New flag to disable an idle runlist optimization that is causing soft
hangs with some diffult-to-reproduce customer workloads. This will
serve as a workaround until the problem can be reproduced and the
root-cause determined.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Restore the code that resets mem->invalid. Othewise so mapping
userptrs after they got an MMU notifiers would always be skipped.
This also avoids unnecessarily calling get_user_pages on BOs that
have not been invalidated since the last try.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We don't want eviction fences to trigger when waiting for page table
updates to complete during restore. In theory there shouldn't be any
unsignaled eviction fences in the PD reservation object, but I'm
seeing them in instrumented code for reasons not fully understood.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Oversubscription of queues or processes results in poor performance
mostly because HWS blinbly schedules busy and idle queues, resulting
in poor occupancy if many queues are idle.
Let users know with a warning message when transitioning from a
non-oversubscribed to an oversubscribed runlist.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
As we have dropped the final reference to the object, we do not need to
wait until after the rcu grace period to drop its pages. We still require
struct_mutex to completely unbind the object to release the pages, so we
still need a free-worker to manage that from process context. By
scheduling the release of pages before waiting for the rcu should mean
that we are not trapping those pages from beyond the reach of the
shrinker.
v2: Pass along the request to skip if the vma is busy to the underlying
unbind routine, to avoid checking the reservation underneath the
i915->mm.obj_lock which may be used from inside irq context.
v3: Flip the bit for unbinding while active, for later convenience.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111035
Fixes: a93615f900 ("drm/i915: Throw away the active object retirement complexity")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190703091726.11690-6-chris@chris-wilson.co.uk
PRIME functionality is now provided by GEM object functions. The driver
callback functions are obsolete. So this patch renames them and turns
them into static internal functions of the VRAM helper library. The
implementation of gem_prime_mmap is now unused and the patch removes it.
v3:
* kept each renamed function at its original location within file
* kept documentation
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Gerd Hoffmann <kraxel@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190702115012.4418-6-tzimmermann@suse.de
PRIME functionality is now provided via the callback functions in
struct drm_gem_object_funcs. The driver-structure functions are obsolete.
As a side effect of this patch, VRAM-based drivers get basic PRIME
support automatically without having to set any flags or additional
fields.
v2:
- use existing PRIME functions for object's table
v3:
- move object table to EOF so it can refer to internal interfaces
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Gerd Hoffmann <kraxel@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190702115012.4418-2-tzimmermann@suse.de
Christoph Hellwig says:
====================
Below is a series that cleans up the dev_pagemap interface so that it is
more easily usable, which removes the need to wrap it in hmm and thus
allowing to kill a lot of code
Changes since v3:
- pull in "mm/swap: Fix release_pages() when releasing devmap pages" and
rebase the other patches on top of that
- fold the hmm_devmem_add_resource into the DEVICE_PUBLIC memory removal
patch
- remove _vm_normal_page as it isn't needed without DEVICE_PUBLIC memory
- pick up various ACKs
Changes since v2:
- fix nvdimm kunit build
- add a new memory type for device dax
- fix a few issues in intermediate patches that didn't show up in the end
result
- incorporate feedback from Michal Hocko, including killing of
the DEVICE_PUBLIC memory type entirely
Changes since v1:
- rebase
- also switch p2pdma to the internal refcount
- add type checking for pgmap->type
- rename the migrate method to migrate_to_ram
- cleanup the altmap_valid flag
- various tidbits from the reviews
====================
Conflicts resolved by:
- Keeping Ira's version of the code in swap.c
- Using the delete for the section in hmm.rst
- Using the delete for the devmap code in hmm.c and .h
* branch 'hmm-devmem-cleanup.4': (24 commits)
mm: don't select MIGRATE_VMA_HELPER from HMM_MIRROR
mm: remove the HMM config option
mm: sort out the DEVICE_PRIVATE Kconfig mess
mm: simplify ZONE_DEVICE page private data
mm: remove hmm_devmem_add
mm: remove hmm_vma_alloc_locked_page
nouveau: use devm_memremap_pages directly
nouveau: use alloc_page_vma directly
PCI/P2PDMA: use the dev_pagemap internal refcount
device-dax: use the dev_pagemap internal refcount
memremap: provide an optional internal refcount in struct dev_pagemap
memremap: replace the altmap_valid field with a PGMAP_ALTMAP_VALID flag
memremap: remove the data field in struct dev_pagemap
memremap: add a migrate_to_ram method to struct dev_pagemap_ops
memremap: lift the devmap_enable manipulation into devm_memremap_pages
memremap: pass a struct dev_pagemap to ->kill and ->cleanup
memremap: move dev_pagemap callbacks into a separate structure
memremap: validate the pagemap type passed to devm_memremap_pages
mm: factor out a devm_request_free_mem_region helper
mm: export alloc_pages_vma
...
