Commit 69c4938249 ("drm/nouveau/instmem/gk20a: use direct CPU access")
tried to be smart while using the DMA-API by managing the CPU mappings of
buffers allocated with the DMA-API by itself. In doing so, it relied
on dma_to_phys() which is an architecture-private function not
available everywhere. This broke the build on several architectures.
Since there is no reliable and portable way to obtain the physical
address of a DMA-API buffer, stop trying to be smart and just use the
CPU mapping that the DMA-API can provide. This means that buffers will
be CPU-mapped for all their life as opposed to when we need them, but
anyway using the DMA-API here is a fallback for when no IOMMU is
available so we should not expect optimal behavior.
This makes the IOMMU and DMA-API implementations of instmem diverge
enough that we should maybe put them into separate files...
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
The LRU list used for recycling CPU mappings was handling concurrency
very poorly. For instance, if an instobj was acquired twice before being
released once, it would end up into the LRU list even though there is
still a client accessing it.
This patch fixes this by properly counting how many clients are
currently using a given instobj.
While at it, we also raise errors when inconsistencies are detected, and
factorize some code.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
The offset should be 8 on Kepler and later.
Signed-off-by: Vince Hsu <vinceh@nvidia.com>
Reviewed-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Very rough, no idea how correct it is at this point, but it prevents
getteximage-depth from piglit from hanging the GPU.
v2: updated with NV_PCE_FE_LAUNCHERR_REPORT values provided by NVIDIA
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
The CPU-side tracking of engine runlists was not protected by a lock,
leading to list corruption, eventually causing runlist_update() to
overrun the GPU-side runlist, triggering an OOPS.
Fixes some of the issues noticed during parallel piglit runs.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
We want to unlock nv_devices_mutex in this error path as well.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Without this patch a pixel clock rate above 165 MHz on a TMDS link is
assumed to be dual link. This is true for DVI, but not for HDMI. HDMI
supports no dual link, but it supports pixel clock rates above 165 MHz.
Only activate Dual Link mode when it is actually possible and requested.
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
[imirkin: check for hdmi monitor for computing proto, use sor ctrl to
enable extra config bit]
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This is an oversight that made use of the trip-point-based fan managenent on
cards that never expose those. This led the fan to stay at fan_min.
Fortunately, the emergency code would kick when the temperature would reach
90°C.
Reported-by: Tom Englund <tomenglund26@gmail.com>
Tested-by: Tom Englund <tomenglund26@gmail.com>
Signed-off-by: Martin Peres <martin.peres@free.fr>
Tested-by: Daemon32 <lnf.purple@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92126
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Cc: stable@vger.kernel.org
Just the one commit I mentioned earlier, making the PGOB workaround the
default.
* 'linux-4.4' of https://github.com/skeggsb/linux:
drm/nouveau/pmu: remove whitelist for PGOB-exit WAR, enable by default
NVIDIA have indicated that the workaround is required on all GK10[467]
boards that have the PGOB fuse set.
I've left the commandline option in place for now, as paranoia.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Ben Skeggs wrote:
A couple of regression fixes, some more boards whitelisted for a hw bug
workaround, gr/ucode fixes for hangs a user is seeing.
The changes look larger than they actually are due to the ucode binaries
(*.fucN.h) being regenerated.
* 'linux-4.4' of git://anongit.freedesktop.org/git/nouveau/linux-2.6:
drm/nouveau/volt/pwm/gk104: fix an off-by-one resulting in the voltage not being set
drm/nouveau/nvif: allow userspace access to its own client object
drm/nouveau/gr/gf100-: fix oops when calling zbc methods
drm/nouveau/gr/gf117-: assume no PPC if NV_PGRAPH_GPC_GPM_PD_PES_TPC_ID_MASK is zero
drm/nouveau/gr/gf117-: read NV_PGRAPH_GPC_GPM_PD_PES_TPC_ID_MASK from correct GPC
drm/nouveau/gr/gf100-: split out per-gpc address calculation macro
drm/nouveau/bios: return actual size of the buffer retrieved via _ROM
drm/nouveau/instmem: protect instobj list with a spinlock
drm/nouveau/pci: enable c800 magic for some unknown Samsung laptop
drm/nouveau/pci: enable c800 magic for Clevo P157SM
Each GPCCS unit was reading the mask from GPC0, which causes problems on
boards where some GPCs are missing PPCs.
Part of the fix for fdo#92761.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
There's a few places where we need to access a GPC register from ucode,
but outside of the falcon's io address space. To do this we need to
calculate the offset based on which GPC we're executing on.
