This is part of an effort to move from the vmwgfx_open_hash hashtable to
linux/hashtable implementation.
Refactor the ref_hash hashtable, used for fast lookup of reference objects
associated with a ttm file.
This also exposed a problem related to inconsistently using 32-bit and
64-bit keys with this hashtable. The hash function used changes depending
on the size of the type, and results are not consistent across numbers,
for example, hash_32(329) = 329, but hash_long(329) = 328. This would
cause the lookup to fail for objects already in the hashtable, since keys
of different sizes were being passed during adding and lookup. This was
not an issue before because vmwgfx_open_hash always used hash_long.
Fix this by always using 64-bit keys for this hashtable, which means that
hash_long is always used.
Signed-off-by: Maaz Mombasawala <mombasawalam@vmware.com>
Reviewed-by: Zack Rusin <zackr@vmware.com>
Signed-off-by: Zack Rusin <zackr@vmware.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20221022040236.616490-11-zack@kde.org
Vmwgfx's hashtab implementation needs to be replaced with linux/hashtable
to reduce maintenence burden.
As part of this effort, refactor the res_ht hashtable used for resource
validation during execbuf execution to use linux/hashtable implementation.
This also refactors vmw_validation_context to use vmw_sw_context as the
container for the hashtable, whereas before it used a vmwgfx_open_hash
directly. This makes vmw_validation_context less generic, but there is
no functional change since res_ht is the only instance where validation
context used a hashtable in vmwgfx driver.
Signed-off-by: Maaz Mombasawala <mombasawalam@vmware.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Signed-off-by: Zack Rusin <zackr@vmware.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20221022040236.616490-6-zack@kde.org
There's no point in explicitly trying to align virtual memory to
facilitate huge page table entries or huge page memory in buffer objects
given that they're not being used.
Transparent hugepages support for vram allocations has been gradually
retired over the last two years making alignment of unmapped areas
unneeded and pointless.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220425203152.1314211-1-zack@kde.org
The vmw_user_bo_noref_lookup() function cannot return NULL. If it
could, then this function would return PTR_ERR(NULL) which is success.
Returning success without initializing "*vmw_bo_p = vmw_bo;" would
lead to an uninitialized variable bug in the caller. Smatch complains
about this:
drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c:1177 vmw_translate_mob_ptr() warn: passing zero to 'PTR_ERR'
drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c:1314 vmw_cmd_dx_bind_query() error: uninitialized symbol 'vmw_bo'.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Zack Rusin <zackr@vmware.com>
Link: https://patchwork.freedesktop.org/patch/msgid/YtZ9qrKeBqmmK8Hv@kili
Use pageref->offset instead of page->index for deferred-I/O writeback
where appropriate. Distinguishes between file-mapping offset and video-
memory offset. While at it, also remove unnecessary references to
struct page.
Fbdev's deferred-I/O code uses the two related page->index and
pageref->offset. The former is the page offset in the mapped file,
the latter is the byte offset in the video memory (or fbdev screen
buffer). It's the same value for fbdev drivers, but for DRM the values
can be different. Because GEM buffer objects are mapped at an offset
in the DRM device file, page->index has this offset added to it as well.
We currently don't hit this case in DRM, because all affected mappings
of GEM memory are performed with an internal, intermediate shadow buffer.
The value of page->index is required by page_mkclean(), which we
call to reset the mappings during the writeback phase of the deferred
I/O. The value of pageref->offset is for conveniently getting an offset
into video memory in fb helpers.
v4:
* fix commit message (Javier)
Suggested-by: Javier Martinez Canillas <javierm@redhat.com>
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Javier Martinez Canillas <javierm@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220429100834.18898-6-tzimmermann@suse.de
Rename various instances of pagelist to pagereflist. The list now
stores pageref structures, so the new name is more appropriate.
In their write-back helpers, several fbdev drivers refer to the
pageref list in struct fb_deferred_io instead of using the one
supplied as argument to the function. Convert them over to the
supplied one. It's the same instance, so no change of behavior
occurs.
v4:
* fix commit message (Javier)
Suggested-by: Sam Ravnborg <sam@ravnborg.org>
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Javier Martinez Canillas <javierm@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220429100834.18898-5-tzimmermann@suse.de
Store the per-page state for fbdev's deferred I/O in struct
fb_deferred_io_pageref. Maintain a list of pagerefs for the pages
that have to be written back to video memory. Update all affected
drivers.
As with pages before, fbdev acquires a pageref when an mmaped page
of the framebuffer is being written to. It holds the pageref in a
list of all currently written pagerefs until it flushes the written
pages to video memory. Writeback occurs periodically. After writeback
fbdev releases all pagerefs and builds up a new dirty list until the
next writeback occurs.