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The migrate_vma helper is only used by noveau to migrate device private
pages around. Other HMM_MIRROR users like amdgpu or infiniband don't
need it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
All the mm/hmm.c code is better keyed off HMM_MIRROR. Also let nouveau
depend on it instead of the mix of a dummy dependency symbol plus the
actually selected one. Drop various odd dependencies, as the code is
pretty portable.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The ZONE_DEVICE support doesn't depend on anything HMM related, just on
various bits of arch support as indicated by the architecture. Also
don't select the option from nouveau as it isn't present in many setups,
and depend on it instead.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Remove the clumsy hmm_devmem_page_{get,set}_drvdata helpers, and
instead just access the page directly. Also make the page data
a void pointer, and thus much easier to use.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Just use devm_memremap_pages instead of hmm_devmem_add pages to allow
killing that wrapper which doesn't provide a whole lot of benefits.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
hmm_vma_alloc_locked_page is scheduled to go away, use the proper
mm function directly.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
mode->name is a character array in a structure, checking it's
address is pointless and causes a warning with some compilers:
drivers/gpu/drm/drm_connector.c:144:15: error: address of array 'mode->name' will always evaluate to 'true'
[-Werror,-Wpointer-bool-conversion]
mode->name ? mode->name : "",
~~~~~~^~~~ ~
include/drm/drm_print.h:366:29: note: expanded from macro 'DRM_DEBUG_KMS'
drm_dbg(DRM_UT_KMS, fmt, ##__VA_ARGS__)
^~~~~~~~~~~
Remove the check here.
Fixes: 3aeeb13d89 ("drm/modes: Support modes names on the command line")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Ramalingam C <ramlaingam.c@intel.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190628103925.2686249-1-arnd@arndb.de
Despite what I think the prm recommends, commit f2253bd985
("drm/i915/ringbuffer: EMIT_INVALIDATE after switch context") turned out
to be a huge mistake when enabling Ironlake contexts as the GPU would
hang on either a MI_FLUSH or PIPE_CONTROL immediately following the
MI_SET_CONTEXT of an active mesa context (more vanilla contexts, e.g.
simple rendercopies with igt, do not suffer).
Ville found the following clue,
"[DevCTG+]: For the invalidate operation of the pipe control, the
following pointers are affected. The
invalidate operation affects the restore of these packets. If the pipe
control invalidate operation is completed
before the context save, the indirect pointers will not be restored from
memory.
1. Pipeline State Pointer
2. Media State Pointer
3. Constant Buffer Packet"
which suggests by us emitting the INVALIDATE prior to the MI_SET_CONTEXT,
we prevent the context-restore from chasing the dangling pointers within
the image, and explains why this likely prevents the GPU hang.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190419111749.3910-1-chris@chris-wilson.co.uk
(cherry picked from commit 928f8f4231 in drm-intel-next)
Cc: stable@vger.kernel.org
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111014
Fixes: f2253bd985 ("drm/i915/ringbuffer: EMIT_INVALIDATE after switch context")
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
When a register is readonly there is not much we can tell about its
value (apart from its default value?). This can be covered by tests
exercising the value of the register from userspace.
For PS_INVOCATION_COUNT we've got the following piglit tests :
KHR-GL45.pipeline_statistics_query_tests_ARB.functional_fragment_shader_invocations
Vulkan CTS tests :
dEQP-VK.query_pool.statistics_query.fragment_shader_invocations.*
v2: Use a local to shrink under 80cols.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 86554f48e5 ("drm/i915/selftests: Verify whitelist of context registers")
Tested-by: Anuj Phogat <anuj.phogat@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20190629131350.31185-1-chris@chris-wilson.co.uk
Since amdgpu has always requested PCIE atomics, kfd don't
need duplicated PCIE atomics enablement. Referring to amdgpu
request result is enough.
Signed-off-by: Jack Xiao <Jack.Xiao@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
GPU atomics operation depends on PCIE atomics support.
Always enable PCIE atomics ops support in case that
it hasn't been enabled.
Signed-off-by: Jack Xiao <Jack.Xiao@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The new field in amdgpu device is used to record whether the
system has PCIE atomics support. The field can be exposed to
UMD or kfd whether PCIE atomics have supported.
Signed-off-by: Jack Xiao <Jack.Xiao@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
