RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
Fantastic, I'm integrating it now. Fingers crossed! Kent > -Original Message- > From: Kuehling, Felix > Sent: Tuesday, March 12, 2019 7:31 PM > To: Yang, Philip ; Russell, Kent > ; Koenig, Christian ; > amd-gfx@lists.freedesktop.org > Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > > Never mind. I must have messed up my build. I can't reproduce the problem > any more. The patch I sent out is still needed and valid. AFAICT it should be > all > that's needed to fix GPUVM for KFD. > > I have not seen any faults with KFDCWSRTest.BasicTest on my system with > Fiji or Vega10 with that patch applied. > > Regards, > Felix > > On 2019-03-12 5:19 p.m., Felix Kuehling wrote: > > I'm also still seeing VM faults in the eviction test even with my fix, > > and even with SDMA page table updates. There is still something else > > going wrong. :/ > > > > Thanks, > > Felix > > > > On 2019-03-12 5:13 p.m., Yang, Philip wrote: > >> vm fault happens about 1/10 for KFDCWSRTest.BasicTest for me. I am > >> using SDMA for page table update. I don't try CPU page table update. > >> > >> Philip > >> > >> On 2019-03-12 11:12 a.m., Russell, Kent wrote: > >>> Peculiar, I hit it immediately when I ran it . Can you try use > >>> --gtest_filter=KFDCWSRTest.BasicTest . That one hung every time for > me. > >>> > >>> Kent > >>> > >>>> -----Original Message- > >>>> From: Christian König > >>>> Sent: Tuesday, March 12, 2019 11:09 AM > >>>> To: Russell, Kent ; Koenig, Christian > >>>> ; Kuehling, Felix > >>>> ; amd-gfx@lists.freedesktop.org > >>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on > demand > >>>> > >>>> Yeah, same problem here. > >>>> > >>>> I removed libhsakmt package and installed it manually and now it > >>>> seems to work. > >>>> > >>>> Doing some testing now, but at least of hand I can't seem to > >>>> reproduce the VM fault on a Vega10. > >>>> > >>>> Christian. > >>>> > >>>> Am 12.03.19 um 16:01 schrieb Russell, Kent: > >>>>> Oh right, I remember that issue. I had that happen to me once, > >>>>> where my > >>>> installed libhsakmt didn't match up with the latest source code, so > >>>> I ended up having to remove the libhsakmt package and pointing it > >>>> to the folders instead. > >>>>> Kent > >>>>> > >>>>>> -Original Message- > >>>>>> From: Koenig, Christian > >>>>>> Sent: Tuesday, March 12, 2019 10:49 AM > >>>>>> To: Russell, Kent ; Kuehling, Felix > >>>>>> ; amd-gfx@lists.freedesktop.org > >>>>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on > >>>>>> demand > >>>>>> > >>>>>> Yeah, the problem is I do have the libhsakmt installed. > >>>>>> > >>>>>> Going to give it a try to specify the directory directly. > >>>>>> > >>>>>> Christian. > >>>>>> > >>>>>> Am 12.03.19 um 15:47 schrieb Russell, Kent: > >>>>>>> The README.txt file inside the tests/kfdtest folder has > >>>>>>> instructions on how > >>>>>> to do it if you don't have the libhsakmt package installed on > >>>>>> your system: > >>>>>>> export LIBHSAKMT_PATH=/*your local libhsakmt folder*/ With > that, > >>>>>>> the headers and libraries are searched under > >>>>>>> LIBHSAKMT_PATH/include and LIBHSAKMT_PATH/lib respectively. > >>>>>>> > >>>>>>> So if you try export LIBHSAKMT_PATH as the root ROCT folder (the > >>>>>>> one > >>>>>> containing include, src, tests, etc), then that should cover it. > >>>>>>> Kent > >>>>>>> > >>>>>>> > >>>>>>>> -Original Message- > >>>>>>>> From: Christian König > >>>>>>>> Sent: Tuesday, March 12, 2019 9:13 AM > >>>>>>>> To: Russell, Kent ;
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
Never mind. I must have messed up my build. I can't reproduce the problem any more. The patch I sent out is still needed and valid. AFAICT it should be all that's needed to fix GPUVM for KFD. I have not seen any faults with KFDCWSRTest.BasicTest on my system with Fiji or Vega10 with that patch applied. Regards, Felix On 2019-03-12 5:19 p.m., Felix Kuehling wrote: > I'm also still seeing VM faults in the eviction test even with my fix, > and even with SDMA page table updates. There is still something else > going wrong. :/ > > Thanks, > Felix > > On 2019-03-12 5:13 p.m., Yang, Philip wrote: >> vm fault happens about 1/10 for KFDCWSRTest.BasicTest for me. I am using >> SDMA for page table update. I don't try CPU page table update. >> >> Philip >> >> On 2019-03-12 11:12 a.m., Russell, Kent wrote: >>> Peculiar, I hit it immediately when I ran it . Can you try use >>> --gtest_filter=KFDCWSRTest.BasicTest . That one hung every time for me. >>> >>> Kent >>> >>>> -Original Message- >>>> From: Christian König >>>> Sent: Tuesday, March 12, 2019 11:09 AM >>>> To: Russell, Kent ; Koenig, Christian >>>> ; Kuehling, Felix ; >>>> amd-gfx@lists.freedesktop.org >>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>>> >>>> Yeah, same problem here. >>>> >>>> I removed libhsakmt package and installed it manually and now it >>>> seems to >>>> work. >>>> >>>> Doing some testing now, but at least of hand I can't seem to >>>> reproduce the >>>> VM fault on a Vega10. >>>> >>>> Christian. >>>> >>>> Am 12.03.19 um 16:01 schrieb Russell, Kent: >>>>> Oh right, I remember that issue. I had that happen to me once, >>>>> where my >>>> installed libhsakmt didn't match up with the latest source code, so >>>> I ended up >>>> having to remove the libhsakmt package and pointing it to the folders >>>> instead. >>>>> Kent >>>>> >>>>>> -Original Message- >>>>>> From: Koenig, Christian >>>>>> Sent: Tuesday, March 12, 2019 10:49 AM >>>>>> To: Russell, Kent ; Kuehling, Felix >>>>>> ; amd-gfx@lists.freedesktop.org >>>>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>>>>> >>>>>> Yeah, the problem is I do have the libhsakmt installed. >>>>>> >>>>>> Going to give it a try to specify the directory directly. >>>>>> >>>>>> Christian. >>>>>> >>>>>> Am 12.03.19 um 15:47 schrieb Russell, Kent: >>>>>>> The README.txt file inside the tests/kfdtest folder has >>>>>>> instructions >>>>>>> on how >>>>>> to do it if you don't have the libhsakmt package installed on >>>>>> your system: >>>>>>> export LIBHSAKMT_PATH=/*your local libhsakmt folder*/ With that, >>>>>>> the >>>>>>> headers and libraries are searched under LIBHSAKMT_PATH/include and >>>>>>> LIBHSAKMT_PATH/lib respectively. >>>>>>> >>>>>>> So if you try export LIBHSAKMT_PATH as the root ROCT folder (the >>>>>>> one >>>>>> containing include, src, tests, etc), then that should cover it. >>>>>>> Kent >>>>>>> >>>>>>> >>>>>>>> -Original Message- >>>>>>>> From: Christian König >>>>>>>> Sent: Tuesday, March 12, 2019 9:13 AM >>>>>>>> To: Russell, Kent ; Kuehling, Felix >>>>>>>> ; Koenig, Christian >>>>>>>> ; amd-gfx@lists.freedesktop.org >>>>>>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on >>>> demand >>>>>>>> Hi guys, >>>>>>>> >>>>>>>> so found a few minutes today to compile kfdtest. >>>>>>>> >>>>>>>> Problem is that during the compile I get a lots of this: >>>>>>>>> CMakeFiles/kfdtest.dir/src/BaseQueue.cpp.o: In Funktion >>>>>>>>> »BaseQueue::Create(unsigned int, unsigned int, unsigned l
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
I'm also still seeing VM faults in the eviction test even with my fix, and even with SDMA page table updates. There is still something else going wrong. :/ Thanks, Felix On 2019-03-12 5:13 p.m., Yang, Philip wrote: > vm fault happens about 1/10 for KFDCWSRTest.BasicTest for me. I am using > SDMA for page table update. I don't try CPU page table update. > > Philip > > On 2019-03-12 11:12 a.m., Russell, Kent wrote: >> Peculiar, I hit it immediately when I ran it . Can you try use >> --gtest_filter=KFDCWSRTest.BasicTest . That one hung every time for me. >> >>Kent >> >>> -Original Message- >>> From: Christian König >>> Sent: Tuesday, March 12, 2019 11:09 AM >>> To: Russell, Kent ; Koenig, Christian >>> ; Kuehling, Felix ; >>> amd-gfx@lists.freedesktop.org >>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>> >>> Yeah, same problem here. >>> >>> I removed libhsakmt package and installed it manually and now it seems to >>> work. >>> >>> Doing some testing now, but at least of hand I can't seem to reproduce the >>> VM fault on a Vega10. >>> >>> Christian. >>> >>> Am 12.03.19 um 16:01 schrieb Russell, Kent: >>>> Oh right, I remember that issue. I had that happen to me once, where my >>> installed libhsakmt didn't match up with the latest source code, so I ended >>> up >>> having to remove the libhsakmt package and pointing it to the folders >>> instead. >>>> Kent >>>> >>>>> -Original Message- >>>>> From: Koenig, Christian >>>>> Sent: Tuesday, March 12, 2019 10:49 AM >>>>> To: Russell, Kent ; Kuehling, Felix >>>>> ; amd-gfx@lists.freedesktop.org >>>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>>>> >>>>> Yeah, the problem is I do have the libhsakmt installed. >>>>> >>>>> Going to give it a try to specify the directory directly. >>>>> >>>>> Christian. >>>>> >>>>> Am 12.03.19 um 15:47 schrieb