A2XX has its own very simple MMU.
Added a msm_use_mmu() function because we can't rely on iommu_present to
decide to use MMU or not.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Signed-off-by: Rob Clark <robdclark@gmail.com>
For debugging purposes it is useful to assign descriptions
to buffers so that we know what they are used for. Add
a field to the buffer object and use that to name the various
kernel side allocations which ends up looking like like this
in /d/dri/X/gem:
flags id ref offset kaddr size madv name
00040000: I 0 ( 1) 00000000 0000000070b79eca 00004096 memptrs
vmas: [gpu: 01000000,mapped,inuse=1]
00020000: I 0 ( 1) 00000000 0000000031ed4074 00032768 ring0
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add a reference count to track how many times a particular
chunk of iova memory is pinned (mapped) in the iomu and
add msm_gem_unpin_iova to give up references.
It is important to note that msm_gem_unpin_iova replaces
msm_gem_put_iova because the new implicit behavior
that an assigned iova in a given vma is now valid for the
life of the buffer and what we are really focusing on is
the use of that iova.
For now the unmappings are lazy; once the reference counts
go to zero they *COULD* be unmapped dynamically but that
will require an outside force such as a shrinker or
mm_notifiers. For now, we're just focusing on getting
the counting right and setting ourselves up to be ready
for the future.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add a new function to get and pin the iova memory in one
step (basically renaming the old msm_gem_get_iova function)
and switch msm_gem_get_iova() to only allocate an iova but
not map it in the IOMMU. This is only currently used by
msm_ioctl_gem_info() since all other users of of the iova
expect that the memory be immediately available.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Buffer objects allocated with msm_gem_kernel_new() are mostly
freed the same way so we can save a few lines of code with a
common function.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Instead of trying to store all the tagged buffers from a hanging
submit only store the command buffers that were not imported.
This cuts down on the amount of data stored in the GPU state to
the base minimum of useful information.
The downside is that this will make it more difficult to
successfully replay a hang with just the GPU state but there
isn't any reason why that functionality can't be added back
in later once we've figured out how to better communicate
such massive amounts of data.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add trace events to track the progress of a GPU submission
msm_gpu_submit occurs at the beginning of the submissions,
msm_gpu_submit_flush happens when the submission is put on
the ringbuffer and msm_submit_flush_retired is sent when
the operation is retired.
To make it easier to track the operations a unique sequence
number is assigned to each submission and displayed in each
event output so a human or a script can easily associate
the events related to a specific submission.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Allocate the correct buffer size for the GPU memptrs. The incorrect
size hasn't affected us thus far since the incorrect size was larger
than the intended size and we're still stuck on page sized
granularity anyway but technically correct is the best kind of
correct.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Use DRM_DEV_INFO/ERROR/WARN instead of dev_info/err/debug to generate
drm-formatted specific log messages so that it will be easy to
differentiate in case of multiple instances of driver.
Signed-off-by: Mamta Shukla <mamtashukla555@gmail.com>
Signed-off-by: Rob Clark <robdclark@gmail.com>
The devfreq framework requires the drivers to provide busy time estimations.
The GPU driver relies on the hardware performance counteres for the busy time
estimations, but different hardware revisions have counters which can be
sourced from different clocks. So the busy time estimation will be target
dependent. Additionally on targets where the clocks are completely controlled
by the on chip microcontroller, fetching and setting the current GPU frequency
will be different. This patch aims to embrace these differences by re-factoring
the devfreq code a bit.
Signed-off-by: Sharat Masetty <smasetty@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Devfreq turns on and starts recommending power level as soon as it is
initialized. The GPU is still not powered on by the time the devfreq
init happens and this leads to problems on GPU's where register access
is needed to get/set power levels. So we start suspended and only restart
devfreq when GPU is powered on.
Signed-off-by: Sharat Masetty <smasetty@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
When CONFIG_DEV_COREDUMP isn't defined msm_gpu_crashstate_capture
doesn't pass the correct parameters.
drivers/gpu/drm/msm/msm_gpu.c: In function ‘recover_worker’:
drivers/gpu/drm/msm/msm_gpu.c:479:34: error: passing argument 2 of ‘msm_gpu_crashstate_capture’ from incompatible pointer type [-Werror=incompatible-pointer-types]
msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
^~~~~~
drivers/gpu/drm/msm/msm_gpu.c:388:13: note: expected ‘char *’ but argument is of type ‘struct msm_gem_submit *’
static void msm_gpu_crashstate_capture(struct msm_gpu *gpu, char *comm,
^~~~~~~~~~~~~~~~~~~~~~~~~~
drivers/gpu/drm/msm/msm_gpu.c:479:2: error: too many arguments to function ‘msm_gpu_crashstate_capture’
msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
^~~~~~~~~~~~~~~~~~~~~~~~~~
drivers/gpu/drm/msm/msm_gpu.c:388:13: note: declared here
static void msm_gpu_crashstate_capture(struct msm_gpu *gpu, char *comm,
In current code the function msm_gpu_crashstate_capture parameters.
