This gets tricky as we can't do the TLB invalidation until the unbind
operation is done on the hardware and we can't signal the unbind as
complete until the TLB invalidation is done. To work around this we
create an unbind fence which does a TLB invalidation after unbind is
done on the hardware, signals on TLB invalidation completion, and this
fence is installed in the BO dma-resv slot and installed in out-syncs
for the unbind operation.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Suggested-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com
Suggested-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
We can't currently do this due to TLB invalidation done handler
expecting the seqno being received in-order, with the fast-path a TLB
invalidation done could pass one being processed in the slow-path in an
extreme corner case. Remove TLB invalidation done from the fast-path for
now and in a follow up reenable this once the TLB invalidation done
handler can deal with out of order seqno.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Scratch page is in VRAM, and therefore requires 64K GTT layout. In GGTT
world this just means having 16 consecutive entries, with 64K GTT
alignment for the GTT address of the first entry (also matching physical
alignment). However to keep things simple just dump it into system
memory, like we already do for ppGTT. While we are here, also give it
known default value.
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Spec says we need to use 64K VRAM pages for GGTT on platforms like DG2.
In GGTT this just means aligning the GTT address to 64K and ensuring
that we have 16 consecutive entries each pointing to the respective 4K
entry. We already ensure we have 64K pages underneath, so it's just a
case of forcing the GTT alignment.
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
On DG2 when running the xe_vm IGT, the kernel generates loads of CAT
errors and GT resets (sometimes at least). On small-bar systems seems
to trigger a lot more easily (maybe due to difference in allocation
strategy). Appears to be related to scratch, since we seem to use the
64K TLB hint on scratch entries, even though the scratch page is a 4K
vram page. Bumping the scratch page size and physical alignment seems
to fix it. Or at least we no longer hit:
[ 148.872683] xe 0000:03:00.0: [drm] Engine memory cat error: guc_id=0
[ 148.872701] xe 0000:03:00.0: [drm] Engine memory cat error: guc_id=0
[ 148.875108] WARNING: CPU: 0 PID: 953 at drivers/gpu/drm/xe/xe_guc_submit.c:797
However to keep things simple, so we don't have to deal with 64K TLB
hints, just move the scratch page into system memory on platforms that
require 64K VRAM pages.
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
On DGFX this blows up if can call this with a system memory object:
XE_BUG_ON(!mem_type_is_vram(place->mem_type) && place->mem_type != XE_PL_STOLEN);
If we consider dpt it looks like we can already in theory hit this, if
we run out of vram and stolen vram. It at least seems reasonable to
allow calling this on any object which supports CPU access.
Note this also changes the behaviour with stolen VRAM and suspend, such
that we no longer attempt to migrate stolen objects into system memory.
However nothing in stolen should ever need to be restored (same on
integrated), so should be fine. Also on small-bar systems the stolen
portion is pretty much always non-CPU accessible, and currently pinned
objects use plain memcpy when being moved, which doesn't play nicely.
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
The comparison with < 0 suggests that the memory device access
should be signed to handle underflow. This makes it work more reliably.
As a result, the max refcount is now S32_MAX instead.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
At present the interrupts are enabled while initializing the last GT.
But this is incorrect for a Multi-GT platform, as root GT initialization
will fail with interrupt disabled. Interrupts are required for
the GuC submission triggered during initialization.
Enable the interrupt during the root GT initialization.
Signed-off-by: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Some of the tests may benefit from running with ARCH=um, forgoing any
additional setup on the CI build side. Add min config for that.
Tested with:
./tools/testing/kunit/kunit.py build \
--kunitconfig drivers/gpu/drm/xe/.kunitconfig \
--jobs $(nproc) \
--build_dir build_kunit
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Cc: Mauro Carvalho Chehab <mchehab@kernel.org>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Mauro Carvalho Chehab <mchehab@kernel.org>
mem_type field was added in commit d8b52a02cb ("drm/xe: Implement
stolen memory.") to designate the TTM memory type for that mgr. Add
kernel-doc with its description.