Russell, Kent: >>>>>> The README.txt file inside the tests/kfdtest folder has instructions >>>>>> on how >>>>> to do it if you don't have the libhsakmt package installed on your system: >>>>>> export LIBHSAKMT_PATH=/*your local libhsakmt folder*/ With that, the >>>>>> headers and libraries are searched under LIBHSAKMT_PATH/include and >>>>>> LIBHSAKMT_PATH/lib respectively. >>>>>> >>>>>> So if you try export LIBHSAKMT_PATH as the root ROCT folder (the one >>>>> containing include, src, tests, etc), then that should cover it. >>>>>> Kent >>>>>> >>>>>> >>>>>>> -Original Message- >>>>>>> From: Christian König >>>>>>> Sent: Tuesday, March 12, 2019 9:13 AM >>>>>>> To: Russell, Kent ; Kuehling, Felix >>>>>>> ; Koenig, Christian >>>>>>> ; amd-gfx@lists.freedesktop.org >>>>>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on >>> demand >>>>>>> Hi guys, >>>>>>> >>>>>>> so found a few minutes today to compile kfdtest. >>>>>>> >>>>>>> Problem is that during the compile I get a lots of this: >>>>>>>> CMakeFiles/kfdtest.dir/src/BaseQueue.cpp.o: In Funktion >>>>>>>> »BaseQueue::Create(unsigned int, unsigned int, unsigned long*)«: >>>>>>>> /usr/src/ROCT-Thunk-Interface/tests/kfdtest/src/BaseQueue.cpp:57: >>>>>>>> Warnung: undefinierter Verweis auf »hsaKmtCreateQueue« >>>>>>> Any idea? >>>>>>> >>>>>>> Christian. >>>>>>> >>>>>>> Am 11.03.19 um 17:55 schrieb Christian König: >>>>>>>> Hi guys, >>>>>>>> >>>>>>>> well it's most likely some missing handling in the KFD, so I'm >>>>>>>> rather reluctant to revert the change immediately. >>>>>>>> >>>>>>>> Problem is that I don't have time right now to look into it >>>>>>>> immediately. So Kent can you continue to take a look? >>&g
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
vm fault happens about 1/10 for KFDCWSRTest.BasicTest for me. I am using SDMA for page table update. I don't try CPU page table update. Philip On 2019-03-12 11:12 a.m., Russell, Kent wrote: > Peculiar, I hit it immediately when I ran it . Can you try use > --gtest_filter=KFDCWSRTest.BasicTest . That one hung every time for me. > > Kent > >> -Original Message- >> From: Christian König >> Sent: Tuesday, March 12, 2019 11:09 AM >> To: Russell, Kent ; Koenig, Christian >> ; Kuehling, Felix ; >> amd-gfx@lists.freedesktop.org >> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >> >> Yeah, same problem here. >> >> I removed libhsakmt package and installed it manually and now it seems to >> work. >> >> Doing some testing now, but at least of hand I can't seem to reproduce the >> VM fault on a Vega10. >> >> Christian. >> >> Am 12.03.19 um 16:01 schrieb Russell, Kent: >>> Oh right, I remember that issue. I had that happen to me once, where my >> installed libhsakmt didn't match up with the latest source code, so I ended >> up >> having to remove the libhsakmt package and pointing it to the folders >> instead. >>> >>>Kent >>> >>>> -----Original Message----- >>>> From: Koenig, Christian >>>> Sent: Tuesday, March 12, 2019 10:49 AM >>>> To: Russell, Kent ; Kuehling, Felix >>>> ; amd-gfx@lists.freedesktop.org >>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>>> >>>> Yeah, the problem is I do have the libhsakmt installed. >>>> >>>> Going to give it a try to specify the directory directly. >>>> >>>> Christian. >>>> >>>> Am 12.03.19 um 15:47 schrieb Russell, Kent: >>>>> The README.txt file inside the tests/kfdtest folder has instructions >>>>> on how >>>> to do it if you don't have the libhsakmt package installed on your system: >>>>> export LIBHSAKMT_PATH=/*your local libhsakmt folder*/ With that, the >>>>> headers and libraries are searched under LIBHSAKMT_PATH/include and >>>>> LIBHSAKMT_PATH/lib respectively. >>>>> >>>>> So if you try export LIBHSAKMT_PATH as the root ROCT folder (the one >>>> containing include, src, tests, etc), then that should cover it. >>>>> Kent >>>>> >>>>> >>>>>> -Original Message- >>>>>> From: Christian König >>>>>> Sent: Tuesday, March 12, 2019 9:13 AM >>>>>> To: Russell, Kent ; Kuehling, Felix >>>>>> ; Koenig, Christian >>>>>> ; amd-gfx@lists.freedesktop.org >>>>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on >> demand >>>>>> >>>>>> Hi guys, >>>>>> >>>>>> so found a few minutes today to compile kfdtest. >>>>>> >>>>>> Problem is that during the compile I get a lots of this: >>>>>>> CMakeFiles/kfdtest.dir/src/BaseQueue.cpp.o: In Funktion >>>>>>> »BaseQueue::Create(unsigned int, unsigned int, unsigned long*)«: >>>>>>> /usr/src/ROCT-Thunk-Interface/tests/kfdtest/src/BaseQueue.cpp:57: >>>>>>> Warnung: undefinierter Verweis auf »hsaKmtCreateQueue« >>>>>> Any idea? >>>>>> >>>>>> Christian. >>>>>> >>>>>> Am 11.03.19 um 17:55 schrieb Christian König: >>>>>>> Hi guys, >>>>>>> >>>>>>> well it's most likely some missing handling in the KFD, so I'm >>>>>>> rather reluctant to revert the change immediately. >>>>>>> >>>>>>> Problem is that I don't have time right now to look into it >>>>>>> immediately. So Kent can you continue to take a look? >>>>>>> >>>>>>> Sounds like its crashing immediately, so it should be something >> obvious. >>>>>>> >>>>>>> Christian. >>>>>>> >>>>>>> Am 11.03.19 um 10:49 schrieb Russell, Kent: >>>>>>>> From what I've been able to dig through, the VM Fault seems to >>>>>>>> occur right after a doorbell mmap, but that's as far as I got. I >>>>>>>> can try to revert it in
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
The root cause is that we don't wait after calling amdgpu_vm_clear_bo in amdgpu_vm_alloc_pts. Waiting for the page table BOs to be idle for CPU page table updates is done in amdgpu_vm_bo_update_mapping. That is now *before* the page tables are actually allocated and cleared in amdgpu_vm_update_ptes. We'll need to move the waiting for page tables to be idle into amdgpu_vm_alloc_pts or amdgpu_vm_update_ptes. Regards, Felix On 2019-03-12 3:02 p.m., Felix Kuehling wrote: > I find that it's related to CPU page table updates. If I force page > table updates with SDMA, I don't get the VM fault. > > Regards, > Felix > > On 2019-03-11 12:55 p.m., Christian König wrote: >> Hi guys, >> >> well it's most likely some missing handling in the KFD, so I'm rather >> reluctant to revert the change immediately. >> >> Problem is that I don't have time right now to look into it >> immediately. So Kent can you continue to take a look? >> >> Sounds like its crashing immediately, so it should be something obvious. >> >> Christian. >> >> Am 11.03.19 um 10:49 schrieb Russell, Kent: >>> From what I've been able to dig through, the VM Fault seems to >>> occur right after a doorbell mmap, but that's as far as I got. I can >>> try to revert it in today's merge and see how things go. >>> >>> Kent >>> >>>> -----Original Message----- >>>> From: Kuehling, Felix >>>> Sent: Friday, March 08, 2019 11:16 PM >>>> To: Koenig, Christian ; Russell, Kent >>>> ; amd-gfx@lists.freedesktop.org >>>> Subject: RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>>> >>>> My concerns were related to eviction fence handing. It would >>>> manifest by >>>> unnecessary eviction callbacks into KFD that aren't cause by real >>>> evictions. I >>>> addressed that with a previous patch series that removed the need to >>>> remove eviction fences and add them back around page table updates in >>>> amdgpu_amdkfd_gpuvm.c. >>>> >>>> I don't know what's going on here. I can probably take a look on >>>> Monday. I >>>> haven't considered what changed with respect to PD updates. >>>> >>>> Kent, can we temporarily revert the offending change in >>>> amd-kfd-staging >>>> just to unblock the merge? >>>> >>>> Christian, I think KFD is currently broken on amd-staging-drm-next. >>>> If we're >>>> serious about supporting KFD upstream, you may also want to consider >>>> reverting your change there for now. Also consider building the >>>> Thunk and >>>> kfdtest so you can do quick smoke tests locally whenever you make >>>> amdgpu_vm changes that can affect KFD. >>>> https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface >>>> >>>> Regards, >>>> Felix >>>> >>>> -Original Message- >>>> From: amd-gfx On Behalf Of >>>> Christian König >>>> Sent: Friday, March 08, 2019 9:14 AM >>>> To: Russell, Kent ; >>>> amd-gfx@lists.freedesktop.org >>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>>> >>>> My best guess is that we forget somewhere to update the PDs. What >>>> hardware is that on? >>>> >>>> Felix already mentioned that this could be problematic for the KFD. >>>> >>>> Maybe he has an idea, >>>> Christian. >>>> >>>> Am 08.03.19 um 15:04 schrieb Russell, Kent: >>>>> Hi Christian, >>>>> >>>>> This patch ended up causing a VM Fault in KFDTest. Reverting just >>>>> this >>>> patch addressed the issue: >>>>> [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 >>>>> 0x480c for >>>> process pid 0 thread pid 0 >>>>> [ 82.703512] amdgpu :0c:00.0: >>>> VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x1000 >>>>> [ 82.703516] amdgpu :0c:00.0: >>>> VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x1004800C >>>>> [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid >>>>> 32769) at >>>> page 4096, read from 'TC0' (0x54433000) (72) >>>>> [ 82.703585] Evicting PASID 32769 queues >>>>> >>>>> I am look
