[PATCH v5 00/27] RFC Support hot device unplug in amdgpu
Andrey Grodzovsky
andrey.grodzovsky at amd.com
Wed Apr 28 15:11:40 UTC 2021
Until now extracting a card either by physical extraction (e.g. eGPU with
thunderbolt connection or by emulation through syfs -> /sys/bus/pci/devices/device_id/remove)
would cause random crashes in user apps. The random crashes in apps were
mostly due to the app having mapped a device backed BO into its address
space was still trying to access the BO while the backing device was gone.
To answer this first problem Christian suggested to fix the handling of mapped
memory in the clients when the device goes away by forcibly unmap all buffers the
user processes has by clearing their respective VMAs mapping the device BOs.
Then when the VMAs try to fill in the page tables again we check in the fault
handlerif the device is removed and if so, return an error. This will generate a
SIGBUS to the application which can then cleanly terminate.This indeed was done
but this in turn created a problem of kernel OOPs were the OOPSes were due to the
fact that while the app was terminating because of the SIGBUSit would trigger use
after free in the driver by calling to accesses device structures that were already
released from the pci remove sequence.This was handled by introducing a 'flush'
sequence during device removal were we wait for drm file reference to drop to 0
meaning all user clients directly using this device terminated.
v2:
Based on discussions in the mailing list with Daniel and Pekka [1] and based on the document
produced by Pekka from those discussions [2] the whole approach with returning SIGBUS and
waiting for all user clients having CPU mapping of device BOs to die was dropped.
Instead as per the document suggestion the device structures are kept alive until
the last reference to the device is dropped by user client and in the meanwhile all existing and new CPU mappings of the BOs
belonging to the device directly or by dma-buf import are rerouted to per user
process dummy rw page.Also, I skipped the 'Requirements for KMS UAPI' section of [2]
since i am trying to get the minimal set of requirements that still give useful solution
to work and this is the'Requirements for Render and Cross-Device UAPI' section and so my
test case is removing a secondary device, which is render only and is not involved
in KMS.
v3:
More updates following comments from v2 such as removing loop to find DRM file when rerouting
page faults to dummy page,getting rid of unnecessary sysfs handling refactoring and moving
prevention of GPU recovery post device unplug from amdgpu to scheduler layer.
On top of that added unplug support for the IOMMU enabled system.
v4:
Drop last sysfs hack and use sysfs default attribute.
Guard against write accesses after device removal to avoid modifying released memory.
Update dummy pages handling to on demand allocation and release through drm managed framework.
Add return value to scheduler job TO handler (by Luben Tuikov) and use this in amdgpu for prevention
of GPU recovery post device unplug
Also rebase on top of drm-misc-mext instead of amd-staging-drm-next
v5:
The most significant in this series is the improved protection from kernel driver accessing MMIO ranges that were allocated
for the device once the device is gone. To do this, first a patch 'drm/amdgpu: Unmap all MMIO mappings' is introduced.
This patch unamps all MMIO mapped into the kernel address space in the form of BARs and kernel BOs with CPU visible VRAM mappings.
This way it helped to discover multiple such access points because a page fault would be immediately generated on access. Most of them
were solved by moving HW fini code into pci_remove stage (patch drm/amdgpu: Add early fini callback) and for some who
were harder to unwind drm_dev_enter/exit scoping was used. In addition all the IOCTLs and all background work and timers
are now protected with drm_dev_enter/exit at their root in an attempt that after drm_dev_unplug is finished none of them
run anymore and the pci_remove thread is the only thread executing which might touch the HW. To prevent deadlocks in such
case against threads stuck on various HW or SW fences patches 'drm/amdgpu: Finalise device fences on device remove'
and drm/amdgpu: Add rw_sem to pushing job into sched queue' take care of force signaling all such existing fences
and rejecting any newly added ones.
With these patches I am able to gracefully remove the secondary card using sysfs remove hook while glxgears is running off of secondary
card (DRI_PRIME=1) without kernel oopses or hangs and keep working with the primary card or soft reset the device without hangs or oopses.
Also as per Daniel's comment I added 3 tests to IGT [4] to core_hotunplug test suite - remove device while commands are submitted,
exported BO and exported fence (not pushed yet).
Also now it's possible to plug back the device after unplug
Also some users now can successfully use those patches with eGPU boxes[3].
TODOs for followup work:
Convert AMDGPU code to use devm (for hw stuff) and drmm (for sw stuff and allocations) (Daniel)
Add support for 'Requirements for KMS UAPI' section of [2] - unplugging primary, display connected card.
[1] - Discussions during v4 of the patchset https://lists.freedesktop.org/archives/amd-gfx/2021-January/058595.html
[2] - drm/doc: device hot-unplug for userspace https://www.spinics.net/lists/dri-devel/msg259755.html
[3] - Related gitlab ticket https://gitlab.freedesktop.org/drm/amd/-/issues/1081
[4] - https://gitlab.freedesktop.org/agrodzov/igt-gpu-tools/-/commits/master
Andrey Grodzovsky (27):
drm/ttm: Remap all page faults to per process dummy page.
drm/ttm: Expose ttm_tt_unpopulate for driver use
drm/amdgpu: Split amdgpu_device_fini into early and late
drm/amdkfd: Split kfd suspend from devie exit
drm/amdgpu: Add early fini callback
drm/amdgpu: Handle IOMMU enabled case.
drm/amdgpu: Remap all page faults to per process dummy page.
