Restrict reading of FE object table to the number of entries
available.
Change-Id: I11275ecd14e53f0b763d00d65042adb4b1e8ae6f
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: http://git-master/r/449306
Runlist event is not sent in gm20b for updated runlist. Polling is
the preferred way also for gk20a.
Bug 1555239
Change-Id: I60de084db69f848f63451f1f3078f183ca51ba50
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: http://git-master/r/500241
Add poll interface and control ioctls for waiting for GPU job completion
via semaphores.
Poll on a gk20a channel file waits for events from pending semaphore
interrupts (stalling) of that channel. New ioctls enable and disable the
events, and clear a single interrupt event so that next poll doesn't
wake up for it again.
Bug 1528781
Change-Id: I5c6238966b5d0900c8ab263c6a7f8f2611901f33
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: http://git-master/r/497750
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Tested-by: Terje Bergstrom <tbergstrom@nvidia.com>
Expose supported nvgpu ioctls to userspace via bits in the flags field
of nvhost_gpu_characteristics; currently define two bits for special
memory allocation support.
Bug 1539747
Change-Id: I1bc9333b12825d07a00b7a4136ae9d35816a5855
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: http://git-master/r/495942
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Tested-by: Terje Bergstrom <tbergstrom@nvidia.com>
L2 bypass registers have moved in gm20b. Move the code to
ltc_common.c, which gets compiled once per chip version.
Change-Id: I0ab4dd03c78b8ad8abc7a7b18c094b6002827587
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: http://git-master/r/499220
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Deepak Nibade <dnibade@nvidia.com>
Use TSG specific API gk20a_fifo_recover_tsg() in following cases :
- IOCTL_CHANNEL_FORCE_RESET
to force reset a channel in TSG, reset all the channels
- handle pbdma intr
while resetting in case of pbdma intr, if channel is part of
TSG, recover entire TSG
- TSG preempt failure
when TSG preempt times out, use TSG recover API
Use preempt_tsg() API to preempt if channel is part of TSG
Add below two generic APIs which will take care of preempting/
recovering either of channel or TSG as required
gk20a_fifo_preempt()
gk20a_fifo_force_reset_ch()
Bug 1470692
Change-Id: I8d46e252af79136be85a9a2accf8b51bd924ca8c
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/497875
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
- add support to handle MMU faults on a channel in TSG
- first get the ID and type of channel that engine is running
- if TSG, abort each channel in it
- if regular channel, abort that channel
- also, add two versions of API set_ctx_mmu_error(), one for
regular channel and another for TSG
Bug 1470692
Change-Id: Ia7b01b81739598459702ed172180adb00e345eba
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/497874
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
- add and export API "gk20a_fifo_recover_tsg()" to
recover a TSG
- if TSG is running on any engine, then trigger MMU fault
on those engines
- otherwise, abort each channel in TSG
- modify channel specific API engines_on_ch() to generic
engines_on_id() which will take an ID and a flag to specify
whether ID is for channel or TSG and return engines running
on that ID
- modify channel specific API get_faulty_channel() to generic
get_faulty_id_type() which will take pointers to ID and type
of ID (either a regular channel or TSG)
- remove runlist update from recover_ch() since
no need to touch runlist during recovery
- set error notifier first and then only abort the channels
for TSG recovery path
- also, add necessary accessors to get engine
status type as TSG
Bug 1470692
Change-Id: I7137f611f80916b3d256d4b0dc6e5cf1e93eef6f
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/497873
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Restored changing GM20B GPCPLL post-divider in flight with the
following limitation: post divider transition is glitch-less only if
there is common "1" in binary representation of old and new settings.
Transitions that may create glitch are implemented in glitch-less steps
with minimum possible interim divider value (for example, 1 <=> 2
transition has interim value 3: 1 <=> 3 <=> 2).
Steps allowed for glitch-less transitions may not always have frequency
jump at/below VCO min/2 (in the example above 1st step jumps 2/3 of
VCOmin). Enabled external linear divider at 1:2 during such steps.
Used extra write of the same data when changing GM20b linear divider.
Bug 1552225
Change-Id: Ie8fba2fbe44afd34ca68f5f355bd302b7426a632
Signed-off-by: Alex Frid <afrid@nvidia.com>
Reviewed-on: http://git-master/r/496319
(cherry picked from commit bdd21e0003032fe664bd20f163dbab9942fd1d1d)
Reviewed-on: http://git-master/r/499193
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Yu-Huan Hsu <yhsu@nvidia.com>
- Skipped PLL re-locking if only post-divider is changing under bypass
- Added 1us delay after switch to bypass clock source
- Changed wait for lock under bypass resolution from 2us to 1us
Change-Id: I259581c00c417752263ef3b2ea057200bb78ecbf
Signed-off-by: Alex Frid <afrid@nvidia.com>
Reviewed-on: http://git-master/r/495473
(cherry picked from commit d90a19b8bf59c608a2a3a891b34ca714dfe990e9)
Reviewed-on: http://git-master/r/499192
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Yu-Huan Hsu <yhsu@nvidia.com>
Provide sysedp_dynamic_capping with the instantaneous GPU frequency
when notifying it of the GPU load. Modify the gpu/cpu priority
decision logic to choose CPU priority until GPU frequency gets "near"
the CPU-priority-limited-GPU-fmax. Introduce the priority_bias debugfs
parameter to facilitate tuning of "near". priority_bias takes a value
from 0 to 100.