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
I find that it's related to CPU page table updates. If I force page table updates with SDMA, I don't get the VM fault. Regards, Felix On 2019-03-11 12:55 p.m., Christian König wrote: > Hi guys, > > well it's most likely some missing handling in the KFD, so I'm rather > reluctant to revert the change immediately. > > Problem is that I don't have time right now to look into it > immediately. So Kent can you continue to take a look? > > Sounds like its crashing immediately, so it should be something obvious. > > Christian. > > Am 11.03.19 um 10:49 schrieb Russell, Kent: >> From what I've been able to dig through, the VM Fault seems to occur >> right after a doorbell mmap, but that's as far as I got. I can try to >> revert it in today's merge and see how things go. >> >> Kent >> >>> -Original Message- >>> From: Kuehling, Felix >>> Sent: Friday, March 08, 2019 11:16 PM >>> To: Koenig, Christian ; Russell, Kent >>> ; amd-gfx@lists.freedesktop.org >>> Subject: RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>> >>> My concerns were related to eviction fence handing. It would >>> manifest by >>> unnecessary eviction callbacks into KFD that aren't cause by real >>> evictions. I >>> addressed that with a previous patch series that removed the need to >>> remove eviction fences and add them back around page table updates in >>> amdgpu_amdkfd_gpuvm.c. >>> >>> I don't know what's going on here. I can probably take a look on >>> Monday. I >>> haven't considered what changed with respect to PD updates. >>> >>> Kent, can we temporarily revert the offending change in amd-kfd-staging >>> just to unblock the merge? >>> >>> Christian, I think KFD is currently broken on amd-staging-drm-next. >>> If we're >>> serious about supporting KFD upstream, you may also want to consider >>> reverting your change there for now. Also consider building the >>> Thunk and >>> kfdtest so you can do quick smoke tests locally whenever you make >>> amdgpu_vm changes that can affect KFD. >>> https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface >>> >>> Regards, >>> Felix >>> >>> -Original Message- >>> From: amd-gfx On Behalf Of >>> Christian König >>> Sent: Friday, March 08, 2019 9:14 AM >>> To: Russell, Kent ; amd-gfx@lists.freedesktop.org >>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>> >>> My best guess is that we forget somewhere to update the PDs. What >>> hardware is that on? >>> >>> Felix already mentioned that this could be problematic for the KFD. >>> >>> Maybe he has an idea, >>> Christian. >>> >>> Am 08.03.19 um 15:04 schrieb Russell, Kent: >>>> Hi Christian, >>>> >>>> This patch ended up causing a VM Fault in KFDTest. Reverting just this >>> patch addressed the issue: >>>> [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 >>>> 0x480c for >>> process pid 0 thread pid 0 >>>> [ 82.703512] amdgpu :0c:00.0: >>> VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x1000 >>>> [ 82.703516] amdgpu :0c:00.0: >>> VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x1004800C >>>> [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid >>>> 32769) at >>> page 4096, read from 'TC0' (0x54433000) (72) >>>> [ 82.703585] Evicting PASID 32769 queues >>>> >>>> I am looking into it, but if you have any insight that would be >>>> great in >>> helping to resolve it quickly. >>>> Kent >>>>> -Original Message- >>>>> From: amd-gfx On Behalf Of >>>>> Christian König >>>>> Sent: Tuesday, February 26, 2019 7:47 AM >>>>> To: amd-gfx@lists.freedesktop.org >>>>> Subject: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>>>> >>>>> Let's start to allocate VM PDs/PTs on demand instead of >>>>> pre-allocating them during mapping. >>>>> >>>>> Signed-off-by: Christian König >>>>> Reviewed-by: Felix Kuehling >>>>> --- >>>>> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 10 +- >>>>> drivers/gpu/drm/amd/am
RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
Peculiar, I hit it immediately when I ran it . Can you try use --gtest_filter=KFDCWSRTest.BasicTest . That one hung every time for me. Kent > -Original Message- > From: Christian König > Sent: Tuesday, March 12, 2019 11:09 AM > To: Russell, Kent ; Koenig, Christian > ; Kuehling, Felix ; > amd-gfx@lists.freedesktop.org > Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > > Yeah, same problem here. > > I removed libhsakmt package and installed it manually and now it seems to > work. > > Doing some testing now, but at least of hand I can't seem to reproduce the > VM fault on a Vega10. > > Christian. > > Am 12.03.19 um 16:01 schrieb Russell, Kent: > > Oh right, I remember that issue. I had that happen to me once, where my > installed libhsakmt didn't match up with the latest source code, so I ended up > having to remove the libhsakmt package and pointing it to the folders > instead. > > > > Kent > > > >> -Original Message- > >> From: Koenig, Christian > >> Sent: Tuesday, March 12, 2019 10:49 AM > >> To: Russell, Kent ; Kuehling, Felix > >> ; amd-gfx@lists.freedesktop.org > >> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > >> > >> Yeah, the problem is I do have the libhsakmt installed. > >> > >> Going to give it a try to specify the directory directly. > >> > >> Christian. > >> > >> Am 12.03.19 um 15:47 schrieb Russell, Kent: > >>> The README.txt file inside the tests/kfdtest folder has instructions > >>> on how > >> to do it if you don't have the libhsakmt package installed on your system: > >>> export LIBHSAKMT_PATH=/*your local libhsakmt folder*/ With that, the > >>> headers and libraries are searched under LIBHSAKMT_PATH/include and > >>> LIBHSAKMT_PATH/lib respectively. > >>> > >>> So if you try export LIBHSAKMT_PATH as the root ROCT folder (the one > >> containing include, src, tests, etc), then that should cover it. > >>>Kent > >>> > >>> > >>>> -Original Message- > >>>> From: Christian König > >>>> Sent: Tuesday, March 12, 2019 9:13 AM > >>>> To: Russell, Kent ; Kuehling, Felix > >>>> ; Koenig, Christian > >>>> ; amd-gfx@lists.freedesktop.org > >>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on > demand > >>>> > >>>> Hi guys, > >>>> > >>>> so found a few minutes today to compile kfdtest. > >>>> > >>>> Problem is that during the compile I get a lots of this: > >>>>> CMakeFiles/kfdtest.dir/src/BaseQueue.cpp.o: In Funktion > >>>>> »BaseQueue::Create(unsigned int, unsigned int, unsigned long*)«: > >>>>> /usr/src/ROCT-Thunk-Interface/tests/kfdtest/src/BaseQueue.cpp:57: > >>>>> Warnung: undefinierter Verweis auf »hsaKmtCreateQueue« > >>>> Any idea? > >>>> > >>>> Christian. > >>>> > >>>> Am 11.03.19 um 17:55 schrieb Christian König: > >>>>> Hi guys, > >>>>> > >>>>> well it's most likely some missing handling in the KFD, so I'm > >>>>> rather reluctant to revert the change immediately. > >>>>> > >>>>> Problem is that I don't have time right now to look into it > >>>>> immediately. So Kent can you continue to take a look? > >>>>> > >>>>> Sounds like its crashing immediately, so it should be something > obvious. > >>>>> > >>>>> Christian. > >>>>> > >>>>> Am 11.03.19 um 10:49 schrieb Russell, Kent: > >>>>>> From what I've been able to dig through, the VM Fault seems to > >>>>>> occur right after a doorbell mmap, but that's as far as I got. I > >>>>>> can try to revert it in today's merge and see how things go. > >>>>>> > >>>>>> Kent > >>>>>> > >>>>>>> -Original Message- > >>>>>>> From: Kuehling, Felix > >>>>>>> Sent: Friday, March 08, 2019 11:16 PM > >>>>>>> To: Koenig, Christian ; Russell, Kent > >>>>>>> ; amd-gfx@lists.freedesktop.org > >>>>>>> Subje
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
