The message to tell RM server to unbind channel has to be sent after
client unbinds the channel and before client calls tsg release. The
channel has to belong to a tsg on RM server before client submit a
runlist to remove the channel. Or there's a bare channel problem.
By moving .tsg_unbind_channl one layer lower, gk20a_tsg_unbind_channel()
will be common functions for all chip, and it'll call tsg release after
call .tsg_unbind_channel. So vgpu won't need to worry about tsg was
released before sending msg to RM server.
Bug 200382695
Bug 200382785
Change-Id: I32acc122f3f9d5d0628049ccf673225f9e90c87a
Signed-off-by: Richard Zhao <rizhao@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1645383
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: Konsta Holtta <kholtta@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>
Simplify the copyengine code massively by storing the job post fence
pointers in an array of fences instead of mixing them up in the command
buffer memory. The post fences are used when the ring buffer of a
context gets full and we need to wait for the oldest slot to free up.
NVGPU-43
NVGPU-52
Change-Id: I36969e19676bec0f38de9a6357767a8d5cbcd329
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1646037
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
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>
Adds the skeleton and integration of the GV100 endpoint driver to NVGPU
(1) Adds a OS abstraction layer for the internal nvlink structure.
(2) Adds linux specific integration with Nvlink core driver.
(3) Adds function pointers for nvlink api, initialization and isr process.
(4) Adds initial support for minion.
(5) Adds new GPU enable properties to handle NVLINK presence
(6) Adds new GPU enable properties for SG_PHY bypass (required for NVLINK over
PCI)
(7) Adds parsing of nvlink vbios structures.
(8) Adds logging defines for NVGPU
JIRA: EVLR-2328
Change-Id: I0720a165a15c7187892c8c1a0662ec598354ac06
Signed-off-by: David Nieto <dmartineznie@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1644708
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Data can be speculatively loaded from memory and stay in cache even
when bound check fails. This can lead to unintended information
disclosure via side-channel analysis.
To mitigate this problem insert a speculation barrier.
bug 2039126
CVE-2017-5753
Change-Id: Ib6c4b2f99b85af3119cce3882fe35ab47509c76f
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1640500
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Lots of code paths were split to T19x specific code paths and structs
due to split repository. Now that repositories are merged, fold all of
them back to main code paths and structs and remove the T19x specific
Kconfig flag.
Change-Id: Id0d17a5f0610fc0b49f51ab6664e716dc8b222b6
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1640606
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Added sw method support for SET_BES_CROP_DEBUG4.
In this sw method:
CLAMP_FP_BLEND_TO_MAXVAL forces overflow and
CLAMP_FP_BLEND_TO_INF blend results to clamp to FP maxval.
Added support for this sw method in gp10b/gp106/gv11b
and gv100.
Bug 2046636
Change-Id: I3a9e97587aca76718f7f504ea3b853f87409092a
Signed-off-by: seshendra Gadagottu <sgadagottu@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1641529
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The idling and unidling of deterministic channels in the
do_idle/do_unidle path assume that each deterministic channel holds a
power reference. This is no longer the case if railgating has been
allowed for a channel via the deterministic options ioctl which also
causes the channel to drop the power ref that it holds otherwise during
its lifetime.
All this is happening inside the deterministic_busy rwsem, which also
guards the ioctl changing those deterministic option states.
Bug 200327089
Change-Id: I9ce312bbaa459b3cf4a7541fa369186b78c3afdc
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1642310
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
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>
Fix a race condition where we'd still be booting up the gpu and/or
initializing the driver but elsewhere assume that all is done already.
Some userspace APIs to make sure that we're ready by testing
g->gr.sw_ready, but this flag is set in the middle of bootup; there are
other things after gr initialization. Add a new flag that is enabled
after bootup is fully complete at the end of finalize_poweron, and
change the checks in user API paths to test the new flag only.
These checks are only in the ioctl paths for ctrl, dbg and tsg, and in
the ctrl device's opening path.
The gr.sw_ready flag is still left there to signify whether just gr has
had its bookkeeping initialized.
Bug 200370011
Change-Id: I2995500e06de46430d9b835de1e9d60b3f01744e
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1640124
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Add HAL for dumping ctxsw statistics. The statistics are dependent on
the architecture, and the function that calls this operation needs to
be moved to gk20a.
Bug 1842197
Change-Id: I285c74b8ddc8c7854c85b3fef4cbfc582098919e
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1632681
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Seshendra Gadagottu <sgadagottu@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
On gv11b we can have multiple SMs per TPC. Add sm_per_tpc in
vgpu constants to properly dimension the virtual SM to TPC/GPC
mapping in virtualization case.
Use TEGRA_VGPU_CMD_GET_SMS_MAPPING to query current mapping.
Bug 2039676
Change-Id: I817be18f9a28cfb9bd8af207d7d6341a2ec3994b
Signed-off-by: Thomas Fleury <tfleury@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1631203
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Used chip specific attrib_cb_gfxp_default_size and
attrib_cb_gfxp_size buffer sizes during committing
global callback buffer when gfx preemption is requested.
These sizes are different for gv11b from gp10b.
For gp10b used smaller buffer sizes than specified
value in hw manuals as per sw requirement.
Also used gv11b specific preemption related functions:
gr_gv11b_set_ctxsw_preemption_mode
gr_gv11b_update_ctxsw_preemption_mode
This is required because preemption related buffer
sizes are different for gv11b from gp10b. More optimization
will be done as part of NVGPU-484.
Another issue fixed is: gpu va for preemption buffers
still needs to be 8 bit aligned, even though 49 bits
available now. This done because of legacy implementation
of fecs ucode.
