Previously, in an effort to defer initializing the gpu until firmware
was available (ie. rootfs mounted), the gpu was not loaded at when the
subdevice was bound. Which resulted that clks/etc were requested in a
place that devm couldn't really help unwind if something failed.
Instead move request_firmware() to gpu->hw_init() and construct the gpu
earlier in adreno_bind(). To avoid the rest of the driver needing to
be aware of a gpu that hasn't managed to load firmware and hw_init()
yet, stash the gpu ptr in the adreno device's drvdata, and don't set
priv->gpu() until hw_init() succeeds.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Nearly all of the buffer allocations for kernel allocate an buffer object,
virtual address and GPU iova at the same time. Make a helper function to
handle the details.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
[dropped msm_fbdev conversion to new helper, since it interferes with
display-handover work, where we want to separate allocation and mapping]
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>
msm_gpu's get_timestamp() op (called by the MSM_GET_PARAM ioctl) can
result in register accesses. We need our power domain and clocks to
be active for that. Make sure they are enabled here.
Signed-off-by: Archit Taneja <architt@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>
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>
memptrs->wptr seems to be unused. Remove it to avoid
confusing the upcoming preemption code.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
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>
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>
User space needs to know where the GMEM whole starts so that they
can set up the addressing correctly.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
There are reasons for a memory object to outlive the file descriptor
that created it and so the address space that a buffer object is
attached to must also outlive the file descriptor. Reference count
the address space so that it can remain viable until all the objects
have released their addresses.
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>
Each of the per-generation callbacks was doing this. Lets just simplify
and move it into toplevel show() fxn.
Signed-off-by: Rob Clark <robdclark@gmail.com>
This was never documented or used in upstream dtb. It is used by
downstream bindings from android device kernels. But the quirks are
a property of the gpu revision, and as such are redundant to be listed
separately in dt. Instead, move the quirks to the device table.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Fixes: 9cb07b099fb ("drm/msm: support multiple address spaces")
Reported-by: Riku Voipio <riku.voipio@linaro.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Currently the value written to CP_RB_WPTR is calculated on the fly as
(rb->next - rb->start). But as the code is designed rb->next is wrapped
before writing the commands so if a series of commands happened to
fit perfectly in the ringbuffer, rb->next would end up being equal to
rb->size / 4 and thus result in an out of bounds address to CP_RB_WPTR.
The easiest way to fix this is to mask WPTR when writing it to the
hardware; it makes the hardware happy and the rest of the ringbuffer
math appears to work and there isn't any point in upsetting anything.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
[squash in is_power_of_2() check]
Signed-off-by: Rob Clark <robdclark@gmail.com>
Disable the interrupt during the init sequence to avoid having
interrupts fired for errors and other things that we are not
ready to handle while initializing.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add a new generic function to write a "64" bit value. This isn't
actually a 64 bit operation, it just writes the upper and lower
32 bit of a 64 bit value to a specified LO and HI register. If
a particular target doesn't support one of the registers it can
mark that register as SKIP and writes/reads from that register
will be quietly dropped.
This can be immediately put in place for the ringbuffer base and
the RPTR address. Both writes are converted to use
adreno_gpu_write64() with their respective high and low registers
and the high register appropriately marked as SKIP for both 32 bit
targets (a3xx and a4xx). When a5xx comes it will define valid target
registers for the 'hi' option and everything else will just work.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
When the GPU hardware init function fails (like say, ME_INIT timed
out) return error instead of blindly continuing on. This gives us
a small chance of saving the system before it goes boom.
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>
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>
For some optimizations coming on the userspace side, splitting larger
draw or gmem cmds into multiple cmdstream buffers, we need to support
much more than the previous small/arbitrary limit.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Before we can add vmap shrinking, we really need to know which vmap'ings
are currently being used. So switch to get/put interface. Stubbed put
fxns for now.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Some, but not all, callers of obj->vmap() would check if return
IS_ERR(). So let's actually return an error if vmap() fails. And fixup
the call-sites that were not handling this properly.
Signed-off-by: Rob Clark <robdclark@gmail.com>
At this point, there is nothing left to fail. And submit already has a
fence assigned and is added to the submit_list. Any problems from here
on out are asynchronous (ie. hangcheck/recovery).
Signed-off-by: Rob Clark <robdclark@gmail.com>
It is no longer true that we discard all in-flight submits on recover
(these days we only discard the first one that hung). After the first
re-submitted batch completes it would overwrite the fence with a correct
value, but there would be a window of time which showed all re-submitted
batches as already complete.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Better encapsulate the per-timeline stuff into fence-context. For now
there is just a single fence-context, but eventually we'll also have one
per-CRTC to enable fully explicit fencing.
Signed-off-by: Rob Clark <robdclark@gmail.com>
We need this for GL_TIMESTAMP queries.
Note: currently only supported on a4xx.. a3xx doesn't have this
always-on counter. I think we could emulate it with the one CP
counter that is available, but for now it is of limited usefulness
on a3xx (since we can't seem to do time-elapsed queries in any sane
way with the existing firmware on a3xx, and if you are trying to do
profiling on a tiler you want time-elapsed). We can add that later
if it becomes useful.
Signed-off-by: Rob Clark <robdclark@gmail.com>
As described in the downstream/kgsl driver:
Sometimes the RPTR shadow memory is unreliable causing timeouts
in adreno_idle(). Read it directly from the register instead.
