<html> <head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body text="#000000" bgcolor="#FFFFFF"> <div class="moz-cite-prefix">Am 27.09.2018 um 15:18 schrieb Kuehling, Felix: </div> <blockquote type="cite" cite="mid:DM5PR12MB17072AE77EB0DE3AD22B8D7892140@DM5PR12MB1707.namprd12.prod.outlook.com"> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> <meta name="Generator" content="Microsoft Word 15 (filtered medium)"> <style></style> <div class="WordSection1"> > The problem is here: ><o:p></o:p> > ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);<o:p></o:p> > amdgpu_mn_unlock(p->mn); <o:p></o:p> > > We need to hold the lock until the fence is added to the reservation object. > > Otherwise somebody could have changed the page tables just in the moment between the check of amdgpu_ttm_tt_userptr_needs_pages() and adding the fence to the reservation object. <o:p></o:p> I’m not planning to change that. I don’t think there is any need to change it. </div> </blockquote> Yeah, but when HMM doesn't provide both the start and the end hock of the invalidation this way won't work any more. So we need to find a solution for this, Christian. <blockquote type="cite" cite="mid:DM5PR12MB17072AE77EB0DE3AD22B8D7892140@DM5PR12MB1707.namprd12.prod.outlook.com"> <div class="WordSection1"> <o:p></o:p> <o:p> </o:p> Regards,<o:p></o:p> Felix<o:p></o:p> <o:p> </o:p> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in"> From: Koenig, Christian Sent: Thursday, September 27, 2018 7:24 AM To: Kuehling, Felix <a class="moz-txt-link-rfc2396E" href="mailto:Felix.Kuehling@amd.com"><Felix.Kuehling@amd.com></a> Cc: Yang, Philip <a class="moz-txt-link-rfc2396E" href="mailto:Philip.Yang@amd.com"><Philip.Yang@amd.com></a>; <a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a>; Jerome Glisse <a class="moz-txt-link-rfc2396E" href="mailto:j.glisse@gmail.com"><j.glisse@gmail.com></a> Subject: Re: [PATCH] drm/amdgpu: use HMM mirror callback to replace mmu notifier v4<o:p></o:p> </div> </div> <o:p> </o:p> <div> Am 27.09.2018 um 13:08 schrieb Kuehling, Felix:<o:p></o:p> </div> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> > We double check that there wasn't any page table modification while we prepared the submission and restart the whole process when there actually was some update. > > The reason why we need to do this is here: ><o:p></o:p> > ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence); > amdgpu_mn_unlock(p->mn);<o:p></o:p> > > Only after the new fence is added to the buffer object we can release the lock so that any invalidation will now block on our command submission to finish before it modifies the page table. <o:p></o:p> I don’t see why this requires holding the read-lock until invalidate_range_end. amdgpu_ttm_tt_affect_userptr gets called while the mn read-lock is held in invalidate_range_start notifier.<o:p></o:p> </blockquote> That's not related to amdgpu_ttm_tt_affect_userptr(), this function could actually be called outside the lock. The problem is here: <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);<o:p></o:p> </blockquote> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> amdgpu_mn_unlock(p->mn); <o:p></o:p> </blockquote> We need to hold the lock until the fence is added to the reservation object. Otherwise somebody could have changed the page tables just in the moment between the check of amdgpu_ttm_tt_userptr_needs_pages() and adding the fence to the reservation object. Regards, Christian. <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> <o:p></o:p> Regards,<o:p></o:p> Felix<o:p></o:p> <o:p></o:p> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in"> From: Koenig, Christian Sent: Thursday, September 27, 2018 5:27 AM To: Kuehling, Felix <a href="mailto:Felix.Kuehling@amd.com" moz-do-not-send="true"><Felix.Kuehling@amd.com></a> Cc: Yang, Philip <a href="mailto:Philip.Yang@amd.com" moz-do-not-send="true"><Philip.Yang@amd.com></a>; <a href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a>; Jerome Glisse <a href="mailto:j.glisse@gmail.com" moz-do-not-send="true"><j.glisse@gmail.com></a> Subject: Re: [PATCH] drm/amdgpu: use HMM mirror callback to replace mmu notifier v4<o:p></o:p> </div> </div> <o:p></o:p> <div> That is correct, but take a look what we do when after calling the amdgpu_mn_read_lock(): <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> /* No