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Fix the following error while building for 32b:
In file included from ../drivers/gpu/drm/xe/xe_gt.c:6:
../drivers/gpu/drm/xe/xe_gt.c: In function ‘gt_ttm_mgr_init’:
../include/linux/minmax.h:20:35: error: comparison of distinct pointer types lacks a cast [-Werror]
20 | (!!(sizeof((typeof(x) *)1 == (typeof(y) *)1)))
| ^~
Cast it to u64 so size of the second operand matches the first one when
building it for 32 bits.
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Leave the types as u64, but cast the pointers to unsigned long before
assigning so the compiler doesn't throw warning about casting a pointer
to integer of different size.
Also, size_t should use %zu, not %ld.
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
writeq() and readq() and other functions working on 64 bit variables
are not provided by 32b arch. For that it's needed to choose between
linux/io-64-nonatomic-hi-lo.h and linux/io-64-nonatomic-lo-hi.h,
spliting the read/write in 2 accesses. For xe driver, it doesn't matter
much, so just choose one and include in xe_mmio.h.
This also removes some ifdef CONFIG_64BIT we had around because of the
missing 64bit functions.
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
On DG2 we are now getting:
[ 104.456607] xe 0000:03:00.0: [drm] *ERROR* PCODE timeout, retrying with preemption disabled
Looks like we just need to invert the error check for
xe_pcode_try_request(), which returns zero on success.
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
On DG1, BAR2 is not reliable for reporting Vram size, need to use GSMBASE.
Simplify xe_mmio_total_vram_size to report vram size and usable size.
Signed-off-by: Philippe Lecluse <philippe.lecluse@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
The bo_create ioctl relied on the internal ordering of memory regions to
be the same, make sure we don't allocate stolen instead of VRAM0.
Also remove a debug warning left in.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Reviewed-by: Philippe Lecluse <philippe.lecluse1@gmail.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Starting in 70.6.* GuC firmware the CSS header includes the submission
version, pull this from the CSS header. Prior 70.* versions accidentally
omitted this informatio so hard code to the correct values. This
information will be used by VFs when communicating with the PF.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Philippe Lecluse <philippe.lecluse1@gmail.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
The blob doesn't fully support this yet, so fake for now to ensure our
driver load order is correct.
Once the blob supports pulling gt->info.engine_mask from the blob, this
patch can be removed.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
This adds support for stolen memory, with the same allocator as
vram_mgr. This allows us to skip a whole lot of copy-paste,
by re-using parts of xe_ttm_vram_mgr.
The stolen memory may be bound using VM_BIND, so it performs like any
other memory region.
We should be able to map a stolen BO directly using the physical memory
location instead of through GGTT even on old platforms, but I don't know
what the effects are on coherency.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
When a job is inflight we may access memory to read the hardware seqno.
All user jobs have VM open which has a ref but kernel jobs do not
require VM so it is possible to not have memory ref. To avoid this, take
a memory ref on kernel job creation.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Expand xe_mmio_wait32 to accept atomic and then use
that directly when possible, and create own routine to
wait for the pcode status.
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
We don't need any macro for a simple check we can do explicitly
and clear.
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
To make it simpler, all of the status checks also waits and
times out.
Also, no ktime precision is needed in this case, and we
can use usleep_range because we are not in atomic paths here.
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Rather than a constant check on proto and wait not busy,
let's wait for the expected success and then check the
protocol afterwards.
With this, we can now use the regular xe_mmio_wait32
and kill this local need for the wait_for.
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Possible now that the wait function returns the last read value.
So we can remove the users of i915's wait_for one by one...
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
This is already useful because it avoids some extra reads
where registers might have changed after the timeout decision.
But also, it will be important to end the kill of i915's wait_for.
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