Yeah, same problem here. I removed libhsakmt package and installed it manually and now it seems to work. Doing some testing now, but at least of hand I can't seem to reproduce the VM fault on a Vega10. Christian. Am 12.03.19 um 16:01 schrieb Russell, Kent: Oh right, I remember that issue. I had that happen to me once, where my installed libhsakmt didn't match up with the latest source code, so I ended up having to remove the libhsakmt package and pointing it to the folders instead. Kent -Original Message- From: Koenig, Christian Sent: Tuesday, March 12, 2019 10:49 AM To: Russell, Kent ; Kuehling, Felix ; amd-gfx@lists.freedesktop.org Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand Yeah, the problem is I do have the libhsakmt installed. Going to give it a try to specify the directory directly. Christian. Am 12.03.19 um 15:47 schrieb Russell, Kent: The README.txt file inside the tests/kfdtest folder has instructions on how to do it if you don't have the libhsakmt package installed on your system: export LIBHSAKMT_PATH=/*your local libhsakmt folder*/ With that, the headers and libraries are searched under LIBHSAKMT_PATH/include and LIBHSAKMT_PATH/lib respectively. So if you try export LIBHSAKMT_PATH as the root ROCT folder (the one containing include, src, tests, etc), then that should cover it. Kent -Original Message- From: Christian König Sent: Tuesday, March 12, 2019 9:13 AM To: Russell, Kent ; Kuehling, Felix ; Koenig, Christian ; amd-gfx@lists.freedesktop.org Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand Hi guys, so found a few minutes today to compile kfdtest. Problem is that during the compile I get a lots of this: CMakeFiles/kfdtest.dir/src/BaseQueue.cpp.o: In Funktion »BaseQueue::Create(unsigned int, unsigned int, unsigned long*)«: /usr/src/ROCT-Thunk-Interface/tests/kfdtest/src/BaseQueue.cpp:57: Warnung: undefinierter Verweis auf »hsaKmtCreateQueue« Any idea? Christian. Am 11.03.19 um 17:55 schrieb Christian König: Hi guys, well it's most likely some missing handling in the KFD, so I'm rather reluctant to revert the change immediately. Problem is that I don't have time right now to look into it immediately. So Kent can you continue to take a look? Sounds like its crashing immediately, so it should be something obvious. Christian. Am 11.03.19 um 10:49 schrieb Russell, Kent: From what I've been able to dig through, the VM Fault seems to occur right after a doorbell mmap, but that's as far as I got. I can try to revert it in today's merge and see how things go. Kent -Original Message- From: Kuehling, Felix Sent: Friday, March 08, 2019 11:16 PM To: Koenig, Christian ; Russell, Kent ; amd-gfx@lists.freedesktop.org Subject: RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand My concerns were related to eviction fence handing. It would manifest by unnecessary eviction callbacks into KFD that aren't cause by real evictions. I addressed that with a previous patch series that removed the need to remove eviction fences and add them back around page table updates in amdgpu_amdkfd_gpuvm.c. I don't know what's going on here. I can probably take a look on Monday. I haven't considered what changed with respect to PD updates. Kent, can we temporarily revert the offending change in amd-kfd-staging just to unblock the merge? Christian, I think KFD is currently broken on amd-staging-drm-next. If we're serious about supporting KFD upstream, you may also want to consider reverting your change there for now. Also consider building the Thunk and kfdtest so you can do quick smoke tests locally whenever you make amdgpu_vm changes that can affect KFD. https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface Regards, Felix -Original Message- From: amd-gfx On Behalf Of Christian König Sent: Friday, March 08, 2019 9:14 AM To: Russell, Kent ; amd-gfx@lists.freedesktop.org Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand My best guess is that we forget somewhere to update the PDs. What hardware is that on? Felix already mentioned that this could be problematic for the KFD. Maybe he has an idea, Christian. Am 08.03.19 um 15:04 schrieb Russell, Kent: Hi Christian, This patch ended up causing a VM Fault in KFDTest. Reverting just this patch addressed the issue: [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 0x480c for process pid 0 thread pid 0 [ 82.703512] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x1000 [ 82.703516] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x1004800C [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid 32769) at page 4096, read from 'TC0' (0x54433000) (72) [ 82.703585] Evicting PASID 32769 queues I am looking into it, but if you have any insight that would be great in helping
RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
Oh right, I remember that issue. I had that happen to me once, where my installed libhsakmt didn't match up with the latest source code, so I ended up having to remove the libhsakmt package and pointing it to the folders instead. Kent > -Original Message- > From: Koenig, Christian > Sent: Tuesday, March 12, 2019 10:49 AM > To: Russell, Kent ; Kuehling, Felix > ; amd-gfx@lists.freedesktop.org > Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > > Yeah, the problem is I do have the libhsakmt installed. > > Going to give it a try to specify the directory directly. > > Christian. > > Am 12.03.19 um 15:47 schrieb Russell, Kent: > > The README.txt file inside the tests/kfdtest folder has instructions on how > to do it if you don't have the libhsakmt package installed on your system: > > > > export LIBHSAKMT_PATH=/*your local libhsakmt folder*/ With that, the > > headers and libraries are searched under LIBHSAKMT_PATH/include and > > LIBHSAKMT_PATH/lib respectively. > > > > So if you try export LIBHSAKMT_PATH as the root ROCT folder (the one > containing include, src, tests, etc), then that should cover it. > > > > Kent > > > > > >> -Original Message- > >> From: Christian König > >> Sent: Tuesday, March 12, 2019 9:13 AM > >> To: Russell, Kent ; Kuehling, Felix > >> ; Koenig, Christian > >> ; amd-gfx@lists.freedesktop.org > >> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > >> > >> Hi guys, > >> > >> so found a few minutes today to compile kfdtest. > >> > >> Problem is that during the compile I get a lots of this: > >>> CMakeFiles/kfdtest.dir/src/BaseQueue.cpp.o: In Funktion > >>> »BaseQueue::Create(unsigned int, unsigned int, unsigned long*)«: > >>> /usr/src/ROCT-Thunk-Interface/tests/kfdtest/src/BaseQueue.cpp:57: > >>> Warnung: undefinierter Verweis auf »hsaKmtCreateQueue« > >> Any idea? > >> > >> Christian. > >> > >> Am 11.03.19 um 17:55 schrieb Christian König: > >>> Hi guys, > >>> > >>> well it's most likely some missing handling in the KFD, so I'm > >>> rather reluctant to revert the change immediately. > >>> > >>> Problem is that I don't have time right now to look into it > >>> immediately. So Kent can you continue to take a look? > >>> > >>> Sounds like its crashing immediately, so it should be something obvious. > >>> > >>> Christian. > >>> > >>> Am 11.03.19 um 10:49 schrieb Russell, Kent: > >>>> From what I've been able to dig through, the VM Fault seems to > >>>> occur right after a doorbell mmap, but that's as far as I got. I > >>>> can try to revert it in today's merge and see how things go. > >>>> > >>>> Kent > >>>> > >>>>> -Original Message- > >>>>> From: Kuehling, Felix > >>>>> Sent: Friday, March 08, 2019 11:16 PM > >>>>> To: Koenig, Christian ; Russell, Kent > >>>>> ; amd-gfx@lists.freedesktop.org > >>>>> Subject: RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on > demand > >>>>> > >>>>> My concerns were related to eviction fence handing. It would > >>>>> manifest by unnecessary eviction callbacks into KFD that aren't > >>>>> cause by real evictions. I addressed that with a previous patch > >>>>> series that removed the need to remove eviction fences and add > >>>>> them back around page table updates in amdgpu_amdkfd_gpuvm.c. > >>>>> > >>>>> I don't know what's going on here. I can probably take a look on > >>>>> Monday. I haven't considered what changed with respect to PD > >>>>> updates. > >>>>> > >>>>> Kent, can we temporarily revert the offending change in > >>>>> amd-kfd-staging just to unblock the merge? > >>>>> > >>>>> Christian, I think KFD is currently broken on amd-staging-drm-next. > >>>>> If we're > >>>>> serious about supporting KFD upstream, you may also want to > >>>>> consider reverting your change there for now. Also consider > >>>>> building the Thunk and kfdtest so you can do quick smoke tests > >>>>> lo
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
Yeah, the problem is I do have the libhsakmt installed. Going to give it a try to specify the directory directly. Christian. Am 12.03.19 um 15:47 schrieb Russell, Kent: > The README.txt file inside the tests/kfdtest folder has instructions on how > to do it if you don't have the libhsakmt package installed on your system: > > export LIBHSAKMT_PATH=/*your local libhsakmt folder*/ > With that, the headers and libraries are searched under > LIBHSAKMT_PATH/include and LIBHSAKMT_PATH/lib respectively. > > So if you try export LIBHSAKMT_PATH as the root ROCT folder (the one > containing include, src, tests, etc), then that should cover it. > > Kent > > >> -Original Message- >> From: Christian König >> Sent: Tuesday, March 12, 2019 9:13 AM >> To: Russell, Kent ; Kuehling, Felix >> ; Koenig, Christian >> ; amd-gfx@lists.freedesktop.org >> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >> >> Hi guys, >> >> so found a few minutes today to compile kfdtest. >> >> Problem is that during the compile I get a lots of this: >>> CMakeFiles/kfdtest.dir/src/BaseQueue.cpp.o: In Funktion >>> »BaseQueue::Create(unsigned int, unsigned int, unsigned long*)«: >>> /usr/src/ROCT-Thunk-Interface/tests/kfdtest/src/BaseQueue.cpp:57: >>> Warnung: undefinierter Verweis auf »hsaKmtCreateQueue« >> Any idea? >> >> Christian. >> >> Am 11.03.19 um 17:55 schrieb Christian König: >>> Hi guys, >>> >>> well it's most likely some missing handling in the KFD, so I'm rather >>> reluctant to revert the change immediately. >>> >>> Problem is that I don't have time right now to look into it >>> immediately. So Kent can you continue to take a look? >>> >>> Sounds like its crashing immediately, so it should be something obvious. >>> >>> Christian. >>> >>> Am 11.03.19 um 10:49 schrieb Russell, Kent: >>>> From what I've been able to dig through, the VM Fault seems to occur >>>> right after a doorbell mmap, but that's as far as I got. I can try to >>>> revert it in today's merge and see how things go. >>>> >>>> Kent >>>> >>>>> -Original Message- >>>>> From: Kuehling, Felix >>>>> Sent: Friday, March 08, 2019 11:16 PM >>>>> To: Koenig, Christian ; Russell, Kent >>>>> ; amd-gfx@lists.freedesktop.org >>>>> Subject: RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>>>> >>>>> My concerns were related to eviction fence handing. It would >>>>> manifest by unnecessary eviction callbacks into KFD that aren't >>>>> cause by real evictions. I addressed that with a previous patch >>>>> series that removed the need to remove eviction fences and add them >>>>> back around page table updates in amdgpu_amdkfd_gpuvm.c. >>>>> >>>>> I don't know what's going on here. I can probably take a look on >>>>> Monday. I haven't considered what changed with respect to PD >>>>> updates. >>>>> >>>>> Kent, can we temporarily revert the offending change in >>>>> amd-kfd-staging just to unblock the merge? >>>>> >>>>> Christian, I think KFD is currently broken on amd-staging-drm-next. >>>>> If we're >>>>> serious about supporting KFD upstream, you may also want to consider >>>>> reverting your change there for now. Also consider building the >>>>> Thunk and kfdtest so you can do quick smoke tests locally whenever >>>>> you make amdgpu_vm changes that can affect KFD. >>>>> https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface >>>>> >>>>> Regards, >>>>> Felix >>>>> >>>>> -Original Message- >>>>> From: amd-gfx On Behalf Of >>>>> Christian König >>>>> Sent: Friday, March 08, 2019 9:14 AM >>>>> To: Russell, Kent ; >>>>> amd-gfx@lists.freedesktop.org >>>>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >>>>> >>>>> My best guess is that we forget somewhere to update the PDs. What >>>>> hardware is that on? >>>>> >>>>> Felix already mentioned that this could be problematic for the KFD. >>>>> >>>>> Maybe he has an
RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
The README.txt file inside the tests/kfdtest folder has instructions on how to do it if you don't have the libhsakmt package installed on your system: export LIBHSAKMT_PATH=/*your local libhsakmt folder*/ With that, the headers and libraries are searched under LIBHSAKMT_PATH/include and LIBHSAKMT_PATH/lib respectively. So if you try export LIBHSAKMT_PATH as the root ROCT folder (the one containing include, src, tests, etc), then that should cover it. Kent > -Original Message- > From: Christian König > Sent: Tuesday, March 12, 2019 9:13 AM > To: Russell, Kent ; Kuehling, Felix > ; Koenig, Christian > ; amd-gfx@lists.freedesktop.org > Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > > Hi guys, > > so found a few minutes today to compile kfdtest. > > Problem is that during the compile I get a lots of this: > > CMakeFiles/kfdtest.dir/src/BaseQueue.cpp.o: In Funktion > > »BaseQueue::Create(unsigned int, unsigned int, unsigned long*)«: > > /usr/src/ROCT-Thunk-Interface/tests/kfdtest/src/BaseQueue.cpp:57: > > Warnung: undefinierter Verweis auf »hsaKmtCreateQueue« > > Any idea? > > Christian. > > Am 11.03.19 um 17:55 schrieb Christian König: > > Hi guys, > > > > well it's most likely some missing handling in the KFD, so I'm rather > > reluctant to revert the change immediately. > > > > Problem is that I don't have time right now to look into it > > immediately. So Kent can you continue to take a look? > > > > Sounds like its crashing immediately, so it should be something obvious. > > > > Christian. > > > > Am 11.03.19 um 10:49 schrieb Russell, Kent: > >> From what I've been able to dig through, the VM Fault seems to occur > >> right after a doorbell mmap, but that's as far as I got. I can try to > >> revert it in today's merge and see how things go. > >> > >> Kent > >> > >>> -Original Message- > >>> From: Kuehling, Felix > >>> Sent: Friday, March 08, 2019 11:16 PM > >>> To: Koenig, Christian ; Russell, Kent > >>> ; amd-gfx@lists.freedesktop.org > >>> Subject: RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > >>> > >>> My concerns were related to eviction fence handing. It would > >>> manifest by unnecessary eviction callbacks into KFD that aren't > >>> cause by real evictions. I addressed that with a previous patch > >>> series that removed the need to remove eviction fences and add them > >>> back around page table updates in amdgpu_amdkfd_gpuvm.c. > >>> > >>> I don't know what's going on here. I can probably take a look on > >>> Monday. I haven't considered what changed with respect to PD > >>> updates. > >>> > >>> Kent, can we temporarily revert the offending change in > >>> amd-kfd-staging just to unblock the merge? > >>> > >>> Christian, I think KFD is currently broken on amd-staging-drm-next. > >>> If we're > >>> serious about supporting KFD upstream, you may also want to consider > >>> reverting your change there for now. Also consider building the > >>> Thunk and kfdtest so you can do quick smoke tests locally whenever > >>> you make amdgpu_vm changes that can affect KFD. > >>> https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface > >>> > >>> Regards, > >>> Felix > >>> > >>> -Original Message- > >>> From: amd-gfx On Behalf Of > >>> Christian König > >>> Sent: Friday, March 08, 2019 9:14 AM > >>> To: Russell, Kent ; > >>> amd-gfx@lists.freedesktop.org > >>> Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > >>> > >>> My best guess is that we forget somewhere to update the PDs. What > >>> hardware is that on? > >>> > >>> Felix already mentioned that this could be problematic for the KFD. > >>> > >>> Maybe he has an idea, > >>> Christian. > >>> > >>> Am 08.03.19 um 15:04 schrieb Russell, Kent: > >>>> Hi Christian, > >>>> > >>>> This patch ended up causing a VM Fault in KFDTest. Reverting just > >>>> this > >>> patch addressed the issue: > >>>> [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 > >>>> 0x480c for > >>> process pid 0 thread pid 0 > >
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
Hi guys, so found a few minutes today to compile kfdtest. Problem is that during the compile I get a lots of this: CMakeFiles/kfdtest.dir/src/BaseQueue.cpp.o: In Funktion »BaseQueue::Create(unsigned int, unsigned int, unsigned long*)«: /usr/src/ROCT-Thunk-Interface/tests/kfdtest/src/BaseQueue.cpp:57: Warnung: undefinierter Verweis auf »hsaKmtCreateQueue« Any idea? Christian. Am 11.03.19 um 17:55 schrieb Christian König: Hi guys, well it's most likely some missing handling in the KFD, so I'm rather reluctant to revert the change immediately. Problem is that I don't have time right now to look into it immediately. So Kent can you continue to take a look? Sounds like its crashing immediately, so it should be something obvious. Christian. Am 11.03.19 um 10:49 schrieb Russell, Kent: From what I've been able to dig through, the VM Fault seems to occur right after a doorbell mmap, but that's as far as I got. I can try to revert it in today's merge and see how things go. Kent -Original Message- From: Kuehling, Felix Sent: Friday, March 08, 2019 11:16 PM To: Koenig, Christian ; Russell, Kent ; amd-gfx@lists.freedesktop.org Subject: RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand My concerns were related to eviction fence handing. It would manifest by unnecessary eviction callbacks into KFD that aren't cause by real evictions. I addressed that with a previous patch series that removed the need to remove eviction fences and add them back around page table updates in amdgpu_amdkfd_gpuvm.c. I don't know what's going on here. I can probably take a look on Monday. I haven't considered what changed with respect to PD updates. Kent, can we temporarily revert the offending change in amd-kfd-staging just to unblock the merge? Christian, I think KFD is currently broken on amd-staging-drm-next. If we're serious about supporting KFD upstream, you may also want to consider reverting your change there for now. Also consider building the Thunk and kfdtest so you can do quick smoke tests locally whenever you make amdgpu_vm changes that can affect KFD. https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface Regards, Felix -Original Message- From: amd-gfx On Behalf Of Christian König Sent: Friday, March 08, 2019 9:14 AM To: Russell, Kent ; amd-gfx@lists.freedesktop.org Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand My best guess is that we forget somewhere to update the PDs. What hardware is that on? Felix