Fixes: cdb95931de ("drm/msm/gpu: Add the buffer objects from the submit to the crash dump")
Signed-off-by: Anders Roxell <anders.roxell@linaro.org>
Reviewed-By: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add support for the A6XX family of Adreno GPUs. The biggest addition
is the GMU (Graphics Management Unit) which takes over most of the
power management of the GPU itself but in a ironic twist of fate
needs a goodly amount of management itself. Add support for the
A6XX core code, the GMU and the HFI (hardware firmware interface)
queue that the CPU uses to communicate with the GMU.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add a helper function to parse the clock names and set up
the bulk data so we can take advantage of the bulk clock
functions instead of rolling our own. This is added
as a helper function so the upcoming a6xx GMU code can
also take advantage of it.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
All users of do_gettimeofday() have been removed, but this one recently
crept in, along with an incorrect printing of the microseconds portion.
This converts it to using ktime_get_real_timespec64() as a direct
replacement, and adds the leading zeroes. I considered using monotonic
times (ktime_get()) instead, but as this timestamp appears to only
be used for humans rather than compared with other timestamps, the
real time domain is probably good enough.
Fixes: e43b045e2c82 ("drm/msm/gpu: Capture the state of the GPU")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Rob Clark <robdclark@gmail.com>
For hangs, dump copy out the contents of the buffer objects attached to the
guilty submission and print them in the crash dump report.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Capture the GPU state on a GPU hang and store it for later playback
via the devcoredump facility. Only one crash state is stored at a
time on the assumption that the first hang is usually the most
interesting. The existing crash state can be cleared after capturing
it and then a new one will be captured on the next hang.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Do a bit of cleanup to prepare for upcoming changes to pass the
hanging task comm and cmdline to the crash dump function.
v2: Use GFP_ATOMIC while holding the rcu lock per Chris Wilson
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
If we fail to allocate gpu->grp_clks reset the number of available
clocks to zero to avoid referencing the missing array later.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
drm_gem_object_{reference,unreference,unreference_unlocked} are
deprecated functions, and merely alias to the get/put functions.
Switch to the new names.
Signed-off-by: Steve Kowalik <steven@wedontsleep.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add support for devfreq to dynamically control the GPU frequency.
By default try to use the 'simple_ondemand' governor which can
adjust the frequency based on GPU load.
v2: Fix __aeabi_uldivmod issue from the 0 day bot and use
devfreq_recommended_opp() as suggested by Rob.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Remove the downstream bus scaling code. It isn't needed for for
compatibility with a downstream or vendor kernel. Get it out of the
way to clear space for devfreq support.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
The fault recovery code tries to sync fences on all possible rings
instead of only the rings that actually exist which will fault the
kernel when the number of rings are less than the maximum amount.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
The return type of ARRAY_SIZE() is size_t, so we have to use
%zu instead of %lu to avoid this warning:
drivers/gpu/drm/msm/msm_gpu.c: In function 'msm_gpu_init':
drivers/gpu/drm/msm/msm_gpu.c:742:31: error: format '%lu' expects argument of type 'long unsigned int', but argument 7 has type 'unsigned int' [-Werror=format=]
The warning it otherwise harmless as size_t is always the
same size as unsigned long in all supported architectures,
but gcc doesn't know that.
Fixes: c2fceabca6d5 ("drm/msm: Support multiple ringbuffers")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Note we need to move update_fences() to after msm_rd_dump_submit(),
otherwise the bo's referenced by the submit may no longer be valid.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Split into two instances, the existing $debugfs/rd which continues to
dump all submits, and $debugfs/hangrd which will be used to dump just
submits that cause gpu hangs (and eventually faults, but that will
require some iommu framework enhancements).
Signed-off-by: Rob Clark <robdclark@gmail.com>
Now that freedreno gallium driver defaults to using submit_queue task
(render reordering), just showing task->comm is not so useful (ie. it is
always "flush_queue:0"), so also dump the cmdline. This should also be
more useful for piglit/shader_runner.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Implement preemption for A5XX targets - this allows multiple
ringbuffers for different priorities with automatic preemption
of a lower priority ringbuffer if a higher one is ready.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add the infrastructure to support the idea of multiple ringbuffers.
Assign each ringbuffer an id and use that as an index for the various
ring specific operations.
The biggest delta is to support legacy fences. Each fence gets its own
sequence number but the legacy functions expect to use a unique integer.
