- defined platform agnostic wrapper for mempool
mapping and unmapping.
- used platform agnositc wrapper for device
tree parsing.
- modified css_gr_gk20a to include special
handling incase of rm-server
JIRA: VQRM:3699
Change-Id: I08fd26052edfa1edf45a67be57f7d27c38ad106a
Signed-off-by: Antony Clince Alex <aalex@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1733576
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
VBIOS link_disable_mask should be sufficient to find the connected
links. As VBIOS is not updated with correct mask, we parse the DT
node where we hardcode the link_id. DT method is not scalable as same
DT node is used for different dGPUs connected over PCIE. Remove the
DT parsing of link id and use HAL to get link_mask based on the GPU.
JIRA NVLINK-162
Change-Id: Idb7b639962928ce48711a0d7fc277c4c324bee91
Signed-off-by: Tejal Kudav <tkudav@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1738967
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Sysfs nodes for GR stats are created on GR init. If nvgpu
module is removed without any ops, then it tries to remove
sysfs nodes which do not exist resulting in kernel panic.
Fix this issue by removing sysfs nodes only if ecc counters
are initialized.
Bug 1987855
Change-Id: I3f967ee92ec02ad19ffbd9bfa8bace5bfd229dd2
Signed-off-by: Nitin Kumbhar <nkumbhar@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1730536
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Earlier implementation of railgate disable config is disabling
runtime pm during pm_init. This is causing multiple issues:
1. gpu rail will be on as soon as nvgpu driver probe is called.
Actual gpu hw init may happen at much later point of time.
2. This is breaking railgate_enable sysfs node functionality.
railgate_enable is not working if runtime pm is disabled.
To avoid all these issues for railgate disable, enable runtime pm
during pm_init and set auto-suspend delay to negative (-1), which
will disable runtime pm suspend calls.
Also fixed following issues along with this:
1. Updated railgate_enable debugfs implementation to use auto-suspend delay.
To disable railgating:
Set auto-suspend delay with negative value(-1) which will disable runtime
pm suspend.
To enable railgating:
Set auto-suspend delay with railgate_delay value.
Also removed redundant user_railgate_disabled gk20a device data and
replaced with can_railgate, where ever it is applicable.
2. Initialized default railgate_delay to 500msec to avoid railgate
on/off transitions with railigate enable from disabled state.
3. Created railgate_residency debug fs node irrespective of can_railgate
initial state. This is helping with the case, where initial state of
railgate state off and then railgate enable is done through sysfs node.
Bug 2073029
Change-Id: I531da6d93ba8907e806f65a1de2a447c1ec2665c
Signed-off-by: seshendra Gadagottu <sgadagottu@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1694944
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Each GPU platform has different DMA limitations. For older
chips the maximum size of a DMA buffer was more limited than
newer SoCs (read: Xavier) and discrete GPUs.
This patch adds support to set the DMA mask for a GPU on a
per platform basis by adding a platform field that is populated
with the maximum allowed DMA mask. That mask is programmed by
the driver common code. If no mask is specified then the
default mask size is 16GB (34 bits).
Bug 2043276
Change-Id: I9c3c76c86bac6c485eb1197326e662516fbcaa41
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1700980
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Add g->fifo_eng_timeout_us to define engine timeout in microseconds.
It is initialized with GRFIFO_TIMEOUT_CHECK_PERIOD_US. In RM server
case, it can be overriden with value defined in device tree.
Jira EVLR-2674
Change-Id: I69ac2ce779fe575566c8ba48e8cd2d0e6b2d93cf
Signed-off-by: Thomas Fleury <tfleury@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1728391
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Add gk20a_fifo_profile_snapshot() to store the submit time in a
profiling entry that was acquired from gk20a_fifo_profile_acquire().
Also get rid of ifdef CONFIG_DEBUG_FS by stubbing the acquire and free
functions when debugfs is not enabled. This reduces some cyclomatic
complexity in the submit path.
Jira NVGPU-708
Change-Id: I39829a6475cfe3aa582620219e420bde62228e52
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1729545
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The number of samples has to be at least the number of percentile ranges
(here 20) for the reporting to work as expected and also to not cause
negative indices in reading the sorted profile data. If there are not
enough samples, just report all zeroes.
Change-Id: Ie893859d95074f5ceabf6abe873941873668861d
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1721892
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The forced PRAMIN reads and writes for sysmem buffers haven't worked in
a while since the PRAMIN access code was refactored to work with
vidmem-only sgt allocs. This feature was only ever meant for testing and
debugging PRAMIN access and early dGPU support, but that is stable
enough now so just delete the broken feature instead of fixing it.
