<html>
  <head>
    <meta content="text/html; charset=windows-1252"
      http-equiv="Content-Type">
  </head>
  <body bgcolor="#FFFFFF" text="#000000">
    <div class="moz-cite-prefix">Hi all,<br>
      <br>
      I've reset the cluster to ofed 3.12-1 and rerun rsockets with the
      updated librdmacm-1.0.21 installed.  Everything passed
      heterogeneously.  <br>
      <br>
      I'll now do the same for nfsordma with the hacked together
      connectathon plan and move this arbitration request forward.  <br>
      <br>
      Please let me know if you have any questions<br>
      <br>
      Thanks,<br>
      Dave<br>
      <br>
      <br>
      <br>
      <br>
      <br>
      On 6/5/15 4:36 PM, Dave Wyman wrote:<br>
    </div>
    <blockquote
      cite="mid:400F6B2D-1C11-45B0-A0F1-1079D8D2A3A9@iol.unh.edu"
      type="cite">
      <meta http-equiv="Content-Type" content="text/html;
        charset=windows-1252">
      <base href="x-msg://2999/">Hello,
      <div><br>
      </div>
      <div>I've done a test run with the updated librdmacm-1.0.21
        installed on hosts with QLE7340 and QLE7342 against each other
        and the set of Mellanox devices included in the January Logo
        Event with each host acting as both server and client.  The
        installed OFED was 3.18-rc2.  All passed per the test plan
        (rstream -T [sabn] -S all).  This is great news for the upcoming
        logo event given the fix has been submitted for 3.18.  I plan to
        reset the hosts to 3.12-1 and again update the Intel hosts with
        librdmacm-1.0.21 and test again to verify this against the
        pending arbitration request.  </div>
      <div><br>
      </div>
      <div>Thanks,</div>
      <div>Dave</div>
      <div><br>
      </div>
      <div><br>
      </div>
      <div><br>
      </div>
      <div><br>
      </div>
      <div><br>
        <div>
          <div>On Jun 3, 2015, at 3:10 PM, "Calciano, Jess" <<a
              moz-do-not-send="true"
              href="mailto:jess.calciano@intel.com">jess.calciano@intel.com</a>>
            wrote:</div>
          <br class="Apple-interchange-newline">
          <blockquote type="cite">
            <div link="blue" vlink="purple" style="font-family:
              Helvetica; font-size: medium; font-style: normal;
              font-variant: normal; font-weight: normal; letter-spacing:
              normal; line-height: normal; orphans: 2; text-align:
              -webkit-auto; text-indent: 0px; text-transform: none;
              white-space: normal; widows: 2; word-spacing: 0px;
              -webkit-text-size-adjust: auto; -webkit-text-stroke-width:
              0px; " lang="EN-US">
              <div class="WordSection1" style="page: WordSection1; ">
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Hello,<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Since the
                    original arbitration request was submitted, there’s
                    been some further discussion about the RSockets
                    failure. With the fix for librdmacm described in the
                    original request, rstream ran successfully for most
                    message sizes, but still hung with -S 1024.<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Additional
                    investigation traced the new problem to an
                    incompatibility between the qib driver and the
                    ibv_create_qp() function. A workaround (described
                    below) is available for the current OFED version and
                    a permanent fix to librdmacm will be included in the
                    next OFED 3.18 release.<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Details:<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt 0.5in; font-size:
                  12pt; font-family: 'Times New Roman', serif; ">The
                  ultimate issue is still related to the qib driver
                  being non-compliant with the ibv_create_qp()
                  definition:<br>
                  <br>
                  <span class="apple-tab-span">           <span
                      class="Apple-converted-space"> </span></span>The
                  function ibv_create_qp() will update the
                  qp_init_attr->cap<br>
                  <span class="apple-tab-span">           <span
                      class="Apple-converted-space"> </span></span>struct
                  with the actual QP values of the QP that was created;<br>
                  <span class="apple-tab-span">           <span
                      class="Apple-converted-space"> </span></span>***
                  the values will be greater than or equal to the values
                  requested. ***<br>
                  <span class="apple-tab-span">           <span
                      class="Apple-converted-space"> </span></span><br>
                  Specifically, the qib driver will return an inline
                  size that is smaller than that requested.  Rsockets
                  has code to trap for this, but the rsockets code looks