Change-Id: Ia2cba36b8ea024fb8b01b5ba195dcf6550e38121
Signed-off-by: Matt Longnecker <mlongnecker@nvidia.com>
Reviewed-on: http://git-master/r/481720
GVS: Gerrit_Virtual_Submit
Reviewed-on: http://git-master/r/498912
Reviewed-by: Timo Alho <talho@nvidia.com>
The compression state tracking user space API already accepts and
returns the ZBC color used for the surface. Actually store the color
in kernel so that the feature works.
Bug 1536227
Bug 1524301
Change-Id: I264e1eeb90f0c4d40fe35fc2479b0ce83e19a7d7
Signed-off-by: Lauri Peltonen <lpeltonen@nvidia.com>
Reviewed-on: http://git-master/r/497476
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Tested-by: Jussi Rasanen <jrasanen@nvidia.com>
Reviewed-by: Arto Merilainen <amerilainen@nvidia.com>
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Defer CDE app initialization to the point where we actually need to
launch the app. This allows us to use the compression state API also on
T124 where we never use the CDE app.
Also return the error code correctly from gk20a_prepare_compressible_read.
Bug 1524301
Change-Id: If79fbe161e8dc9353b9f5fa0dfcd7f30b00d29b4
Signed-off-by: Lauri Peltonen <lpeltonen@nvidia.com>
Reviewed-on: http://git-master/r/497351
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Jussi Rasanen <jrasanen@nvidia.com>
Tested-by: Jussi Rasanen <jrasanen@nvidia.com>
Reviewed-by: Arto Merilainen <amerilainen@nvidia.com>
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
dma_buf_get returns PTR_ERRs, so fix checking for null to proper IS_ERR
in gk20a_vm_map_buffer. Buffer mapping from user space with ioctls would
also have paniced here if an improper handle would be passed.
Change-Id: I245fe41cd209e49fc9265e56340c1c8215ffb1d2
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: http://git-master/r/498320
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
If a channel is already scheduled as regular channel, we should not
allow it to be marked as TSG since it will fail book keeping of
number of active channels in a TSG
This way we can force to bind the channels first and then only
make them active
Also, remove duplicate function declaration added during branch merge
and one unnecessary comparison with zero
Bug 1470692
Change-Id: I88f9678919e4b76de472c6dda21e4537520241c4
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/497903
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Disable illegal compstat access interrupt. We access compstat backing
store to handle CDE swizzling.
Also change the magic number for evicted_cb to use a generated value.
Change-Id: I79b299abbffcb90497690ba4fc55d8517a3dbd87
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: http://git-master/r/496444
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Arto Merilainen <amerilainen@nvidia.com>
Reviewed-by: Jussi Rasanen <jrasanen@nvidia.com>
Tested-by: Jussi Rasanen <jrasanen@nvidia.com>
Add refcounting for TSGs and manage the refcounts as below :
- initialize ref when TSG is opened
- get ref when channel is bound to TSG
- drop the ref when channel is unbound (i.e. during channel close)
- drop the ref when TSG is closed
- when refcount drops to zero, we free the TSG
This refcounting makes it possible to close channels or TSG
in any order
Bug 1470692
Change-Id: Ia4b39164a4582c8169da62a91b9131094c67f5f8
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/495667
(cherry picked from commit be1e198a663d2102e9674978f3d2cca0f2327a6b)
Reviewed-on: http://git-master/r/495955
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Tested-by: Terje Bergstrom <tbergstrom@nvidia.com>
gk20a_channel_release() should bail if filp->private_data is
NULL. This can happen as a result of gk20a_channel_release()
being called when __gk20a_channel_open() fails in
NVHOST_IOCTL_CHANNEL_OPEN.
Bug 200014898
Change-Id: I32cc957aca46fcd4265a8052ac5be355b644b9f7
Signed-off-by: Aingara Paramakuru <aparamakuru@nvidia.com>
Reviewed-on: http://git-master/r/496138
Reviewed-by: Ken Adams <kadams@nvidia.com>
Tested-by: Ken Adams <kadams@nvidia.com>
Add ioctls to enable/disable/preempt a TSG
1. NVGPU_IOCTL_TSG_ENABLE
enable all the channels in a TSG
2. NVGPU_IOCTL_TSG_DISABLE
disable all the channels in a TSG
3. NVGPU_IOCTL_TSG_PREEMPT
preempt a TSG with given id
Bug 1514064
Change-Id: I2db6404b8536e872243cbe57b99e7c6d14243fa5
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/494671
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Add API gk20a_fifo_preempt_tsg() which takes ID of tsg
and preempts it
Bug 1514064
Bug 1470692
Change-Id: I1d52c1dd7a9aecc1314b0f223fe4eedecc033629
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/495583
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Remove static variable gk20a_handle used to store pointer to
struct gk20a
Use below device tree APIs to get pointer to platform_device
and then struct gk20a
device_node = of_find_matching_node()
platform_device = of_find_device_by_node(device_node)
Also, use two versions of do_idle()/do_unidle() APIs -
one will receive void (to be used from outside GPU driver)
and another will receive platform_device (to be used within
GPU driver where it is available)
Change-Id: I9f2c7610646c5fbcd3d99a1b03dc0364201272a8
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/496508
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-by: Ken Adams <kadams@nvidia.com>
GVS: Gerrit_Virtual_Submit
gk20a_pmu_disable_elpg can be called before the PMU driver has
received and processed the INIT message from the PMU. If change
ensures that the ELPG ref-count has been initialized to zero
before that can happen.