Using pagerefs has a number of benefits.
For pages of the framebuffer, the deferred I/O code used struct
page.lru as an entry into the list of dirty pages. The lru field is
owned by the page cache, which makes deferred I/O incompatible with
some memory pages (e.g., most notably DRM's GEM SHMEM allocator).
struct fb_deferred_io_pageref now provides an entry into a list of
dirty framebuffer pages, freeing lru for use with the page cache.
Drivers also assumed that struct page.index is the page offset into
the framebuffer. This is not true for DRM buffers, which are located
at various offset within a mapped area. struct fb_deferred_io_pageref
explicitly stores an offset into the framebuffer. struct page.index
is now only the page offset into the mapped area.
These changes will allow DRM to use fbdev deferred I/O without an
intermediate shadow buffer.
v3:
* use pageref->offset for sorting
* fix grammar in comment
v2:
* minor fixes in commit message
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Javier Martinez Canillas <javierm@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220429100834.18898-3-tzimmermann@suse.de
The fbdev mmap function fb_mmap() unconditionally overrides the
driver's implementation if deferred I/O has been activated. This
makes it hard to implement mmap with anything but a vmalloc()'ed
software buffer. That is specifically a problem for DRM, where
video memory is maintained by a memory manager.
Leave the mmap handling to drivers and expect them to call the
helper for deferred I/O by thmeselves.
v4:
* unlock mm_lock in fb_mmap() error path (Dan)
v3:
* fix warning if fb_mmap is missing (kernel test robot)
v2:
* print a helpful error message if the defio setup is
incorrect (Javier)
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Javier Martinez Canillas <javierm@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220429100834.18898-2-tzimmermann@suse.de
Linux 5.18-rc5
There was a build fix for arm I wanted in drm-next, so backmerge rather then cherry-pick.
Signed-off-by: Dave Airlie <airlied@redhat.com>
The buffer objects created by cotables were missing fence reservations.
They are created from vmw_validation_res_validate which makes them miss
the ttm_eu_reserve_buffers which is called from vmw_validation_bo_reserve.
Cotables are the only resources which create a buffer object in the
create callback so make sure the code also reserves the slots.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Co-developed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Fixes: c8d4c18bfb ("dma-buf/drivers: make reserving a shared slot mandatory v4")
Link: https://patchwork.freedesktop.org/patch/msgid/20220422161342.1142584-1-zack@kde.org
v2: Add the last part of the ref count fix which was spotted by
Philipp Sieweck where the ref count of cpu writers is off due to
ERESTARTSYS or EBUSY during bo waits.
The initial GEM port broke refcounting on shareable (prime) surfaces and
memory evictions. The prime surfaces broke because the parent surfaces
weren't increasing the ref count on GEM surfaces, which meant that
the memory backing textures could have been deleted while the texture
was still accessible. The evictions broke due to a typo, the code was
supposed to exit if the passed buffers were not vmw_buffer_object
not if they were. They're tied because the evictions depend on having
memory to actually evict.
This fixes crashes with XA state tracker which is used for xrender
acceleration on xf86-video-vmware, apps/tests which use a lot of
memory (a good test being the piglit's streaming-texture-leak) and
desktops.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Fixes: 8afa13a058 ("drm/vmwgfx: Implement DRIVER_GEM")
Reported-by: Philipp Sieweck <psi@informatik.uni-kiel.de>
Cc: <stable@vger.kernel.org> # v5.17+
Reviewed-by: Maaz Mombasawala <mombasawalam@vmware.com>
Reviewed-by: Martin Krastev <krastevm@vmware.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220420040328.1007409-1-zack@kde.org
Add an usage for kernel submissions. Waiting for those are mandatory for
dynamic DMA-bufs.
As a precaution this patch also changes all occurrences where fences are
added as part of memory management in TTM, VMWGFX and i915 to use the
new value because it now becomes possible for drivers to ignore fences
with the WRITE usage.
v2: use "must" in documentation, fix whitespaces
v3: separate out some driver changes and better document why some
changes should still be part of this patch.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20220407085946.744568-5-christian.koenig@amd.com
Instead of distingting between shared and exclusive fences specify
the fence usage while adding fences.