Bug 1976694
Change-Id: I2dc923340d34d0dc5fe45419200d0cf4f53cdb23
Signed-off-by: seshendra Gadagottu <sgadagottu@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1635027
GVS: Gerrit_Virtual_Submit
Reviewed-by: Richard Zhao <rizhao@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>
This reverts commit caf168e33e.
Might be causing an intermittency in quill-c03 graphics submit. Super
weird since the only change that seems like it could affect it is the
header file update but that seems rather safe.
Bug 2044830
Change-Id: I14809d4945744193b9c2d7729ae8a516eb3e0b21
Signed-off-by: Alex Waterman <alexw@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1634349
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: Timo Alho <talho@nvidia.com>
Tested-by: Timo Alho <talho@nvidia.com>
Used chip specific attrib_cb_gfxp_default_size and
attrib_cb_gfxp_size buffer sizes during committing
global callback buffer when gfx preemption is requested.
These sizes are different for gv11b from gp10b.
Also used gv11b specific preemption related functions:
gr_gv11b_set_ctxsw_preemption_mode
gr_gv11b_update_ctxsw_preemption_mode
This is required because preemption related buffer
sizes are different for gv11b from gp10b. More optimization
will be done as part of NVGPU-484.
Another issue fixed is: gpu va for preemption buffers
still needs to be 8 bit aligned, even though 49 bits
available now. This done because of legacy implementation
of fecs ucode.
Bug 1976694
Change-Id: I284e29e0815d205c150998b07d0757b5089d3267
Signed-off-by: seshendra Gadagottu <sgadagottu@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1630520
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-by: Richard Zhao <rizhao@nvidia.com>
Tested-by: Richard Zhao <rizhao@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
During bringup and before nvlink is up GV100 on the DDPX platform operates
with a very, very slow sysmem link. In order to get sysmem test to pass
it is neccesary to significantly increase most timeouts by an order the
magnitude.
Bug 2040544
Change-Id: I26858afde4ae80c70f86b47cfff674b6b00b5bf8
Signed-off-by: David Nieto <dmartineznie@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1627417
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
With a recent rework we moved gk20a_get() call to nvgpu_ioctl_tsg_open(),
but corresponding gk20a_put() call remained in gk20a_tsg_release()
So if a TSG is opened and released from within kernel with APIs
gk20a_tsg_open()/gk20a_tsg_release() we mistakenly drop extra refcount
through gk20a_put()
Fix this by moving gk20a_put() call to nvgpu_ioctl_tsg_release() which
balances gk20a_get() call in nvgpu_ioctl_tsg_open()
Bug 200374011
Change-Id: Id0cec0426e6231309dc530ab5c934dacaba9f8da
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1630969
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@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>
In gk20a_ce_create_context(), if gk20a_tsg_open() or gk20a_open_new_channel()
fails, we bail out from the function without setting the error code
This could mislead the caller and report incorrect success
Fix this by setting error code explicitly in failure cases
Bug 200374011
Change-Id: Idf6cba4a57740107bada698295745352f7b5d5ac
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1631506
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@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>
In gk20a_ce_delete_gpu_context(), we unbind the channel from TSG and
close the channel. But we do not drop the TSG refcount leaking
the TSG reference
Fix this by explicitly dropping TSG refcount
Also, do not explicitly unbind the channel from TSG
gk20a_channel_close() will internally unbind the channel from TSG
Bug 200374011
Change-Id: Ie4aa32f1d0bff4231f41aa2b33743cdc63e967c7
Signed-off-by: Deepak Nibade <dnibade@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1629972
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@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>
Remove nvgpu internal flag indicating that TSGs are required. We now
require TSGs always. This also fixes a regression where CE channels
were back to using bare channels on gp106.
Bug 1842197
Change-Id: Id359e5a455fb324278636bb8994b583936490ffd
Signed-off-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1628481
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Deepak Nibade <dnibade@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: Seshendra Gadagottu <sgadagottu@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
This CL handles
- erroneous use of boot_0 function pointer
before being assigned in __nvgpu_check_gpu_state
- And proper handling of error returned from gk20a_readl
in gk20a_mc_boot_0
With these fixes crash is not seen in case mc_boot_0 read
returns 0 in gk20a_mc_boot_0
- And also this handles the recursion caused by mc.boot_0()
calling nvgpu_readl and nvgpu_readl in turn
calling mc.boot_0 in case of read failure
Bug 2010966
Change-Id: Ia087811c67d88948b7fc5fff35e0fabc6ea91989
Signed-off-by: Supriya <ssharatkumar@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1616274
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
During finalize power on, resume channels only after
complete hw initialization is done. Otherwise it will
cause issues with unexpected usage of hw. During first
boot will not see these issues because there will no channels.
But after rail gate/ungate or suspend/resume these issues
can be seen if channels are present before rail-gate/suspend.
Bug 2039195
Change-Id: Ie96e2f2b91902ba18b37e9a167344eeae07ba8c2
Signed-off-by: seshendra Gadagottu <sgadagottu@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1625506
Reviewed-by: Automatic_Commit_Validation_User
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>
Check the availability of ecc units by checking
relevant ecc fuse and fuse overrides.
During gpu boot, initialize ecc units by scrubbing
individual ecc units available. ECC initialization
should be done before gr initialization.
Following ecc units are scrubbed:
SM LRF
SM L1 DATA
SM L1 TAG
SM CBU
SM ICACHE
Bug 200339497
Change-Id: I54bf8cc1fce639a9993bf80984dafc28dca0dba3
Signed-off-by: Deepak Goyal <dgoyal@nvidia.com>
Signed-off-by: seshendra Gadagottu <sgadagottu@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1612734
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