PCI: add support for dev_groups to struct pci_device_driver
dmr/amdgpu: Move some sysfs attrs creation to default_attr
drm/amdgpu: Guard against write accesses after device removal
drm/sched: Make timeout timer rearm conditional.
drm/amdgpu: Prevent any job recoveries after device is unplugged.
drm/amdgpu: When filizing the fence driver. stop scheduler first.
drm/amdgpu: Fix hang on device removal.
drm/scheduler: Fix hang when sched_entity released
drm/amdgpu: Unmap all MMIO mappings
drm/amdgpu: Add rw_sem to pushing job into sched queue
drm/sched: Expose drm_sched_entity_kill_jobs
drm/amdgpu: Finilise device fences on device remove.
drm: Scope all DRM IOCTLs with drm_dev_enter/exit
drm/amdgpu: Add support for hot-unplug feature at DRM level.
drm/amd/display: Scope all DM queued work with drm_dev_enter/exit
drm/amd/powerplay: Scope all PM queued work with drm_dev_enter/exit
drm/amdkfd: Scope all KFD queued work with drm_dev_enter/exit
drm/amdgpu: Scope all amdgpu queued work with drm_dev_enter/exit
drm/amd/display: Remove superflous drm_mode_config_cleanup
drm/amdgpu: Verify DMA opearations from device are done
drivers/gpu/drm/amd/amdgpu/amdgpu.h | 18 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 13 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_atombios.c | 17 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 13 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 353 ++++++++++++++----
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 34 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 34 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c | 3 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_gart.h | 1 +
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c | 9 +
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c | 25 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 228 +++++------
drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c | 61 ++-
drivers/gpu/drm/amd/amdgpu/amdgpu_irq.h | 3 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 33 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c | 28 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 12 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 41 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_object.h | 7 +
drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c | 115 +++---
drivers/gpu/drm/amd/amdgpu/amdgpu_psp.h | 3 +
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 56 ++-
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 70 ++++
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 52 +--
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 21 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c | 74 ++--
drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c | 45 ++-
drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 83 ++--
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 7 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 14 +-
drivers/gpu/drm/amd/amdgpu/cik_ih.c | 3 +-
drivers/gpu/drm/amd/amdgpu/cz_ih.c | 3 +-
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 10 +-
drivers/gpu/drm/amd/amdgpu/iceland_ih.c | 3 +-
drivers/gpu/drm/amd/amdgpu/navi10_ih.c | 5 +-
drivers/gpu/drm/amd/amdgpu/psp_v11_0.c | 44 +--
drivers/gpu/drm/amd/amdgpu/psp_v12_0.c | 8 +-
drivers/gpu/drm/amd/amdgpu/psp_v3_1.c | 8 +-
drivers/gpu/drm/amd/amdgpu/si_ih.c | 3 +-
drivers/gpu/drm/amd/amdgpu/tonga_ih.c | 3 +-
drivers/gpu/drm/amd/amdgpu/vce_v4_0.c | 26 +-
drivers/gpu/drm/amd/amdgpu/vcn_v3_0.c | 22 +-
drivers/gpu/drm/amd/amdgpu/vega10_ih.c | 5 +-
drivers/gpu/drm/amd/amdgpu/vega20_ih.c | 2 +-
drivers/gpu/drm/amd/amdkfd/kfd_device.c | 3 +-
drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c | 14 +-
.../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 13 +-
.../amd/display/amdgpu_dm/amdgpu_dm_hdcp.c | 124 +++---
.../drm/amd/display/amdgpu_dm/amdgpu_dm_irq.c | 24 +-
drivers/gpu/drm/amd/include/amd_shared.h | 2 +
drivers/gpu/drm/amd/pm/amdgpu_dpm.c | 44 ++-
.../drm/amd/pm/powerplay/smumgr/smu7_smumgr.c | 2 +
drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c | 26 +-
drivers/gpu/drm/drm_ioctl.c | 15 +-
drivers/gpu/drm/scheduler/sched_entity.c | 6 +-
drivers/gpu/drm/scheduler/sched_main.c | 35 +-
drivers/gpu/drm/ttm/ttm_bo_vm.c | 79 +++-
drivers/gpu/drm/ttm/ttm_tt.c | 1 +
drivers/pci/pci-driver.c | 1 +
include/drm/drm_drv.h | 6 +
include/drm/gpu_scheduler.h | 1 +
include/drm/ttm/ttm_bo_api.h | 2 +
include/linux/pci.h | 3 +
64 files changed, 1388 insertions(+), 633 deletions(-)
--
2.25.1
More information about the dri-devel
mailing list