                  like this:<br>
                  <br>
                  <span class="apple-tab-span">           <span
                      class="Apple-converted-space"> </span></span>inline_size
                  = SOME_DEFAULT_LIKE_64<br>
                  <span class="apple-tab-span">           <span
                      class="Apple-converted-space"> </span></span>rs_init_bufs(...);<br>
                  <span class="apple-tab-span">           <span
                      class="Apple-converted-space"> </span></span>...<br>
                  <span class="apple-tab-span">           <span
                      class="Apple-converted-space"> </span></span>rs_create_qp(...);<br>
                  <span class="apple-tab-span">           <span
                      class="Apple-converted-space"> </span></span>inline_size
                  = qp_cap->max_inline_size;<br>
                  <br>
                  The issue is that rs_init_bufs(), which allocates the
                  buffers and registers the memory, uses the default
                  inline size.  The net result is that rsockets ends up
                  referencing memory that is outside of the registered
                  memory region when sending credit updates.  The lost
                  credit update is causing the hang that you see.<br>
                  <br>
                  A quick check shows that I can move the rs_init_bufs()
                  call after the qp has been created and have the test
                  work.  You should also be able to override the
                  inline_size by writing the value 0 into a config file.
                   This will set the inline_size to 0 as the default.
                   To do this, you need to write a 0 into
                  /etc/rdma/rsocket/inline_default.  (The actual path
                  will depend on your configuration, so it could be
                  under /usr/etc/rdma/... for example.)  Updating the
                  config file should work with the current version.<br>
                  <br>
                  I will provide an update to the librdmacm to handle
                  this.  That update will find its way into the 3.18
                  release.<br>
                  <br>
                  <span style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "><o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Thanks,<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Jess Calciano<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><a
                    moz-do-not-send="true" name="_MailEndCompose"><span
                      style="font-size: 11pt; font-family: Calibri,
                      sans-serif; color: rgb(31, 73, 125); "> </span></a></div>
                <div>
                  <div style="border-style: solid none none;
                    border-top-width: 1pt; border-top-color: rgb(225,
                    225, 225); padding: 3pt 0in 0in; ">
                    <div style="margin: 0in 0in 0.0001pt; font-size:
                      12pt; font-family: 'Times New Roman', serif; "><b><span
                          style="font-size: 11pt; font-family: Calibri,
                          sans-serif; ">From:</span></b><span
                        style="font-size: 11pt; font-family: Calibri,
                        sans-serif; "><span
                          class="Apple-converted-space"> </span>Calciano,
                        Jess<span class="Apple-converted-space"> </span><br>
                        <b>Sent:</b><span class="Apple-converted-space"> </span>Wednesday,
                        April 08, 2015 2:39 PM<br>
                        <b>To:</b><span class="Apple-converted-space"> </span><a
                          moz-do-not-send="true"
                          href="mailto:iwg-arbitration-committee@openfabrics.org"
                          style="color: purple; text-decoration:
                          underline; ">iwg-arbitration-committee@openfabrics.org</a><br>
                        <b>Cc:</b><span class="Apple-converted-space"> </span>OFA
                        Lab Mailing List; Dave Wyman; Rupert Dance <<a
                          moz-do-not-send="true"
                          href="mailto:rsdance@soft-forge.com"
                          style="color: purple; text-decoration:
                          underline; ">rsdance@soft-forge.com</a>> (<a
                          moz-do-not-send="true"
                          href="mailto:rsdance@soft-forge.com"
                          style="color: purple; text-decoration:
                          underline; ">rsdance@soft-forge.com</a>);
                        Cole, Cliff; Mascarenhas, Edward; Sharma, Karun;
                        Thete, Swapna; Hefty, Sean; Yan, Philip W;
                        Flores, Jose F<br>
                        <b>Subject:</b><span
                          class="Apple-converted-space"> </span>Arbitration
                        request for Intel QLE7340 & QLE7342 HCAs