already mentioned that this could be problematic for the KFD. Maybe he has an idea, Christian. Am 08.03.19 um 15:04 schrieb Russell, Kent: Hi Christian, This patch ended up causing a VM Fault in KFDTest. Reverting just this patch addressed the issue: [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 0x480c for process pid 0 thread pid 0 [ 82.703512] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x1000 [ 82.703516] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x1004800C [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid 32769) at page 4096, read from 'TC0' (0x54433000) (72) [ 82.703585] Evicting PASID 32769 queues I am looking into it, but if you have any insight that would be great in helping to resolve it quickly. Kent -Original Message- From: amd-gfx On Behalf Of Christian König Sent: Tuesday, February 26, 2019 7:47 AM To: amd-gfx@lists.freedesktop.org Subject: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand Let's start to allocate VM PDs/PTs on demand instead of pre-allocating them during mapping. Signed-off-by: Christian König Reviewed-by: Felix Kuehling --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 9 -- drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 -- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 136 + - drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 3 - 5 files changed, 39 insertions(+), 129 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index 31e3953dcb6e..088e9b6b765b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -410,15 +410,7 @@ static int add_bo_to_vm(struct amdgpu_device *adev, struct kgd_mem *mem, if (p_bo_va_entry) *p_bo_va_entry = bo_va_entry; - /* Allocate new page tables if needed and validate - * them. - */ - ret = amdgpu_vm_alloc_pts(adev, vm, va, amdgpu_bo_size(bo)); - if (ret) { - pr_err("Failed to allocate pts, err=%d\n", ret); - goto err_alloc_pts; - } - + /* Allocate validate page tables if needed */ ret = vm_validate_pt_pd_bos(vm); if (ret) { pr_err("validate_pt_pd_bos() failed\n
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
Hi guys, well it's most likely some missing handling in the KFD, so I'm rather reluctant to revert the change immediately. Problem is that I don't have time right now to look into it immediately. So Kent can you continue to take a look? Sounds like its crashing immediately, so it should be something obvious. Christian. Am 11.03.19 um 10:49 schrieb Russell, Kent: From what I've been able to dig through, the VM Fault seems to occur right after a doorbell mmap, but that's as far as I got. I can try to revert it in today's merge and see how things go. Kent -Original Message- From: Kuehling, Felix Sent: Friday, March 08, 2019 11:16 PM To: Koenig, Christian ; Russell, Kent ; amd-gfx@lists.freedesktop.org Subject: RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand My concerns were related to eviction fence handing. It would manifest by unnecessary eviction callbacks into KFD that aren't cause by real evictions. I addressed that with a previous patch series that removed the need to remove eviction fences and add them back around page table updates in amdgpu_amdkfd_gpuvm.c. I don't know what's going on here. I can probably take a look on Monday. I haven't considered what changed with respect to PD updates. Kent, can we temporarily revert the offending change in amd-kfd-staging just to unblock the merge? Christian, I think KFD is currently broken on amd-staging-drm-next. If we're serious about supporting KFD upstream, you may also want to consider reverting your change there for now. Also consider building the Thunk and kfdtest so you can do quick smoke tests locally whenever you make amdgpu_vm changes that can affect KFD. https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface Regards, Felix -Original Message- From: amd-gfx On Behalf Of Christian König Sent: Friday, March 08, 2019 9:14 AM To: Russell, Kent ; amd-gfx@lists.freedesktop.org Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand My best guess is that we forget somewhere to update the PDs. What hardware is that on? Felix already mentioned that this could be problematic for the KFD. Maybe he has an idea, Christian. Am 08.03.19 um 15:04 schrieb Russell, Kent: Hi Christian, This patch ended up causing a VM Fault in KFDTest. Reverting just this patch addressed the issue: [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 0x480c for process pid 0 thread pid 0 [ 82.703512] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x1000 [ 82.703516] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x1004800C [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid 32769) at page 4096, read from 'TC0' (0x54433000) (72) [ 82.703585] Evicting PASID 32769 queues I am looking into it, but if you have any insight that would be great in helping to resolve it quickly. Kent -Original Message- From: amd-gfx On Behalf Of Christian König Sent: Tuesday, February 26, 2019 7:47 AM To: amd-gfx@lists.freedesktop.org Subject: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand Let's start to allocate VM PDs/PTs on demand instead of pre-allocating them during mapping. Signed-off-by: Christian König Reviewed-by: Felix Kuehling --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 9 -- drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 -- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c| 136 + - drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h| 3 - 5 files changed, 39 insertions(+), 129 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index 31e3953dcb6e..088e9b6b765b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -410,15 +410,7 @@ static int add_bo_to_vm(struct amdgpu_device *adev, struct kgd_mem *mem, if (p_bo_va_entry) *p_bo_va_entry = bo_va_entry; - /* Allocate new page tables if needed and validate -* them. -*/ - ret = amdgpu_vm_alloc_pts(adev, vm, va, amdgpu_bo_size(bo)); - if (ret) { - pr_err("Failed to allocate pts, err=%d\n", ret); - goto err_alloc_pts; - } - + /* Allocate validate page tables if needed */ ret = vm_validate_pt_pd_bos(vm); if (ret) { pr_err("validate_pt_pd_bos() failed\n"); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c index 7e22be7ca68a..54dd02a898b9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c @@ -92,15 +92,6 @@ int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm, return -ENOMEM; } - r = amdgpu_vm_allo
RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
From what I've been able to dig through, the VM Fault seems to occur right after a doorbell mmap, but that's as far as I got. I can try to revert it in today's merge and see how things go. Kent > -Original Message- > From: Kuehling, Felix > Sent: Friday, March 08, 2019 11:16 PM > To: Koenig, Christian ; Russell, Kent > ; amd-gfx@lists.freedesktop.org > Subject: RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > > My concerns were related to eviction fence handing. It would manifest by > unnecessary eviction callbacks into KFD that aren't cause by real evictions. I > addressed that with a previous patch series that removed the need to > remove eviction fences and add them back around page table updates in > amdgpu_amdkfd_gpuvm.c. > > I don't know what's going on here. I can probably take a look on Monday. I > haven't considered what changed with respect to PD updates. > > Kent, can we temporarily revert the offending change in amd-kfd-staging > just to unblock the merge? > > Christian, I think KFD is currently broken on amd-staging-drm-next. If we're > serious about supporting KFD upstream, you may also want to consider > reverting your change there for now. Also consider building the Thunk and > kfdtest so you can do quick smoke tests locally whenever you make > amdgpu_vm changes that can affect KFD. > https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface > > Regards, > Felix > > -Original Message- > From: amd-gfx On Behalf Of > Christian König > Sent: Friday, March 08, 2019 9:14 AM > To: Russell, Kent ; amd-gfx@lists.freedesktop.org > Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > > My best guess is that we forget somewhere to update the PDs. What > hardware is that on? > > Felix already mentioned that this could be problematic for the KFD. > > Maybe he has an idea, > Christian. > > Am 08.03.19 um 15:04 schrieb Russell, Kent: > > Hi Christian, > > > > This patch ended up causing a VM Fault in KFDTest. Reverting just this > patch addressed the issue: > > [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 0x480c for > process pid 0 thread pid 0 > > [ 82.703512] amdgpu :0c:00.0: > VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x1000 > > [ 82.703516] amdgpu :0c:00.0: > VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x1004800C > > [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid 32769) at > page 4096, read from 'TC0' (0x54433000) (72) > > [ 82.703585] Evicting PASID 32769 queues > > > > I am looking into it, but if you have any