<html>
    <head>
      <base href="https://bugs.freedesktop.org/">
    </head>
    <body>
      <p>
        <div>
            <b><a class="bz_bug_link 
          bz_status_NEW "
   title="NEW - [GPU Passthrough]GPU (Polaris) not reinitialized with Linux VM (Reset bug)"
   href="https://bugs.freedesktop.org/show_bug.cgi?id=106111#c2">Comment # 2</a>
              on <a class="bz_bug_link 
          bz_status_NEW "
   title="NEW - [GPU Passthrough]GPU (Polaris) not reinitialized with Linux VM (Reset bug)"
   href="https://bugs.freedesktop.org/show_bug.cgi?id=106111">bug 106111</a>
              from <span class="vcard"><a class="email" href="mailto:alex.williamson@redhat.com" title="Alex Williamson <alex.williamson@redhat.com>"> <span class="fn">Alex Williamson</span></a>
</span></b>
        <pre>The IOMMU looks to be unhappy first:

[   40.201258] vfio_ecap_init: 0000:0a:00.0 hiding ecap 0x19@0x270
[   40.201271] vfio_ecap_init: 0000:0a:00.0 hiding ecap 0x1b@0x2d0
[   40.201279] vfio_ecap_init: 0000:0a:00.0 hiding ecap 0x1e@0x370
[  159.958402] AMD-Vi: Completion-Wait loop timed out
[  160.118777] AMD-Vi: Completion-Wait loop timed out
[  160.799864] AMD-Vi: Event logged [
[  160.799868] IOTLB_INV_TIMEOUT device=0a:00.0 address=0x000000043e8e8550]
[  160.799872] AMD-Vi: Event logged [
[  160.799874] IOTLB_INV_TIMEOUT device=0a:00.0 address=0x000000043e8e8570]
[  160.799876] AMD-Vi: Event logged [
[  160.799878] IOTLB_INV_TIMEOUT device=0a:00.0 address=0x000000043e8e8590]
[  161.801729] AMD-Vi: Event logged [
[  161.801732] IOTLB_INV_TIMEOUT device=0a:00.0 address=0x000000043e8e85e0]
[  180.096365] AMD-Vi: Completion-Wait loop timed out
[  180.256758] AMD-Vi: Completion-Wait loop timed out
[  180.417182] AMD-Vi: Completion-Wait loop timed out
[  180.577636] AMD-Vi: Completion-Wait loop timed out

Can you try a v4.17-rc1 kernel?  Specifically, these two updates:

6bd06f5a486c vfio/type1: Adopt fast IOTLB flush interface when unmap IOVAs
eb5ecd1a40e2 iommu/amd: Add support for fast IOTLB flushing

Something about AMD GPUs get unhappy if the IOMMU sends out too many
invalidations and the above two patches can reduce the number of those
invalidations by up to a factor of 512.

<a href="https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=6bd06f5a486c06023a618a86e8153b91d26f75f4">https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=6bd06f5a486c06023a618a86e8153b91d26f75f4</a>
<a href="https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=eb5ecd1a40e2098f805fb63cb07817ac48826e40">https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=eb5ecd1a40e2098f805fb63cb07817ac48826e40</a></pre>
        </div>
      </p>


      <hr>
      <span>You are receiving this mail because:</span>

      <ul>
          <li>You are the assignee for the bug.</li>
      </ul>
    </body>
</html>