<html><head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body> <div class="moz-cite-prefix">On 2023-09-15 6:19, Christian König wrote: </div> <blockquote type="cite" cite="mid:39f0c4d9-0959-73ed-9bca-43a342fb906a@gmail.com">Am 15.09.23 um 10:53 schrieb Lang Yu: <blockquote type="cite">On 09/14/ , Felix Kuehling wrote: <blockquote type="cite">On 2023-09-14 10:02, Christian König wrote: </blockquote> Do we still need to use legacy flush to emulate heavyweight flush if we don't use SVM? And can I push this now? </blockquote> Felix needs to decide that. From what I understand the KFD needs heavyweight flushes for secure SVM operation. </blockquote> Yes. We need to be able to guarantee to the kernel, that the GPU will not access unmapped memory. There are two strategies in the driver to do this: <ol> <li>Preempt GPU queues (which implies a heavy-weight TLB flush in the scheduler firmware)</li> <li>Invalidate page table entries and flush TLBs </li> </ol> #1 happens during MMU notifiers with XNACK off. #2 happens in MMU notifiers with XNACK on (not supported on GFX10.x) and when unified memory us munmapped. It's that last part I'm worried about. When memory is munmapped and given back to the OS, we need to be able to guarantee that the GPU won't access it any more. The same is true when GTT BOs and userptr BOs are freed. After unmapping them from the GPU page tables, we need a heavy-weight flush. I believe the same should apply to the graphics driver, but maybe that's implied through the CS and fence mechanisms that keep memory allocated while the GPU is accessing it. A legacy flush has a slim chance of not being sufficient because memory accesses using old addresses can still be in flight in the GPU. <blockquote type="cite" cite="mid:39f0c4d9-0959-73ed-9bca-43a342fb906a@gmail.com"> If heavyweight flushes are buggy papering over that by using legacy flushes is only a mediocre workaround. </blockquote> I agree. I'd like to avoid half-baked workarounds that will cause more headaches later on. I started an internal email thread with Tony to understand the requirements for heavy-weight flushes on the affected GPUs and find a better workaround. Regards, Felix <blockquote type="cite" cite="mid:39f0c4d9-0959-73ed-9bca-43a342fb906a@gmail.com"> Regards, Christian. <blockquote type="cite"> Regards, Lang <blockquote type="cite"> <blockquote type="cite">Am 14.09.23 um 15:59 schrieb Felix Kuehling: <blockquote type="cite">On 2023-09-14 9:39, Christian König wrote: <blockquote type="cite">Is a single legacy flush sufficient to emulate an heavyweight flush as well? On previous generations we needed to issue at least two legacy flushes for this. </blockquote> I assume you are referring to the Vega20 XGMI workaround. That is a very different issue. Because PTEs would be cached in L2, we had to always use a heavy-weight flush that would also flush the L2 cache as well, and follow that with another legacy flush to deal with race conditions where stale PTEs could be re-fetched from L2 before the L2 flush was complete. </blockquote> No, we also have another (badly documented) workaround which issues a legacy flush before each heavy weight on some hw generations. See the my TLB flush cleanup patches. <blockquote type="cite">A heavy-weight flush guarantees that there are no more possible memory accesses using the old PTEs. With physically addressed caches on GFXv9 that includes a cache flush because the address translation happened before putting data into the cache. I think the address translation and cache architecture works differently on GFXv10. So maybe the cache-flush is not required here. But even then a legacy flush probably allows for in-flight memory accesses with old physical addresses to complete after the TLB flush. So there is a small risk of memory corruption that was assumed to not be accessed by the GPU any more. Or when using IOMMU device isolation it would result in IOMMU faults if the DMA mappings are invalidated slightly too early. </blockquote> Mhm, that's quite bad. Any idea how to avoid that? </blockquote> A few ideas * Add an arbitrary delay and hope that it is longer than the FIFOs in the HW * Execute an atomic operation to memory on some GPU engine that could act as a fence, maybe just a RELEASE_MEM on the CP to some writeback location would do the job * If needed, RELEASE_MEM could also perform a cache flush Regards, Felix <blockquote type="cite">Regards, Christian. <blockquote type="cite">Regards, Felix <blockquote type="cite">And please don't push before getting an rb from Felix as well. Regards, Christian. Am 14.09.23 um 11:23 schrieb Lang Yu: <blockquote type="cite">cyan_skilfish has problems with other flush types. v2: fix incorrect ternary conditional operator usage.(Yifan) Signed-off-by: Lang Yu <a class="moz-txt-link-rfc2396E" href="mailto:Lang.Yu@amd.com"><Lang.Yu@amd.com></a> Cc: <a class="moz-txt-link-rfc2396E" href="mailto:stable@vger.kernel.org"><stable@vger.kernel.org></a> # v5.15+ --- drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c index d3da13f4c80e..c6d11047169a 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c @@ -236,7 +236,8 @@ static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid, { bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub); struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; - u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type); + u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, + (adev->asic_type != CHIP_CYAN_SKILLFISH) ? flush_type : 0); u32 tmp; /* Use register 17 for GART */ const unsigned int eng = 17; @@ -331,6 +332,8 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, int r; + flush_type = (adev->asic_type != CHIP_CYAN_SKILLFISH) ? flush_type : 0; + /* flush hdp cache */ adev->hdp.funcs->flush_hdp(adev, NULL); @@ -426,6 +429,8 @@ static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, struct amdgpu_ring *ring = &adev->gfx.kiq[0].ring; struct amdgpu_kiq *kiq = &adev->gfx.kiq[0]; + flush_type = (adev->asic_type != CHIP_CYAN_SKILLFISH) ? flush_type : 0; + if (amdgpu_emu_mode == 0 && ring->sched.ready) { spin_lock(&adev->gfx.kiq[0].ring_lock); /* 2 dwords flush + 8 dwords fence */ </blockquote> </blockquote> </blockquote> </blockquote> </blockquote> </blockquote> </blockquote> </body> </html>