Signed-off-by: Craig Stout <cstout@chromium.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>
We need this in userspace for interpreting some of the perf ctrs.
Note possibly not quite sufficient if we had some frequency mgmt
approach other than race-to-idle. Not really sure what the best
thing to do if we did. Although displaying results as a percentage
of max frequence seems sensible(ish) if we did.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
A few spots in the driver have support for downstream android
CONFIG_MSM_BUS_SCALING. This is mainly to simplify backporting the
driver for various devices which do not have sufficient upstream
kernel support. But the intentionally dead code seems to cause
some confusion. Rename the #define to make this more clear.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Dump a bit more info when the GPU hangs, without having hang_debug
enabled (which dumps a *lot* of registers). Also dump the scratch
registers, as they are useful for determining where in the cmdstream
the GPU hung (and they seem always safe to read when GPU has hung).
Note that the freedreno gallium driver emits increasing counter values
to SCRATCH6 (to identify tile #) and SCRATCH7 (to identify draw #), so
these two in particular can be used to "triangulate" where in the
cmdstream the GPU hung.
Signed-off-by: Rob Clark <robdclark@gmail.com>
In error paths, this was being called without struct_mutex held.
Leading to panics like:
msm 1a00000.qcom,mdss_mdp: No memory protection without IOMMU
Kernel panic - not syncing: BUG!
CPU: 0 PID: 1409 Comm: cat Not tainted 4.0.0-dirty #4
Hardware name: Qualcomm Technologies, Inc. APQ 8016 SBC (DT)
Call trace:
[<ffffffc000089c78>] dump_backtrace+0x0/0x118
[<ffffffc000089da0>] show_stack+0x10/0x20
[<ffffffc0006686d4>] dump_stack+0x84/0xc4
[<ffffffc0006678b4>] panic+0xd0/0x210
[<ffffffc0003e1ce4>] drm_gem_object_free+0x5c/0x60
[<ffffffc000402870>] adreno_gpu_cleanup+0x60/0x80
[<ffffffc0004035a0>] a3xx_destroy+0x20/0x70
[<ffffffc0004036f4>] a3xx_gpu_init+0x84/0x108
[<ffffffc0004018b8>] adreno_load_gpu+0x58/0x190
[<ffffffc000419dac>] msm_open+0x74/0x88
[<ffffffc0003e0a48>] drm_open+0x168/0x400
[<ffffffc0003e7210>] drm_stub_open+0xa8/0x118
[<ffffffc0001a0e84>] chrdev_open+0x94/0x198
[<ffffffc000199f88>] do_dentry_open+0x208/0x310
[<ffffffc00019a4c4>] vfs_open+0x44/0x50
[<ffffffc0001aa26c>] do_last.isra.14+0x2c4/0xc10
[<ffffffc0001aac38>] path_openat+0x80/0x5e8
[<ffffffc0001ac354>] do_filp_open+0x2c/0x98
[<ffffffc00019b60c>] do_sys_open+0x13c/0x228
[<ffffffc00019b72c>] SyS_openat+0xc/0x18
CPU1: stopping
But there isn't any particularly good reason to hold struct_mutex for
teardown, so just standardize on calling it without the mutex held and
use the _unlocked() versions for GEM obj unref'ing
Signed-off-by: Rob Clark <robdclark@gmail.com>
The release_firmware() function tests whether its argument is NULL
and then returns immediately. Thus the test around the call is not needed.
This issue was detected by using the Coccinelle software.
Signed-off-by: Markus Elfring <elfring@users.sourceforge.net>
Reviewed-by: Thierry Reding <treding@nvidia.com>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Register offsets have changed between a3xx and a4xx GPUs.
To be able access these registers in common code, we create
a lookup table, and set of read-write APIs to access the
register through the lookup table.
Signed-off-by: Aravind Ganesan <aravindg@codeaurora.org>
[robclark: remove REG_ADRENO_UNDEFINED, just use zero, and minor
tweaks for latest generated headers]
Signed-off-by: Rob Clark <robdclark@gmail.com>
Move anything that can fail after call to base class msm_gpu_init().
This way, if we fail, active_list has already been initialized so we
don't trip 'WARN_ON(!list_empty(&gpu->active_list))' in
msm_gpu_cleanup().
Signed-off-by: Rob Clark <robdclark@gmail.com>
Add ptr to list of interesting registers to 'struct adreno_gpu' and use
that to move most of the debugfs show and register dump bits down into
adreno_gpu. This will avoid duplication as support for additional
adreno generations is added.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Push a few bits down into adreno_gpu so they won't have to be duplicated
as support for additional adreno generations is added.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Somewhere along the way, the firmware loader sprouted another lock
dependency, resulting in possible deadlock scenario:
&dev->struct_mutex --> &sb->s_type->i_mutex_key#2 --> &mm->mmap_sem
which is problematic vs things like gem mmap.
So introduce a separate mutex to synchronize gpu init.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Downstream kernel IOMMU had a non-standard way of dealing with multiple
devices and multiple ports/contexts. We don't need that on upstream
kernel, so rip out the crazy.
Note that we have to move the pinning of the ringbuffer to after the
IOMMU is attached. No idea how that managed to work properly on the
downstream kernel.
For now, I am leaving the IOMMU port name stuff in place, to simplify
things for folks trying to backport latest drm/msm to device kernels.
Once we no longer have to care about pre-DT kernels, we can drop this
and instead backport upstream IOMMU driver.
Signed-off-by: Rob Clark <robdclark@gmail.com>