Rework all drivers to use this interface instead and deprecate the old one.
v2: some kerneldoc comments suggested by Daniel
v3: fix a missing case in radeon
v4: rebase on nouveau changes, fix lockdep and temporary disable warning
v5: more documentation updates
v6: separate internal dma_resv changes from this patch, avoids to
disable warning temporary, rebase on upstream changes
v7: fix missed case in lima driver, minimize changes to i915_gem_busy_ioctl
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20220407085946.744568-3-christian.koenig@amd.com
This change adds the dma_resv_usage enum and allows us to specify why a
dma_resv object is queried for its containing fences.
Additional to that a dma_resv_usage_rw() helper function is added to aid
retrieving the fences for a read or write userspace submission.
This is then deployed to the different query functions of the dma_resv
object and all of their users. When the write paratermer was previously
true we now use DMA_RESV_USAGE_WRITE and DMA_RESV_USAGE_READ otherwise.
v2: add KERNEL/OTHER in separate patch
v3: some kerneldoc suggestions by Daniel
v4: some more kerneldoc suggestions by Daniel, fix missing cases lost in
the rebase pointed out by Bas.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20220407085946.744568-2-christian.koenig@amd.com
With very limited vram on svga3 it's difficult to handle all the surface
migrations. Without gbobjects, i.e. the ability to store surfaces in
guest mobs, there's no reason to support intermediate svga2 features,
especially because we can fall back to fb traces and svga3 will never
support those in-between features.
On svga3 we wither want to use fb traces or screen targets
(i.e. gbobjects), nothing in between. This fixes presentation on a lot
of fusion/esxi tech previews where the exposed svga3 caps haven't been
finalized yet.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Fixes: 2cd80dbd35 ("drm/vmwgfx: Add basic support for SVGA3")
Cc: <stable@vger.kernel.org> # v5.14+
Reviewed-by: Martin Krastev <krastevm@vmware.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220318174332.440068-5-zack@kde.org
Writes to SVGA_REG_CURSOR_MOBID did not wait for the buffers to be fully
populated. This sometimes results in the device not being aware of
the buffer when the cursor mob register was written.
Properly wait for the buffer to be fully populated before setting it
as a cursor mob.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Fixes: 485d98d472 ("drm/vmwgfx: Add support for CursorMob and CursorBypass 4")
Reviewed-by: Martin Krastev <krastevm@vmware.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220318174332.440068-3-zack@kde.org
vmw_move assumed that buffers to be moved would always be
vmw_buffer_object's but after introduction of new placement for mob
pages that's no longer the case.
The resulting invalid read didn't have any practical consequences
because the memory isn't used unless the object actually is a
vmw_buffer_object.
Fix it by moving the cast to the spot where the results are used.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Fixes: f6be23264b ("drm/vmwgfx: Introduce a new placement for MOB page tables")
Reported-by: Chuck Lever III <chuck.lever@oracle.com>
Reviewed-by: Martin Krastev <krastevm@vmware.com>
Tested-by: Chuck Lever <chuck.lever@oracle.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220318174332.440068-2-zack@kde.org
Initial version of guest backed objects in the host had some performance
issues that made using surface-dma's instead of direct copies faster.
Surface dma's force a migration to vram which at best is slow and at
worst is impossible (e.g. on svga3 where there's not enough vram
to migrate fb's to it).
Slowly migrate away from surface dma's to direct copies by limiting
their usage to systems with more than 32MB of vram.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Maaz Mombasawala <mombasawalam@vmware.com>
Reviewed-by: Martin Krastev <krastevm@vmware.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220302152426.885214-9-zack@kde.org
SVGAv3 deprecates legacy interrupts and adds support for MSI/MSI-X. With
MSI the driver visible side remains largely unchanged but with MSI-X
each interrupt gets delivered on its own vector.
Add support for MSI/MSI-X while preserving the old functionality for
SVGAv2. Code between the SVGAv2 and SVGAv3 is exactly the same, only
the number of available vectors changes, in particular between legacy
and MSI-X interrupts.
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Martin Krastev <krastevm@vmware.com>
Reviewed-by: Maaz Mombasawala <mombasawalam@vmware.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220307162412.1183049-1-zack@kde.org
Port of the vmwgfx to SVGAv3 lacked support for fencing. SVGAv3 removed
FIFO's and replaced them with command buffers and extra registers.
The initial version of SVGAv3 lacked support for most advanced features
(e.g. 3D) which made fences unnecessary. That is no longer the case,
especially as 3D support is being turned on.
Switch from FIFO commands and capabilities to command buffers and extra
registers to enable fences on SVGAv3.
Fixes: 2cd80dbd35 ("drm/vmwgfx: Add basic support for SVGA3")
Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Martin Krastev <krastevm@vmware.com>
Reviewed-by: Maaz Mombasawala <mombasawalam@vmware.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220302152426.885214-5-zack@kde.org
