<html xmlns:v="urn:schemas-microsoft-com:vml" xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40"> <head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> <meta name="Generator" content="Microsoft Word 15 (filtered medium)"> <style></style> </head> <body lang="EN-US" link="#467886" vlink="#96607D" style="word-wrap:break-word"> <div class="WordSection1"> Hi all,<o:p></o:p> Sorry for my late reply. I don't know if you still remember this thread, let me give a quick summary:<o:p></o:p> <ol start="1" type="1"> <li style="mso-list:l1 level1 lfo2">We want to implement the dGPU prime feature in guest VM. But we encountered this issue: virtio-gpu doesn’t have ->get_sg_table implemented which is required by drm_gem_map_attach(). This is modified by: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”).<o:p></o:p></li><li style="mso-list:l1 level1 lfo2">To fix this, I override the function virtgpu_gem_device_attach() to not call drm_gem_map_attach() for vram object so drm_gem_map_attach() will not return -ENOSYS for not having ->get_sg_table.<o:p></o:p></li><li style="mso-list:l1 level1 lfo2">Then you think this is incorrect and drm_gem_map_attach() requires get_sg_table to be implemented is intentional. I should either implement ->attach or ->get_sg_table for virtio-gpu.<o:p></o:p></li><li style="mso-list:l1 level1 lfo2">As discussed, I implemented ->attach for virtio-gpu, but you suggested that I should check peer2peer flag first.<o:p></o:p></li><li style="mso-list:l1 level1 lfo2">Now I have the implementation to get p2p_distance and check the p2p flag already, but I found that Rob Clark merged a patch to fix above patch: 207395da5a97 (“drm/prime: reject DMA-BUF attach when get_sg_table is missing”)<o:p></o:p></li><ol start="1" type="a"> <li style="mso-list:l1 level2 lfo2">Rob’s patch: <a href="https://patchwork.freedesktop.org/patch/584318/"> https://patchwork.freedesktop.org/patch/584318/</a><o:p></o:p></li></ol> <li style="mso-list:l1 level1 lfo2">With Rob’s patch, ->get_sg_table isn’t required for virtio-gpu anymore and it seems p2p flag also doesn’t need to be checked anymore.<o:p></o:p></li></ol> So I want to rediscuss if we still need to do p2p checking now? <o:p></o:p> If so, I will send out my implementation soon.<o:p></o:p> Best regards,<o:p></o:p> Julia<o:p></o:p> <o:p> </o:p> <div> On 2024/1/31 22:32, Christian König wrote:<o:p></o:p> </div> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> Am 31.01.24 um 11:20 schrieb Zhang, Julia: <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>On 2024/1/30 22:23, Christian König wrote:<o:p></o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>Am 30.01.24 um 12:16 schrieb Daniel Vetter:<o:p></o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>On Tue, Jan 30, 2024 at 12:10:31PM +0100, Daniel Vetter wrote:<o:p></o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> [SNIP] <o:p></o:p> </blockquote> </blockquote> </blockquote> <pre>Hi Sima, Christian,<o:p></o:p></pre> <pre><o:p> </o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>Yeah, that is really just speculative. All importers need to set the peer2peer flag just in case.<o:p></o:p></pre> </blockquote> <pre>I see, I will modify this.<o:p></o:p></pre> <pre><o:p> </o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>What happens under the hood is that IOMMU redirects the "VRAM" memory access to whatever address the DMA-buf on the host is pointing to (system, VRAM, doorbell, IOMMU, whatever).<o:p></o:p></pre> <pre><o:p> </o:p></pre> <pre>I'm also not 100% sure if all the cache snooping is done correctly in all cases, but for now it seems to work.<o:p></o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>Frankly the more I look at the original patch that added vram export<o:p></o:p></pre> <pre>support the more this just looks like a "pls revert, this is just too<o:p></o:p></pre> <pre>broken".<o:p></o:p></pre> </blockquote> <pre>The commit I mean is this one: ea5ea3d8a117 ("drm/virtio: support mapping<o:p></o:p></pre> <pre>exported vram"). The commit message definitely needs to cite that one, and<o:p></o:p></pre> <pre>also needs a cc: stable because not rejecting invalid imports is a pretty<o:p></o:p></pre> <pre>big deal.<o:p></o:p></pre> </blockquote> <pre>Yeah, I've pointed out that commit in an internal discussion as well. I was just not aware that it's that severely broken.<o:p></o:p></pre> <pre><o:p> </o:p></pre> </blockquote> <pre>Yeah we have mentioned this patch before, but I don't totally understand why this is too broken. Without exporting vram objects, dGPU prime feature would not be realized.<o:p></o:p></pre> <pre>Would you mind to explain more about it. Thanks!<o:p></o:p></pre> </blockquote> One reason is that using sg tables without struct pages is actually a hack we came up with because we couldn't hope to clean up the sg table structure any time soon to not include struct page pointers. Another reason is that using this with devices which don't expect a DMA address pointing into a virtual PCI BAR. So doing this without checking the peer2peer flag can most likely cause quite a bit of trouble. Regards, Christian. <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>Best regards,<o:p></o:p></pre> <pre>Julia<o:p></o:p></pre> <pre><o:p> </o:p></pre> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <pre>Regards,<o:p></o:p></pre> <pre>Christian.<o:p></o:p></pre> <pre><o:p> </o:p></pre> </blockquote> </blockquote> </blockquote> </div> </body> </html>