                        (Jan 2015 OFA Interop Logo Event)<o:p></o:p></span></div>
                  </div>
                </div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><o:p> </o:p></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Hello,<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Intel would
                    like to file an arbitration request for the January
                    2015 OFA Interop Logo Event results for the Intel
                    QLE7340 and QLE7342 HCAs.<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">The provided
                    report (attached for reference) shows two failing
                    tests:<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt 0.5in; font-size:
                  12pt; font-family: 'Times New Roman', serif;
                  text-indent: -0.25in; "><span style="font-size: 11pt;
                    font-family: Calibri, sans-serif; color: rgb(31, 73,
                    125); "><span>1)<span style="font-style: normal;
                        font-variant: normal; font-weight: normal;
                        font-size: 7pt; line-height: normal;
                        font-family: 'Times New Roman'; ">     <span
                          class="Apple-converted-space"> </span></span></span></span><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">TI NFS over
                    RDMA<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt 0.5in; font-size:
                  12pt; font-family: 'Times New Roman', serif;
                  text-indent: -0.25in; "><span style="font-size: 11pt;
                    font-family: Calibri, sans-serif; color: rgb(31, 73,
                    125); "><span>2)<span style="font-style: normal;
                        font-variant: normal; font-weight: normal;
                        font-size: 7pt; line-height: normal;
                        font-family: 'Times New Roman'; ">     <span
                          class="Apple-converted-space"> </span></span></span></span><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">TI RSockets<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">The Intel
                    team has investigated these results and determined
                    that the failures are due to bugs in non-Intel
                    components.<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">NFSoRDMA:<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt 0.25in; font-size:
                  12pt; font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">The failure
                    is due to a known Connectathon issue, documented
                    here:<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt 0.25in; font-size:
                  12pt; font-family: 'Times New Roman', serif; "><a
                    moz-do-not-send="true"
                    href="http://www.spinics.net/lists/linux-nfs/msg16460.html"
                    style="color: purple; text-decoration: underline; "><span
                      style="font-size: 11pt; font-family: Calibri,
                      sans-serif; ">http://www.spinics.net/lists/linux-nfs/msg16460.html</span></a><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; "><o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">RSockets:<o:p></o:p></span></div>
                <p class="MsoNormal" style="margin: 0in 0in 12pt 0.25in;
                  font-size: 12pt; font-family: 'Times New Roman',
                  serif; "><span style="font-size: 11pt; font-family:
                    Calibri, sans-serif; color: rgb(0, 32, 96); ">The
                    issue is that ibv_modify_qp() is failing.  The
                    problem is that an incorrect bit is set in the
                    qp_attr_mask, which is returned from the kernel.
                     With Intel, bit 21 of the qp_attr_mask is set.
                     This is not the case for a Mellanox HCA.<br>
                    <br>
                    Bit 21 is not defined for userspace.  However, it
                    was defined in the kernel as IB_QP_SMAC.<br>
                    <br>
                    If the librdmacm is modified to mask out this bit,
                    the call succeeds and rstream runs successfully.</span><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "><o:p></o:p></span></p>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Please let me
                    know if the arbitration committee needs any
                    additional information on the analysis.<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); "> </span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Thanks,<o:p></o:p></span></div>
                <div style="margin: 0in 0in 0.0001pt; font-size: 12pt;
                  font-family: 'Times New Roman', serif; "><span
                    style="font-size: 11pt; font-family: Calibri,
                    sans-serif; color: rgb(31, 73, 125); ">Jess Calciano<o:p></o:p></span></div>
              </div>
            </div>
          </blockquote>
        </div>
        <br>
      </div>
    </blockquote>
    <br>
  </body>
</html>