This used to be done manually, but we've since found a "base" offset
that can be added by the hardware. To use this, an extra bit needs to
be set in the register address, which is what this macro achieves.
There should be no functional change from this commit.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
No locking is required for the traversal of this list, as it only
happens during suspend/resume where nothing else can be executing.
Fixes some of the issues noticed during parallel piglit runs.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
gk20a is an ARM only GPU, so we can just do the correct thing on
ARM but fail on other architectures. The other option was to use
SWIOTLB as the define, which means phys_to_page exists, but
this seems clearer.
Signed-off-by: Dave Airlie <airlied@redhat.com>
This patch uses an approach closer to the nvidia driver to configure
both PLLs for high gddr5 memory clocks (usually above 2400MHz)
Previously nouveau used the one PLL as it was used for the lower clocks
and just adjusted the second PLL to get as close as possible to the
requested clock. This means for my card, that I got a 4050 MHz clock
although 4008 MHz was requested.
Now the driver iterates over a list of PLL configuration also used by
the nvidia driver and then adjust the second PLL to get near the
requested clock. Also it hold to some restriction I found while
analyzing the PLL configurations
This won't fix all gddr5 high clock issues itself, but it should be
fine on hybrid gpu systems as found on many laptops these days. Also
switching while normal desktop usage should be a lot more stable than
before.
v2: move the pll code into ramgk104
Signed-off-by: Karol Herbst <nouveau@karolherbst.de>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Your milage may vary, as it's only been tested on a single G94 and one G96.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Avoids waiting for VBLANKS that never arrive on headless or otherwise
unconventional set-ups. Strategy taken from MEMX.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
10053c is not even read on some cards, and I have no idea exactly what the
criteria are. Likely NVIDIA pre-scans the VBIOS and in their driver disables
all features that are never used. The practical effect should be the same
as this implementation though.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Like Pierre's G94. We might want to structure Kepler similarly in a follow-up.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Does not seem to be necessary for NVA0, hence untested by me.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Seems to be mostly equal to DDR3 on < GT218, should improve stability for
DDR2 reclocks.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
In preparation of changing FBVDDQ, as observed on at least one GDDR3 card.
While at it, adhere to func.log[1] properly for consistency.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
If the hardware supports extended tag field (8-bit ones), then enable it.
This is usually done by the VBIOS, but not on some MBPs (see fdo#86537).
In case extended tag field is not supported, 5-bit tag field is used which
limits the possible number of requests to 32. Apparently bits 7:0 of
0x08841c stores some number of outstanding requests, so cap it to 32 if
extended tag is unsupported.
Fixes: fdo#86537
v2: Restrict changes to chipsets >= 0x84
v3:
* Add nvkm_pci_mask to pci.h
* Mask bit 8 before setting it
v4:
* Rename `add` argument of nvkm_pci_mask to `value`
* Move code from nvkm_pci_init to g84_pci_init and remove PCIe and chipset
checks
v5:
* Rebase code on latest PCI structure
* Restore PCIe check
* Fix namings in nvkm_pci_mask
* Rephrase part of the commit message
Signed-off-by: Pierre Moreau <pierre.morrow@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
These nvkm_object_func structures are never modified. All other
nvkm_object_func structures are declared as const.
Done with the help of Coccinelle.
Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
GF110+ supports both the A and B compute classes, make sure to accept
both.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
NVIDIA provided the documentation for mp error 0x10, INVALID_ADDR_SPACE,
which apparently happens when trying to use an atomic operation on
local or shared memory (instead of global memory).
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
coverity.com reported that memset was using a buffer of size 0, on
checking the code it turned out that the function was not being used. So
remove it.
Signed-off-by: Sudip Mukherjee <sudip@vectorindia.org>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Was not able to obtain a trace of NVRM due to kernel version annoyances,
however, experimentally confirmed that the WAR we use on NV50/G8x boards
works here too.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Increase clock timeout of some unknown engines in order to avoid failure
at high gpcclk rate.
This fixes IBUS read faults on my GF119 when reclocking is manually
enabled. Note that memory reclocking is completely broken and NvMemExec
has to be disabled to allow core clock reclocking only.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
I got confirmation that we can read and change the voltage with the same code.
The divider is also computed correctly on the gm204 we got our hands on.
Thanks to Yoshimo on IRC for executing the tests on his gm204!
Signed-off-by: Martin Peres <martin.peres@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Let's ignore the other desktop Maxwells until I get my hands on one and confirm
that we still can change the voltage.