memory allocation is allowed while holding the mn lock */ amdgpu_mn_lock(p->mn); amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) { struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo); if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm)) { r = -ERESTARTSYS; goto error_abort; } }<o:p></o:p> </blockquote> We double check that there wasn't any page table modification while we prepared the submission and restart the whole process when there actually was some update. The reason why we need to do this is here: <o:p></o:p> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence); amdgpu_mn_unlock(p->mn);<o:p></o:p> </blockquote> Only after the new fence is added to the buffer object we can release the lock so that any invalidation will now block on our command submission to finish before it modifies the page table. The only other option would be to add the fence first and then check if there was any update to the page tables. The issue with that approach is that adding a fence can't be made undone, so if we find that there actually was an update to the page tables we would need to somehow turn the CS into a dummy (e.g. overwrite all IBs with NOPs or something like that) and still submit it. Not sure if that is actually possible. Regards, Christian. Am 27.09.2018 um 10:47 schrieb Kuehling, Felix:<o:p></o:p> </div> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> So back to my previous question:<o:p></o:p> <o:p></o:p> >> But do we really need another lock for this? Wouldn't the <o:p></o:p> >> re-validation of userptr BOs (currently calling get_user_pages) force <o:p></o:p> >> synchronization with the ongoing page table invalidation through the <o:p></o:p> >> mmap_sem or other MM locks?<o:p></o:p> > <o:p></o:p> > No and yes. We don't hold any other locks while doing command submission, but I expect that HMM has its own mechanism to prevent that.<o:p></o:p> > <o:p></o:p> > Since we don't modify amdgpu_mn_lock()/amdgpu_mn_unlock() we are certainly not using this mechanism correctly.<o:p></o:p> <o:p></o:p> The existing amdgpu_mn_lock/unlock should block the MMU notifier while a command submission is in progress. It should also block command submission while an MMU notifier is in progress.<o:p></o:p> <o:p></o:p> What we lose with HMM is the ability to hold a read-lock for the entire duration of the invalidate_range_start until invalidate_range_end. As I understand it, that lock is meant to prevent new command submissions while the page tables are being updated by the kernel. But my point is, that get_user_pages or hmm_vma_fault should do the same kind of thing. Before the command submission can go ahead, it needs to update the userptr addresses. If the page tables are still being updated, it will block there even without holding the amdgpu_mn_read_lock.<o:p></o:p> <o:p></o:p> Regards,<o:p></o:p> Felix<o:p></o:p> <o:p></o:p> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in"> From: Koenig, Christian Sent: Thursday, September 27, 2018 3:00 AM To: Kuehling, Felix <a href="mailto:Felix.Kuehling@amd.com" moz-do-not-send="true"><Felix.Kuehling@amd.com></a> Cc: Yang, Philip <a href="mailto:Philip.Yang@amd.com" moz-do-not-send="true"><Philip.Yang@amd.com></a>; <a href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a>; Jerome Glisse <a href="mailto:j.glisse@gmail.com" moz-do-not-send="true"><j.glisse@gmail.com></a> Subject: RE: [PATCH] drm/amdgpu: use HMM mirror callback to replace mmu notifier v4<o:p></o:p> </div> </div> <o:p></o:p> <div> <div> No, that won't work. We would still run into lock inversion problems.<o:p></o:p> <div> <o:p></o:p> </div> <div> What we could do with the scheduler is to turn submissions into dummies if we find that the page tables are now outdated.<o:p></o:p> </div> <div> <o:p></o:p> </div> <div> But that would be really hacky and I'm not sure if that would really work in all cases.<o:p></o:p> </div> <div> <o:p></o:p> </div> <div> Christian.