insight that would be great in > helping to resolve it quickly. > > > > Kent > >> -Original Message- > >> From: amd-gfx On Behalf Of > >> Christian König > >> Sent: Tuesday, February 26, 2019 7:47 AM > >> To: amd-gfx@lists.freedesktop.org > >> Subject: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > >> > >> Let's start to allocate VM PDs/PTs on demand instead of > >> pre-allocating them during mapping. > >> > >> Signed-off-by: Christian König > >> Reviewed-by: Felix Kuehling > >> --- > >> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 10 +- > >> drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 9 -- > >> drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 -- > >> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c| 136 + > - > >> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h| 3 - > >> 5 files changed, 39 insertions(+), 129 deletions(-) > >> > >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > >> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > >> index 31e3953dcb6e..088e9b6b765b 100644 > >> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > >> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > >> @@ -410,15 +410,7 @@ static int add_bo_to_vm(struct amdgpu_device > >> *adev, struct kgd_mem *mem, > >>if (p_bo_va_entry) > >>*p_bo_va_entry = bo_va_entry; > >> > >> - /* Allocate new page tables if needed and validate > >> - * them. > >> - */ > >> - ret = amdgpu_vm_alloc_pts(adev, vm, va, amdgpu_bo_size(bo)); > >> - if (ret) { > >> - pr_err("Failed to allocate pts, err=%d\n", ret); > >> - goto err_alloc_pts; > >> - } > >> - > >> + /* Allocate validate page tables if needed */ > >>ret = vm_validate_pt_pd_bos(vm); >
RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
My concerns were related to eviction fence handing. It would manifest by unnecessary eviction callbacks into KFD that aren't cause by real evictions. I addressed that with a previous patch series that removed the need to remove eviction fences and add them back around page table updates in amdgpu_amdkfd_gpuvm.c. I don't know what's going on here. I can probably take a look on Monday. I haven't considered what changed with respect to PD updates. Kent, can we temporarily revert the offending change in amd-kfd-staging just to unblock the merge? Christian, I think KFD is currently broken on amd-staging-drm-next. If we're serious about supporting KFD upstream, you may also want to consider reverting your change there for now. Also consider building the Thunk and kfdtest so you can do quick smoke tests locally whenever you make amdgpu_vm changes that can affect KFD. https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface Regards, Felix -Original Message- From: amd-gfx On Behalf Of Christian König Sent: Friday, March 08, 2019 9:14 AM To: Russell, Kent ; amd-gfx@lists.freedesktop.org Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand My best guess is that we forget somewhere to update the PDs. What hardware is that on? Felix already mentioned that this could be problematic for the KFD. Maybe he has an idea, Christian. Am 08.03.19 um 15:04 schrieb Russell, Kent: > Hi Christian, > > This patch ended up causing a VM Fault in KFDTest. Reverting just this patch > addressed the issue: > [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 0x480c for > process pid 0 thread pid 0 > [ 82.703512] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR > 0x1000 > [ 82.703516] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS > 0x1004800C > [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid 32769) at > page 4096, read from 'TC0' (0x54433000) (72) > [ 82.703585] Evicting PASID 32769 queues > > I am looking into it, but if you have any insight that would be great in > helping to resolve it quickly. > > Kent >> -Original Message- >> From: amd-gfx On Behalf Of >> Christian König >> Sent: Tuesday, February 26, 2019 7:47 AM >> To: amd-gfx@lists.freedesktop.org >> Subject: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand >> >> Let's start to allocate VM PDs/PTs on demand instead of >> pre-allocating them during mapping. >> >> Signed-off-by: Christian König >> Reviewed-by: Felix Kuehling >> --- >> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 10 +- >> drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 9 -- >> drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 -- >> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c| 136 +- >> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h| 3 - >> 5 files changed, 39 insertions(+), 129 deletions(-) >> >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c >> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c >> index 31e3953dcb6e..088e9b6b765b 100644 >> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c >> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c >> @@ -410,15 +410,7 @@ static int add_bo_to_vm(struct amdgpu_device >> *adev, struct kgd_mem *mem, >> if (p_bo_va_entry) >> *p_bo_va_entry = bo_va_entry; >> >> -/* Allocate new page tables if needed and validate >> - * them. >> - */ >> -ret = amdgpu_vm_alloc_pts(adev, vm, va, amdgpu_bo_size(bo)); >> -if (ret) { >> -pr_err("Failed to allocate pts, err=%d\n", ret); >> -goto err_alloc_pts; >> -} >> - >> +/* Allocate validate page tables if needed */ >> ret = vm_validate_pt_pd_bos(vm); >> if (ret) { >> pr_err("validate_pt_pd_bos() failed\n"); diff --git >> a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c >> b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c >> index 7e22be7ca68a..54dd02a898b9 100644 >> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c >> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c >> @@ -92,15 +92,6 @@ int amdgpu_map_static_csa(struct amdgpu_device >> *adev, struct amdgpu_vm *vm, >> return -ENOMEM; >> } >> >> -r = amdgpu_vm_alloc_pts(adev, (*bo_va)->base.vm, csa_addr, >> -size); >> -if (r) { >> -DRM_ERROR("failed to allocate pts for static CSA, err=%d\n", >> r); >> -amdgpu_vm_bo_rmv(adev, *bo_va); >> -ttm_eu_backoff_res
RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
That dmesg was from Fiji, but it occurred on Vega10 as well. Kent > -Original Message- > From: Christian König > Sent: Friday, March 08, 2019 9:14 AM > To: Russell, Kent ; amd-gfx@lists.freedesktop.org > Subject: Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > > My best guess is that we forget somewhere to update the PDs. What > hardware is that on? > > Felix already mentioned that this could be problematic for the KFD. > > Maybe he has an idea, > Christian. > > Am 08.03.19 um 15:04 schrieb Russell, Kent: > > Hi Christian, > > > > This patch ended up causing a VM Fault in KFDTest. Reverting just this > patch addressed the issue: > > [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 0x480c for > process pid 0 thread pid 0 > > [ 82.703512] amdgpu :0c:00.0: > VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x1000 > > [ 82.703516] amdgpu :0c:00.0: > VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x1004800C > > [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid 32769) at > page 4096, read from 'TC0' (0x54433000) (72) > > [ 82.703585] Evicting PASID 32769 queues > > > > I am looking into it, but if you have any insight that would be great in > helping to resolve it quickly. > > > > Kent > >> -Original Message- > >> From: amd-gfx On Behalf Of > >> Christian König > >> Sent: Tuesday, February 26, 2019 7:47 AM > >> To: amd-gfx@lists.freedesktop.org > >> Subject: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > >> > >> Let's start to allocate VM PDs/PTs on demand instead of > >> pre-allocating them during mapping. > >> > >> Signed-off-by: Christian König > >> Reviewed-by: Felix Kuehling > >> --- > >> .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 10 +- > >> drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 9 -- > >> drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 -- > >> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c| 136 + > - > >> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h| 3 - > >> 5 files changed, 39 insertions(+), 129 deletions(-) > >> > >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > >> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > >> index 31e3953dcb6e..088e9b6b765b 100644 > >> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > >> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > >> @@ -410,15 +410,7 @@ static int add_bo_to_vm(struct amdgpu_device > >> *adev, struct kgd_mem *mem, > >>if (p_bo_va_entry) > >>*p_bo_va_entry = bo_va_entry; > >> > >> - /* Allocate new page tables if needed and validate > >> - * them. > >> - */ > >> - ret = amdgpu_vm_alloc_pts(adev, vm, va, amdgpu_bo_size(bo)); > >> - if (ret) { > >> - pr_err("Failed to allocate pts, err=%d\n", ret); > >> - goto err_alloc_pts; > >> - } > >> - > >> + /* Allocate validate page tables if needed */ > >>ret = vm_validate_pt_pd_bos(vm); > >>if (ret) { > >>pr_err("validate_pt_pd_bos() failed\n"); diff --git > >> a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c > >> b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c > >> index 7e22be7ca68a..54dd02a898b9 100644 > >> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c > >> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c > >> @@ -92,15 +92,6 @@ int amdgpu_map_static_csa(struct amdgpu_device > >> *adev, struct amdgpu_vm *vm, > >>return -ENOMEM; > >>} > >> > >> - r = amdgpu_vm_alloc_pts(adev, (*bo_va)->base.vm, csa_addr, > >> - size); > >> - if (r) { > >> - DRM_ERROR("failed to allocate pts for static CSA, err=%d\n", > >> r); > >> - amdgpu_vm_bo_rmv(adev, *bo_va); > >> - ttm_eu_backoff_reservation(&ticket, &list); > >> - return r; > >> - } > >> - > >>r = amdgpu_vm_bo_map(adev, *bo_va, csa_addr, 0, size, > >> AMDGPU_PTE_READABLE | > >> AMDGPU_PTE_WRITEABLE | > >> AMDGPU_PTE_EXECUTABLE); > >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > >> b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > >> index 555285e329ed..fcaaac30e84b 100644 > >> --- a/drivers/