Change-Id: Ib31dae4550f3b6fea3c426a2e4ad126864bf85d2
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1723725
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
On Tegra platforms that have clock management under BPMP, and do not
support Tegra DVFS, GPU driver cannot access Fmax@Vmin (get interface
always returns "0"). Added such access through BPMP DVFS shim driver.
Bug 2045903
Change-Id: I0222f2e2917cda15d18ea3296dd1fe53b2ea6b45
Signed-off-by: Alex Frid <afrid@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1722431
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
-timeouts will be enabled only when timeouts_disabled_refcount
will reach 0
-timeouts_enabled debugfs will change from u32 type to file type
to avoid race enabling/disabling timeout from debugfs and ioctl
-unify setting timeouts_enabled from debugfs and ioctl
Bug 1982434
Change-Id: I54bab778f1ae533872146dfb8d80deafd2a685c7
Signed-off-by: Seema Khowala <seemaj@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1588690
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
For CDE work a sync fence is always requested, but kernel does not need
it and submit flags from userspace will be passed to the submit function
in cde path so a sync fence will get created if necessary. To reduce
some complexity, remove the explicit boolean in favor of just
NVGPU_SUBMIT_FLAGS_SYNC_FENCE.
Jira NVGPU-705
Change-Id: I8aac85288513ed7cc640acd021d892cee86f41d8
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1721785
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The mechanism of posting events to userspace is OS specific.
In linux this works through poll fd, wherein we can make use
of nvgpu_cond variables to poll and trigger the corresponding
wait_queue.
The post event functionality on QNX doesn't work on poll though.
It uses iofunc_notify_trigger to post the events to the calling
process. As such QNX can't work with nvgpu_cond's.
To overcome this issue, it is proposed to create OS specific
interface function for posting clk arb events. Linux can call
nvgpu_cond based implementation, which makes sense since these
are already initialized and poll'ed in Linux specific code only.
QNX can implement this interface to call iofunc_notify_*
functions, as per its need.
Jira VQRM-3741
Change-Id: I7d9f71dae2ae7f6a09cd56662003fd1b7e50324c
Signed-off-by: Sourab Gupta <sourabg@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1709656
GVS: Gerrit_Virtual_Submit
Reviewed-by: Alex Waterman <alexw@nvidia.com>
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Add NVGPU_IOCTL_CHANNEL_RESCHEDULE_RUNLIST ioctl to reschedule runlist,
and optionally check host and FECS status to preempt pending load of
context not belonging to the calling channel on GR engine during context
switch.
This should be called immediately after a submit to decrease worst case
submit to start latency for high interleave channel.
There is less than 0.002% chance that the ioctl blocks up to couple
miliseconds due to race condition of FECS status changing while being read.
For GV11B it will always preempt pending load of unwanted context since
there is no chance that ioctl blocks due to race condition.
Also fix bug with host reschedule for multiple runlists which needs to
write both runlist registers.
Bug 1987640
Bug 1924808
Change-Id: I0b7e2f91bd18b0b20928e5a3311b9426b1bf1848
Signed-off-by: David Li <davli@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1549050
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Add below new HALs
gops.fifo.add_sema_cmd() to insert HOST semaphore acquire/release methods
gops.fifo.get_sema_wait_cmd_size() to get size of acquire command buffer
gops.fifo.get_sema_incr_cmd_size() to get size of release command buffer
Separate out new API gk20a_fifo_add_sema_cmd() to implement semaphore acquire/
release sequence and set it to gops.fifo.add_sema_cmd()
Add gk20a_fifo_get_sema_wait_cmd_size() and gk20a_fifo_get_sema_incr_cmd_size()
to return respective command buffer sizes
Jira NVGPUT-16
Change-Id: Ia81a50921a6a56ebc237f2f90b137268aaa2d749
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1704490
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
sync_gk20a.* files are no longer used by core code and only invoked
from linux specific implementations of the OS_FENCE framework which are
under the common/linux directory. Hence, sync_gk20a.* files are also
moved under common/linux.
JIRA NVGPU-66
Change-Id: If623524611373d2da39b63cfb3c1e40089bf8d22
Signed-off-by: Debarshi Dutta <ddutta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1712900
Reviewed-by: Vijayakumar Subbu <vsubbu@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Currently gpu sysfs retrieves Fmax@Vmin by direct call into Tegra DVFS
driver that introduces compile time dependencies on CONFIG_TEGRA_DVFS.
In addition incorrect clock is used for DVFS information access.
Re-factored sysfs node to use generic GPU clock operation for Fmax@Vmin
read. This would fix a bug in target clock selection, and allows to
remove dependency of sysfs on CONFIG_TEGRA_DVFS.
