the session status because closing > or error recovery paths can change the status. > > For example, the client sends the heart-beat and does not get the > response, it changes the session status and stops IO processing. > It is ok if the status is changed after checking status because > the error recovery path does not free memory and only tries to > reconnection. > > And closing the session changes the session status and flush all IO, > and then free memory. You need to resend it with out the oops message, a patch like this cannot be correctly fixing an oops like it presents. Confirm you do not have the oops now that the other patch is merged. Please explain carefully what it is trying to do Jason
On Tue, Apr 27, 2021 at 9:00 PM Jason Gunthorpe <jgg@ziepe.ca> wrote: > > On Wed, Apr 14, 2021 at 06:59:59AM +0200, Gioh Kim wrote: > > On Wed, Apr 14, 2021 at 12:47 AM Jason Gunthorpe <jgg@nvidia.com> wrote: > > > > > > On Wed, Apr 07, 2021 at 01:34:40PM +0200, Gioh Kim wrote: > > > > This patch set introduces new multipath policy 'min-latency'. > > > > The latency is a time calculated by the heart-beat messages. Whenever > > > > the client sends heart-beat message, it checks the time gap between > > > > sending the heart-beat message and receiving the ACK. So this value > > > > can be changed regularly. > > > > If client has multi-path, it can send IO via a path having the least > > > > latency. > > > > > > > > V3->V2: use sysfs_emit instead of scnprintf > > > > V2->V1: use sysfs_emit instead of sprintf > > > > > > > > Gioh Kim (3): > > > > RDMA/rtrs-clt: Add a minimum latency multipath policy > > > > RDMA/rtrs-clt: new sysfs attribute to print the latency of each path > > > > Documentation/ABI/rtrs-clt: Add descriptions for min-latency policy > > > > > > Applied to for-next, thanks > > > > > > Please be mindful about the subjects, rdma subject start with a > > > capital letter after the tag > > > > > > > Md Haris Iqbal (1): > > > > RDMA/rtrs-clt: Check state of the rtrs_clt_sess before reading its > > > > stats > > > > > > This one was replaced by that other patch > > > > Hi Jason, > > > > "RDMA/rtrs-clt: Check state of the rtrs_clt_sess before reading its stats" > > This patch is still necessary. Please apply that patch. > > > > As I wrote in the description of patch > > "RDMA/rtrs-clt: destroy sysfs after removing session from active list", > > each function still should check the session status because closing > > or error recovery paths can change the status. > > > > For example, the client sends the heart-beat and does not get the > > response, it changes the session status and stops IO processing. > > It is ok if the status is changed after checking status because > > the error recovery path does not free memory and only tries to > > reconnection. > > > > And closing the session changes the session status and flush all IO, > > and then free memory. > > You need to resend it with out the oops message, a patch like this > cannot be correctly fixing an oops like it presents. Confirm you do > not have the oops now that the other patch is merged. Yes, that oops is prevented by another patch. I will change the commit message and send it to you. > > Please explain carefully what it is trying to do Ok, I will. > > Jason
This patch set introduces new multipath policy 'min-latency'. The latency is a time calculated by the heart-beat messages. Whenever the client sends heart-beat message, it checks the time gap between sending the heart-beat message and receiving the ACK. So this value can be changed regularly. If client has multi-path, it can send IO via a path having the least latency. V3->V2: use sysfs_emit instead of scnprintf V2->V1: use sysfs_emit instead of sprintf Gioh Kim (3): RDMA/rtrs-clt: Add a minimum latency multipath policy RDMA/rtrs-clt: new sysfs attribute to print the latency of each path Documentation/ABI/rtrs-clt: Add descriptions for min-latency policy Md Haris Iqbal (1): RDMA/rtrs-clt: Check state of the rtrs_clt_sess before reading its stats .../ABI/testing/sysfs-class-rtrs-client | 12 ++++ drivers/infiniband/ulp/rtrs/rtrs-clt-sysfs.c | 35 ++++++++++- drivers/infiniband/ulp/rtrs/rtrs-clt.c | 60 ++++++++++++++++++- drivers/infiniband/ulp/rtrs/rtrs-clt.h | 1 + drivers/infiniband/ulp/rtrs/rtrs-pri.h | 2 + drivers/infiniband/ulp/rtrs/rtrs.c | 3 + 6 files changed, 109 insertions(+), 4 deletions(-)