Signed-off-by: Martin Peres <martin.peres@free.fr>
Most Keplers actually use the GPIO-based voltage management instead of the new
PWM-based one. Use the GPIO mode as a fallback as it already gracefully handles
the case where no GPIOs exist.
All the Maxwells seem to use the PWM method though.
v2:
- Do not forget to commit the PWM configuration change!
Signed-off-by: Martin Peres <martin.peres@free.fr>
This patch is not ideal but it definitely beats a rewrite of the current
interface and is very self-contained.
Signed-off-by: Martin Peres <martin.peres@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Use the IOMMU bit specified in platform data instead of hardcoding it to
the bit used by current Tegra GPUs.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Current Tegra code taking advantage of the IOMMU assumes a hardcoded
value for the IOMMU bit. Make it a platform property instead for
flexibility.
v2 (Ben Skeggs): remove nvkm dependence on drm structures
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
The Great Nouveau Refactoring Take II brought us a lot of goodness,
including acquire/release methods that are called before and after an
instobj is modified. These functions can be used as synchronization
points to manage CPU/GPU coherency if we modify an instobj using the
CPU.
This patch replaces the legacy and slow PRAMIN access for gk20a instmem
with CPU mappings and writes. A LRU list is used to unmap unused
mappings after a certain threshold (currently 1MB) of mapped instobjs is
reached. This allows mappings to be reused most of the time.
Accessing instobjs using the CPU requires to maintain the GPU L2 cache,
which we do in the acquire/release functions. This triggers a lot of L2
flushes/invalidates, but most of them are performed on an empty cache
(and thus return immediately), and overall context setup performance
greatly benefits from this (from 250ms to 160ms on Jetson TK1 for a
simple libdrm program).
Making L2 management more explicit should allow us to grab some more
performance in the future.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Allow clients to manually flush and invalidate L2. This will be useful
for Tegra systems for which we want to write instmem using the CPU.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
These are useful for systems without a coherent CPU/GPU bus. For such
systems we may need to maintain the L2 ourselves.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Some devices may not have a PMU. Avoid a NULL pointer dereference in
such cases by checking whether the pointer given to nvkm_pmu_pgob() is
valid.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Currently OF bios load fails for a few reasons:
- checksum failure
- bios size too small
- no PCIR header
- bios length not a multiple of 4
In this change, we resolve all of the above by ignoring any checksum
failures (since OF VBIOS tends not to have a checksum), and faking the
PCIR data when loading from OF.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
SiS 761 chipset does not support AGP cards but has AGP capability (for
the onboard video). At least PC Chips A31G board using this chipset has
an AGP-like AGPro slot that's wired to the PCI bus. Enabling AGP will
fail (GPU lockup and software fbcon, X11 hangs).
Add support for matching just the host bridge in nvkm_device_agp_quirks
and add entry for SiS 761 with mode 0 (AGP disabled).
Signed-off-by: Ondrej Zary <linux@rainbow-software.org>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Pull drm fixes from Dave Airlie:
"Just a bunch of fixes to squeeze in before -rc1:
- three nouveau regression fixes
- one qxl regression fix
- a bunch of i915 fixes
... and some core displayport/atomic fixes"
* 'drm-fixes' of git://people.freedesktop.org/~airlied/linux:
drm/nouveau/device: enable c800 quirk for tecra w50
drm/nouveau/clk/gt215: Unbreak engine pausing for GT21x/MCP7x
drm/nouveau/gr/nv04: fix big endian setting on gr context
drm/qxl: validate monitors config modes
drm/i915: Allow DSI dual link to be configured on any pipe
drm/i915: Don't try to use DDR DVFS on CHV when disabled in the BIOS
drm/i915: Fix CSR MMIO address check
drm/i915: Limit the number of loops for reading a split 64bit register
drm/i915: Fix broken mst get_hw_state.
drm/i915: Pass hpd_status_i915[] to intel_get_hpd_pins() in pre-g4x
uapi/drm/i915_drm.h: fix userspace compilation.
drm/i915: Always mark the object as dirty when used by the GPU
drm/dp: Add dp_aux_i2c_speed_khz module param to set the assume i2c bus speed
drm/dp: Adjust i2c-over-aux retry count based on message size and i2c bus speed
drm/dp: Define AUX_RETRY_INTERVAL as 500 us
drm/atomic: Fix bookkeeping with TEST_ONLY, v3.
three nouveau regression fixes.