<o:p></o:p> </div> </div> <div> <o:p></o:p> <div> Am 27.09.2018 08:53 schrieb "Kuehling, Felix" <<a href="mailto:Felix.Kuehling@amd.com" moz-do-not-send="true">Felix.Kuehling@amd.com</a>>:<o:p></o:p> </div> </div> </div> <div> I had a chat with Jerome yesterday. He pointed out that the new blockable parameter can be used to infer whether the MMU notifier is being called in a reclaim operation. So if blockable==true, it should even be safe to take the BO reservation lock without problems. I think with that we should be able to remove the read-write locking completely and go back to locking (or try-locking for blockable==false) the reservation locks in the MMU notifier? Regards, Felix -----Original Message----- From: amd-gfx <<a href="mailto:amd-gfx-bounces@lists.freedesktop.org" moz-do-not-send="true">amd-gfx-bounces@lists.freedesktop.org</a>> On Behalf Of Christian König Sent: Saturday, September 15, 2018 3:47 AM To: Kuehling, Felix <<a href="mailto:Felix.Kuehling@amd.com" moz-do-not-send="true">Felix.Kuehling@amd.com</a>>; Yang, Philip <<a href="mailto:Philip.Yang@amd.com" moz-do-not-send="true">Philip.Yang@amd.com</a>>; <a href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a>; Jerome Glisse <<a href="mailto:j.glisse@gmail.com" moz-do-not-send="true">j.glisse@gmail.com</a>> Subject: Re: [PATCH] drm/amdgpu: use HMM mirror callback to replace mmu notifier v4 Am 14.09.2018 um 22:21 schrieb Felix Kuehling: > On 2018-09-14 01:52 PM, Christian König wrote: >> Am 14.09.2018 um 19:47 schrieb Philip Yang: >>> On 2018-09-14 03:51 AM, Christian König wrote: >>>> Am 13.09.2018 um 23:51 schrieb Felix Kuehling: >>>>> On 2018-09-13 04:52 PM, Philip Yang wrote: >>>>> [SNIP] >>>>>> + amdgpu_mn_read_unlock(amn); >>>>>> + >>>>> amdgpu_mn_read_lock/unlock support recursive locking for multiple >>>>> overlapping or nested invalidation ranges. But if you'r locking >>>>> and unlocking in the same function. Is that still a concern? >>> I don't understand the possible recursive case, but >>> amdgpu_mn_read_lock() still support recursive locking. >>>> Well the real problem is that unlocking them here won't work. >>>> >>>> We need to hold the lock until we are sure that the operation which >>>> updates the page tables is completed. >>>> >>> The reason for this change is because hmm mirror has >>> invalidate_start callback, no invalidate_end callback >>> >>> Check mmu_notifier.c and hmm.c again, below is entire logic to >>> update CPU page tables and callback: >>> >>> mn lock amn->lock is used to protect interval tree access because >>> user may submit/register new userptr anytime. >>> This is same for old and new way. >>> >>> step 2 guarantee the GPU operation is done before updating CPU page >>> table. >>> >>> So I think the change is safe. We don't need hold mn lock until the >>> CPU page tables update is completed. >> No, that isn't even remotely correct. The lock doesn't protects the >> interval tree. >> >>> Old: >>> 1. down_read_non_owner(&amn->lock) >>> 2. loop to handle BOs from node->bos through interval tree >>> amn->object nodes >>> gfx: wait for pending BOs fence operation done, mark user >>> pages dirty >>> kfd: evict user queues of the process, wait for queue >>> unmap/map operation done >>> 3. update CPU page tables >>> 4. up_read(&amn->lock) >>> >>> New, switch step 3 and 4 >>> 1. down_read_non_owner(&amn->lock) >>> 2. loop to handle BOs from node->bos through interval tree >>> amn->object nodes >>> gfx: wait for pending BOs fence operation done, mark user >>> pages dirty >>> kfd: evict user queues of the process, wait for queue >>> unmap/map operation done >>> 3. up_read(&amn->lock) >>> 4. update CPU page tables >> The lock is there to make sure that we serialize page table updates >> with command submission. > As I understand it, the idea is to prevent command submission (adding > new fences to BOs) while a page table invalidation is in progress. Yes, exactly. > But do we really need another lock for this? Wouldn't the > re-validation of userptr BOs (currently calling get_user_pages) force > synchronization with the ongoing page table invalidation through the > mmap_sem or other MM locks? No and yes. We don't hold any other locks while doing command submission, but I expect that HMM has its own mechanism to prevent that. Since we don't modify amdgpu_mn_lock()/amdgpu_mn_unlock() we are certainly not using this mechanism correctly. Regards, Christian. _______________________________________________ amd-gfx mailing list <a href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a> <a href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a><o:p></o:p> </div> </blockquote> <o:p></o:p> </blockquote> <o:p> </o:p> </div> </blockquote> </body> </html>