Re: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
My best guess is that we forget somewhere to update the PDs. What hardware is that on? Felix already mentioned that this could be problematic for the KFD. Maybe he has an idea, Christian. Am 08.03.19 um 15:04 schrieb Russell, Kent: Hi Christian, This patch ended up causing a VM Fault in KFDTest. Reverting just this patch addressed the issue: [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 0x480c for process pid 0 thread pid 0 [ 82.703512] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x1000 [ 82.703516] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x1004800C [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid 32769) at page 4096, read from 'TC0' (0x54433000) (72) [ 82.703585] Evicting PASID 32769 queues I am looking into it, but if you have any insight that would be great in helping to resolve it quickly. Kent -Original Message- From: amd-gfx On Behalf Of Christian König Sent: Tuesday, February 26, 2019 7:47 AM To: amd-gfx@lists.freedesktop.org Subject: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand Let's start to allocate VM PDs/PTs on demand instead of pre-allocating them during mapping. Signed-off-by: Christian König Reviewed-by: Felix Kuehling --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 9 -- drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 -- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c| 136 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h| 3 - 5 files changed, 39 insertions(+), 129 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index 31e3953dcb6e..088e9b6b765b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -410,15 +410,7 @@ static int add_bo_to_vm(struct amdgpu_device *adev, struct kgd_mem *mem, if (p_bo_va_entry) *p_bo_va_entry = bo_va_entry; - /* Allocate new page tables if needed and validate -* them. -*/ - ret = amdgpu_vm_alloc_pts(adev, vm, va, amdgpu_bo_size(bo)); - if (ret) { - pr_err("Failed to allocate pts, err=%d\n", ret); - goto err_alloc_pts; - } - + /* Allocate validate page tables if needed */ ret = vm_validate_pt_pd_bos(vm); if (ret) { pr_err("validate_pt_pd_bos() failed\n"); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c index 7e22be7ca68a..54dd02a898b9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c @@ -92,15 +92,6 @@ int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm, return -ENOMEM; } - r = amdgpu_vm_alloc_pts(adev, (*bo_va)->base.vm, csa_addr, - size); - if (r) { - DRM_ERROR("failed to allocate pts for static CSA, err=%d\n", r); - amdgpu_vm_bo_rmv(adev, *bo_va); - ttm_eu_backoff_reservation(&ticket, &list); - return r; - } - r = amdgpu_vm_bo_map(adev, *bo_va, csa_addr, 0, size, AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE | AMDGPU_PTE_EXECUTABLE); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c index 555285e329ed..fcaaac30e84b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c @@ -625,11 +625,6 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, switch (args->operation) { case AMDGPU_VA_OP_MAP: - r = amdgpu_vm_alloc_pts(adev, bo_va->base.vm, args- va_address, - args->map_size); - if (r) - goto error_backoff; - va_flags = amdgpu_gmc_get_pte_flags(adev, args->flags); r = amdgpu_vm_bo_map(adev, bo_va, args->va_address, args->offset_in_bo, args->map_size, @@ - 645,11 +640,6 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data, args->map_size); break; case AMDGPU_VA_OP_REPLACE: - r = amdgpu_vm_alloc_pts(adev, bo_va->base.vm, args- va_address, - args->map_size); - if (r) - goto error_backoff; - va_flags = amdgpu_gmc_get_pte_flags(adev, args->flags); r = amdgpu_vm_bo_replace_map(adev, bo_va, args- va_address, args->offset_in_bo, args- map_size, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 362436f4e856..dfad543fc000 1006
RE: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand
Hi Christian, This patch ended up causing a VM Fault in KFDTest. Reverting just this patch addressed the issue: [ 82.703503] amdgpu :0c:00.0: GPU fault detected: 146 0x480c for process pid 0 thread pid 0 [ 82.703512] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x1000 [ 82.703516] amdgpu :0c:00.0: VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x1004800C [ 82.703522] amdgpu :0c:00.0: VM fault (0x0c, vmid 8, pasid 32769) at page 4096, read from 'TC0' (0x54433000) (72) [ 82.703585] Evicting PASID 32769 queues I am looking into it, but if you have any insight that would be great in helping to resolve it quickly. Kent > -Original Message- > From: amd-gfx On Behalf Of > Christian König > Sent: Tuesday, February 26, 2019 7:47 AM > To: amd-gfx@lists.freedesktop.org > Subject: [PATCH 3/6] drm/amdgpu: allocate VM PDs/PTs on demand > > Let's start to allocate VM PDs/PTs on demand instead of pre-allocating them > during mapping. > > Signed-off-by: Christian König > Reviewed-by: Felix Kuehling > --- > .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 10 +- > drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 9 -- > drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 10 -- > drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c| 136 +- > drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h| 3 - > 5 files changed, 39 insertions(+), 129 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > index 31e3953dcb6e..088e9b6b765b 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c > @@ -410,15 +410,7 @@ static int add_bo_to_vm(struct amdgpu_device > *adev, struct kgd_mem *mem, > if (p_bo_va_entry) > *p_bo_va_entry = bo_va_entry; > > - /* Allocate new page tables if needed and validate > - * them. > - */ > - ret = amdgpu_vm_alloc_pts(adev, vm, va, amdgpu_bo_size(bo)); > - if (ret) { > - pr_err("Failed to allocate pts, err=%d\n", ret); > - goto err_alloc_pts; > - } > - > + /* Allocate validate page tables if needed */ > ret = vm_validate_pt_pd_bos(vm); > if (ret) { > pr_err("validate_pt_pd_bos() failed\n"); diff --git > a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c > b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c > index 7e22be7ca68a..54dd02a898b9 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c > @@ -92,15 +92,6 @@ int amdgpu_map_static_csa(struct amdgpu_device > *adev, struct amdgpu_vm *vm, > return -ENOMEM; > } > > - r = amdgpu_vm_alloc_pts(adev, (*bo_va)->base.vm, csa_addr, > - size); > - if (r) { > - DRM_ERROR("failed to allocate pts for static CSA, err=%d\n", > r); > - amdgpu_vm_bo_rmv(adev, *bo_va); > - ttm_eu_backoff_reservation(&ticket, &list); > - return r; > - } > - > r = amdgpu_vm_bo_map(adev, *bo_va, csa_addr, 0, size, >AMDGPU_PTE_READABLE | > AMDGPU_PTE_WRITEABLE | >AMDGPU_PTE_EXECUTABLE); > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > index 555285e329ed..fcaaac30e84b 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c > @@ -625,11 +625,6 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, > void *data, > > switch (args->operation) { > case AMDGPU_VA_OP_MAP: > - r = amdgpu_vm_alloc_pts(adev, bo_va->base.vm, args- > >va_address, > - args->map_size); > - if (r) > - goto error_backoff; > - > va_flags = amdgpu_gmc_get_pte_flags(adev, args->flags); > r = amdgpu_vm_bo_map(adev, bo_va, args->va_address, >args->offset_in_bo, args->map_size, @@ - > 645,11 +640,6 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void > *data, > args->map_size); > break; > case AMDGPU_VA_OP_REPLACE: > - r = amdgpu_vm_alloc_pts(adev, bo_va->base.vm, args- > >va_address, > - args->map_size); > - if (r) > - goto error_backoff; > - > va_flags = amdgpu_gmc_get_pte_flags(adev, args->flags); > r = amdgpu_vm_bo_replace_map(adev, bo_va, args- > >va_address, >args->offset_in_bo, args- > >map_size, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c > b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c > index 362436f4e856..dfad543fc000 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c > @@ -504,47 +504,6 @@ static void am