From mboxrd@z Thu Jan 1 00:00:00 1970 Content-Type: multipart/mixed; boundary="===============5626942657747401052==" MIME-Version: 1.0 From: Walker, Benjamin Subject: Re: [SPDK] Content for v18.10.x Date: Wed, 19 Dec 2018 21:36:45 +0000 Message-ID: <5eb9d04d4d2917be23389054cb2e8073dff88630.camel@intel.com> In-Reply-To: 435e9a40-f904-e517-d0ef-6dedcd7ea724@dev.mellanox.co.il List-ID: To: spdk@lists.01.org --===============5626942657747401052== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On Wed, 2018-12-19 at 20:11 +0200, Sasha Kotchubievsky wrote: > Do you see the same pattern in both branches in "master" and in "18.10.1"? Yes - as far as we've been able to tell it occurs in the same way on both m= aster and 18.10.1. > > = > > This is still an open issue, but Seth did figure out why we don't see t= he > > problem on master - the test changed. The plan right now is to move for= ward > > with > > 18.10.1 without addressing this problem. 18.10.1 is far better than 18.= 10 in > > this area already, and because this problem likely exists on master in = the > > same > > form (and may not even ultimately be an SPDK problem), we'd rather not = wait > > to > > get 18.10.1 out any longer. > = > Dose it mean, the patch fixing "ref count" will be merged into 18.10.1? Yes, it's merged now. > > = > > Any assistance with investigating the root cause of the issue on the ma= ster > > branch would be greatly appreciated. It replicates using the 18.10 vers= ion > > of > > the test/nvmf/lvol/nvmf_lvol.sh test, but not the version on master. The > > problem > > appears to be that the connection just drops, and both the initiator an= d the > > target reactor to that correctly. But why the connection is dropping (i= t's > > loopback) is the unknown. > = > Which SW stack is used in the testing: SoftROCE/MOFED/OFED? OS? > = > As I understand, for reproduction, I run test/nvmf/lvol/nvmf_lvol.sh ( = > from 18.10 ) on source coming from "master". Is that correct? Correct, run the 18.10 version of the test against "master" and it will eventually reproduce. It does not hit every time and we so far haven't gott= en it to happen on a development system, but we haven't run it very many times ju= st yet. We're testing on Fedora 28 with whatever RDMA stack it comes with nati= vely using a Mellanox RNIC in loopback. --===============5626942657747401052==--