To handle this we return a unique identifier for each submission but
map it to a specific ring/sequence under the covers. Newer users use
a dma_fence pointer anyway so they don't care about the actual sequence
ID or ring.
The actual mechanics for multiple ringbuffers are very target specific
so this code just allows for the possibility but still only defines
one ringbuffer for each target family.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
When we move to multiple ringbuffers we're going to store the data
in the memptrs on a per-ring basis. In order to prepare for that
move the current memptrs from the adreno namespace into msm_gpu.
This is way cleaner and immediately lets us kill off some sub
functions so there is much less cost later when we do move to
per-ring structs.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Currently the GPU MMU is attached in the adreno_gpu code but as
more and more of the GPU initialization moves to the generic
GPU path we have a need to map and use GPU memory earlier and
earlier. There isn't any reason to defer attaching the MMU
until later so attach it right after the address space is
created so it can be used immediately.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Buffer object specific resources like pages, domains, sg list
need not be protected with struct_mutex. They can be protected
with a buffer object level lock. This simplifies locking and
makes it easier to avoid potential recursive locking scenarios
for SVM involving mmap_sem and struct_mutex. This also removes
unnecessary serialization when creating buffer objects, and also
between buffer object creation and GPU command submission.
Signed-off-by: Sushmita Susheelendra <ssusheel@codeaurora.org>
[robclark: squash in handling new locking for shrinker]
Signed-off-by: Rob Clark <robdclark@gmail.com>
Now that the msm_gem supports an arbitrary number of vma's, we no longer
need to assign an id (index) to each address space. So rip out the
associated code.
Signed-off-by: Rob Clark <robdclark@gmail.com>
No functional change, that will come later. But this will make it
easier to deal with dynamically created address spaces (ie. per-
process pagetables for gpu).
Signed-off-by: Rob Clark <robdclark@gmail.com>
Most, but not all, paths where calling the with struct_mutex held. The
fast-path in msm_gem_get_iova() (plus some sub-code-paths that only run
the first time) was masking this issue.
So lets just always hold struct_mutex for hw_init(). And sprinkle some
WARN_ON()'s and might_lock() to avoid this sort of problem in the
future.
Signed-off-by: Rob Clark <robdclark@gmail.com>
The amount of information that we need to pass into msm_gpu_init()
is steadily increasing, so add a new struct to stabilize the function
call and make it easier to add new configuration down the line.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Otherwise if someone was using old bindings with "core_clk" instead of
"core" as the clock name, we'd never find it and gpu would be stuck at
27MHz (or whatever it's slowest rate is).
Fixes: 98db803 ("msm/drm: gpu: Dynamically locate the clocks from the device tree")
Signed-off-by: Rob Clark <robdclark@gmail.com>
Instead of using a fixed list of clock names use the clock-names
list in the device tree to discover and get the list of clocks
that we need.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Some A3XX and A4XX GPU targets required that the GPU clock be
programmed to a non zero value when it was disabled so
27Mhz was chosen as the "invalid" frequency.
Even though newer targets do not have the same clock restrictions
we still write 27Mhz on clock disable and expect the clock subsystem
to round down to zero.
For unknown reasons even though the slow clock speed is always
27Mhz and it isn't actually a functional level the legacy device tree
frequency tables always defined it and then did gymnastics to work
around it.
Instead of playing the same silly games just hard code the "slow" clock
speed in the code as 27MHz and save ourselves a bit of infrastructure.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
We should be detaching the MMU before destroying the address
space. To do this cleanly, the detach has to happen in
adreno_gpu_cleanup() because it needs access to structs
in adreno_gpu.c. Plus it is better symmetry to have
the attach and detach at the same code level.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Suggested by Rob Herring. We still support the old names for
compatibility with downstream android dt files.
Cc: Rob Herring <robh@kernel.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Acked-by: Rob Herring <robh@kernel.org>
The adreno code inherited a silly workaround from downstream
from the bad old days before decent clock control. grp_clk[0]
(named 'src_clk') doesn't actually exist - it was used as a proxy
for whatever the core clock actually was (usually 'core_clk').
All targets should be able to correctly request 'core_clk' and
get the right thing back so zap the anachronism and directly
use grp_clk[0] to control the clock rate.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
For a5xx the gpu is 64b so we need to change iova to 64b everywhere. On
the display side, iova is still 32b so it can ignore the upper bits.
(Although all the armv8 devices have an iommu that can map 64b pa to 32b
iova.)
Signed-off-by: Rob Clark <robdclark@gmail.com>
We can have various combinations of 64b and 32b address space, ie. 64b
CPU but 32b display and gpu, or 64b CPU and GPU but 32b display. So
best to decouple the device iova's from mmap offset.
Signed-off-by: Rob Clark <robdclark@gmail.com>