Bug 200016313
Change-Id: Ic80ec1ee69b1eb0499effb1abf556f78cb041f5e
Signed-off-by: Neil Gabriel <ngabriel@nvidia.com>
Reviewed-on: http://git-master/r/429161
Reviewed-on: http://git-master/r/433299
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Robert Shih <rshih@nvidia.com>
Tested-by: Robert Shih <rshih@nvidia.com>
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Alloc space writes the page size to a field that requires pgsz_idx.
That can cause corruption in internal kernel structures.
Clear_sparse treated a parameter as page size instead of index.
Bug 1549451
Change-Id: I73ce17b99aae6865056facce72d2ab9ca8b3f81d
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: http://git-master/r/495692
Switch GM20b GPCPLL under bypass when changing post-divider setting
(for now, don't assume that post-divider is glitch-less).
Change-Id: I62b1285c035de0913207a86c41f37b7765da3893
Signed-off-by: Alex Frid <afrid@nvidia.com>
Reviewed-on: http://git-master/r/495300
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Bo Yan <byan@nvidia.com>
We store a reference to common address space of channels
in struct tsg_gk20a without increasing the refcount
This could result in freeing the address space even when
some channel in TSG needs it or when we need to free
common gr_ctx
Fix this by getting ref using gk20a_vm_get() when we store
the VM reference. We drop this reference with
gk20a_vm_put() when closing the TSG
Bug 1470692
Change-Id: Ifc1f29d32cd721810bfbb5a4db96095770318c17
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/495668
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Tested-by: Terje Bergstrom <tbergstrom@nvidia.com>
Add refcounting for TSGs and manage the refcounts as below :
- initialize ref when TSG is opened
- get ref when channel is bound to TSG
- drop the ref when channel is unbound (i.e. during channel close)
- drop the ref when TSG is closed
- when refcount drops to zero, we free the TSG
This refcounting makes it possible to close channels or TSG
in any order
Bug 1470692
Change-Id: Ia4b39164a4582c8169da62a91b9131094c67f5f8
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/495667
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Regenerate clock gating lists. Add new blocks, and takes them into
use. Also moves some clock gating settings to be applied at the
earliest possible moment right after reset.
Change-Id: I21888186c200f7a477c63bd3332e8ed578f63741
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: http://git-master/r/457698
remove config GK20A_PHYS_PAGE_TABLES since all code
dependent on this config is now moved to runtime
selection
Change-Id: I27d2722a9ad91cf4e0537a30943675c9132d6924
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: http://git-master/r/494499
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Shridhar Rasal <srasal@nvidia.com>
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Enable gpu rail gating with INT_MAX delay. This will allow
teams to experiment with different rail-gate entry delay.
Change-Id: I8c696140aba2374c797365282999b6589432047c
Signed-off-by: Seshendra Gadagottu <sgadagottu@nvidia.com>
Reviewed-on: http://git-master/r/491615
GVS: Gerrit_Virtual_Submit
Reviewed-by: Mitch Luban <mluban@nvidia.com>
Display the Falcon Trace prints in the right format.
Embedd the parameters in the string instead of printing it separately.
Bug NA
Change-Id: Ia61fc43384cf6e44a867c7aa9cbb828127146099
Signed-off-by: Vaikundanathan S <vaikuns@nvidia.com>
Reviewed-on: http://git-master/r/488757
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Vijayakumar Subbu <vsubbu@nvidia.com>
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Moved detection of idempotent GPCPLL operations from set_pll_freq()
function to its callers, e.g., explicitly check when enable operation
is called on already enabled PLL, instead of passing same frequency
to set_pll_freq() in such case. Similarly explicitly check when disable
operation is called on already disabled PLL.
Also moved check for GPU powered on from set_pll_freq() to callers,
and skip call to set interface if not.
Added last GPCPLL configuration structure updated after successful
completion of set_pll_freq() function.
Bug 1450787
Change-Id: I8c14b8cab2a8548e98c9b2d223c465c68fb87b61
Signed-off-by: Alex Frid <afrid@nvidia.com>
Reviewed-on: http://git-master/r/488027
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Krishna Reddy <vdumpa@nvidia.com>
The cde shader will only read data from the global compbit backing
store. Map it as read-only to enforce this.
Change-Id: If5be44b8daedd5e7fdee650a6e76befa7bdecfd6
Signed-off-by: Lauri Peltonen <lpeltonen@nvidia.com>
Reviewed-on: http://git-master/r/486679