Updated nvgpu_linux_get_fmax_at_vmin_safe operation itself so it can be
called on platforms that does not support Tegra DVFS, although 0 will
still be returned as Fmax@Vmin on such platforms.
Bug 2045903
Change-Id: I32cce25320df026288c82458c913b0cde9ad4f72
Signed-off-by: Alex Frid <afrid@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1710924
Reviewed-by: Alex Waterman <alexw@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Replaced all instances of sync_fence in gk20a_fence* code with
nvgpu_os_fence. Added the API install_fence for the nvgpu_os_fence
abstraction. sync_fence mechanism and its dependencies are completely
removed from the fence_gk20a methods. Due to the recent os_fence changes
and the changes to fence_gk20a, we can finally get rid of all the CONFIG_SYNCS
present in the submit path.
JIRA NVGPU-66
Change-Id: I3551dab04b93b1e94db83fc102a41872be89e9ed
Signed-off-by: Debarshi Dutta <ddutta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1701245
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
This patch constructs an abstraction to hide the sync_fence
functionality from the common code. struct nvgpu_os_fence acts as an
abstraction for struct sync_fence.
struct nvgpu_os_fence consists of an ops structure named nvgpu_os_fence_ops
which contains an API to do pushbuffer programming to generate wait
commands for the fence.
The current implementation of nvgpu only allows for wait method on a
sync_fence which was generated using a similar backend(i.e. either
Nvhost Syncpoints or Semaphores). In this patch, a
generic API is introduced which will decide the type of the underlying
implementation of the struct nvgpu_os_fence at runtime and run the
corresponding wait implementation on it.
This patch changes the channel_sync_gk20a's semaphore specific
implementation to use the abstract API. A subsequent patch will make
the changes for the nvhost_syncpoint based implementations as well.
JIRA NVGPU-66
Change-Id: If6675bfde5885c3d15d2ca380bb6c7c0e240e734
Signed-off-by: Debarshi Dutta <ddutta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1667218
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
We so far do not support unbinding a channel from TSG explicitly through IOCTL
Channel is unbound from TSG when channel fd is closed
But this could create issue in case process is forked and fd is duplicated
In worst case it is possible that duplicate fd is closed sometime later when
TSG is active and that could lead to corruption of TSG state
Fix this by implementing NVGPU_TSG_IOCTL_UNBIND_CHANNEL API
This API will simply call gk20a_tsg_unbind_channel() API to unbind the
channel
Note that we also mark the channel has_timedout since unbound channel does
not have any context of its own and it cannot serve any job
Remove call to gk20a_disable_channel() while closing the channel.
We do not support bare channels without TSG, and if channel is not part of TSG
then it means channel is already unbound from TSG with call to
NVGPU_TSG_IOCTL_UNBIND_CHANNEL, and there is nothing to do while closing the
channel
Bug 200327095
Jira NVGPU-229
Change-Id: Ib7f6710f0dc8125caafabe7bee737622c3dd9fa3
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1708902
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Split sim initialization to two parts,
first part gets invoked as part of probe and
second part gets invoked in the finalize_poweron
after the hal has been initialized.
This is done because some of the sim init
code uses mm api's which are assigned as
part of hal init.
replaced sim buffer allocation api's
with nvgpu_dma_sys_alloc.
Change-Id: Ib019fbb747bdf6dcc74e7deba732ab41f0869e96
Signed-off-by: Antony Clince Alex <aalex@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1705424
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Switch all logging to nvgpu_log*(). gk20a_dbg* macros are
intentionally left there because of use from other repositories.
Because the new functions do not work without a pointer to struct
gk20a, and piping it just for logging is excessive, some log messages
are deleted.
Change-Id: I00e22e75fe4596a330bb0282ab4774b3639ee31e
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1704148
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
segregated os-agnostic function from linux/sim.c and linux/sim_pci.c
to sim.c and sim_pci.c, while retaining os-specific functions.
renamed all gk20a_* api's to nvgpu_*.
renamed hw_sim_gk20a.h to nvgpu/hw_sim.h
moved hw_sim_pci.h to nvgpu/hw_sim_pci.h
JIRA VQRM-2368
Change-Id: I040a6b12b19111a0b99280245808ea2b0f344cdd
Signed-off-by: Antony Clince Alex <aalex@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1702425
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
PMU ucode is updated to include LDIV slowdown factor in gr_init_param command.
- Defined a new version gr_init_param_v2.
- Updated the PMU FW version code.
- Set the LDIV slowdown factor to 0x1e by default.
- Added sysfs entry to program ldiv_slowdown factor at runtime.
Bug 200391931
Change-Id: Ic66049588c3b20e934faff3f29283f66c30303e4
Signed-off-by: Deepak Goyal <dgoyal@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1674208
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
