Thanks Gal, I think I found the problem for the issue below, attached a patch that should be applied on top of the series. Please let me know if this fixed the issues you are seeing. In qedr we work with only single pages, and this issue will only occur with multiple pages. Thanks, Michal > -----Original Message----- > From: Gal Pressman > Sent: Wednesday, August 21, 2019 1:15 PM > To: Michal Kalderon > Cc: jgg@ziepe.ca; dledford@redhat.com; Ariel Elior ; > bmt@zurich.ibm.com; sleybo@amazon.com; leon@kernel.org; linux- > rdma@vger.kernel.org > Subject: Re: [PATCH v7 rdma-next 0/7] RDMA/qedr: Use the doorbell > overflow recovery mechanism for RDMA > > On 21/08/2019 11:03, Michal Kalderon wrote: > > Hi Gal, > > > > Thanks for the quick testing and feedback! > > > > Can you share some more information on the scenario you're running ? > > It happens on most of our automated tests. > I reproduce it manually by running ib_send_{bw,lat} over SRD. > > > Does this happen each time or intermittently ? > > Happens on most of the runs. > > > Can you send me your .config ? > > Attached. > > > are you running agains rdma-next tree ? > > Yes, commit 77905379e9b2 ("RDMA/hns: Remove unuseful member") with > this series applied on top. > > > Can you reproduce with enabling ib_core module dynamic debug on ? > > Attached a log of ib_send_bw running with ib_core and ib_uverbs dynamic > debug enabled. > > Let me know if there's anything else I can do.