* 'linux-4.3' of git://anongit.freedesktop.org/git/nouveau/linux-2.6:
drm/nouveau/device: enable c800 quirk for tecra w50
drm/nouveau/clk/gt215: Unbreak engine pausing for GT21x/MCP7x
drm/nouveau/gr/nv04: fix big endian setting on gr context
Typo that snuck in with commit 6979c6303a
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Reported-by: Pierre Moreau <pierre.morrow@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Broken since "gr: convert user classes to new-style nvkm_object"
Tested on a PPC64 G5 + NV34
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Pull drm updates from Dave Airlie:
"This is the main pull request for the drm for 4.3. Nouveau is
probably the biggest amount of changes in here, since it missed 4.2.
Highlights below, along with the usual bunch of fixes.
All stuff outside drm should have applicable acks.
Highlights:
- new drivers:
freescale dcu kms driver
- core:
more atomic fixes
disable some dri1 interfaces on kms drivers
drop fb panic handling, this was just getting more broken, as more locking was required.
new core fbdev Kconfig support - instead of each driver enable/disabling it
struct_mutex cleanups
- panel:
more new panels
cleanup Kconfig
- i915:
Skylake support enabled by default
legacy modesetting using atomic infrastructure
Skylake fixes
GEN9 workarounds
- amdgpu:
Fiji support
CGS support for amdgpu
Initial GPU scheduler - off by default
Lots of bug fixes and optimisations.
- radeon:
DP fixes
misc fixes
- amdkfd:
Add Carrizo support for amdkfd using amdgpu.
- nouveau:
long pending cleanup to complete driver,
fully bisectable which makes it larger,
perfmon work
more reclocking improvements
maxwell displayport fixes
- vmwgfx:
new DX device support, supports OpenGL 3.3
screen targets support
- mgag200:
G200eW support
G200e new revision support
- msm:
dragonboard 410c support, msm8x94 support, msm8x74v1 support
yuv format support
dma plane support
mdp5 rotation
initial hdcp
- sti:
atomic support
- exynos:
lots of cleanups
atomic modesetting/pageflipping support
render node support
- tegra:
tegra210 support (dc, dsi, dp/hdmi)
dpms with atomic modesetting support
- atmel:
support for 3 more atmel SoCs
new input formats, PRIME support.
- dwhdmi:
preparing to add audio support
- rockchip:
yuv plane support"
* 'drm-next' of git://people.freedesktop.org/~airlied/linux: (1369 commits)
drm/amdgpu: rename gmc_v8_0_init_compute_vmid
drm/amdgpu: fix vce3 instance handling
drm/amdgpu: remove ib test for the second VCE Ring
drm/amdgpu: properly enable VM fault interrupts
drm/amdgpu: fix warning in scheduler
drm/amdgpu: fix buffer placement under memory pressure
drm/amdgpu/cz: fix cz_dpm_update_low_memory_pstate logic
drm/amdgpu: fix typo in dce11 watermark setup
drm/amdgpu: fix typo in dce10 watermark setup
drm/amdgpu: use top down allocation for non-CPU accessible vram
drm/amdgpu: be explicit about cpu vram access for driver BOs (v2)
drm/amdgpu: set MEC doorbell range for Fiji
drm/amdgpu: implement burst NOP for SDMA
drm/amdgpu: add insert_nop ring func and default implementation
drm/amdgpu: add amdgpu_get_sdma_instance helper function
drm/amdgpu: add AMDGPU_MAX_SDMA_INSTANCES
drm/amdgpu: add burst_nop flag for sdma
drm/amdgpu: add count field for the SDMA NOP packet v2
drm/amdgpu: use PT for VM sync on unmap
drm/amdgpu: make wait_event uninterruptible in push_job
...
The copyright header in nvkm/engine/device/platform.c has been replaced
with the NVIDIA one from drm/nouveau_platform.c, as most of the actual
code is now theirs.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Doesn't fix any known issue, but best be safe in case control is handed
to us from firmware with these left enabled.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This ensures we have a valid mask of disabled engines before we start
trying to execute fini()/init() on the subdevs, potentially touching
devices that don't exist.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
An upcoming commit requires being able to modify the PRAMIN BAR page
tables while already holding the MMU subdev mutex.
To solve this issue, each VM has been given its own mutex. As a nice
side-effect, this also allows separate VMs to be updated concurrently.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Replaces the piece-by-piece (in response to NV_DEVICE ctor args) device
contruction with a once-off all-or-nothing approach, eliminating some
tricky refcounting issues. The partial device init capability was only
required by some tools, and has been moved to probe time instead.
Temporarily removes a workaround for some boards where we need to fiddle
with AGP registers before executing the DEVINIT scripts. A later commit
in this series reinstates it.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
These require an explicit pointers to nvkm_object/nvkm_subdev/nvkm_device,
depending on which macros are used. This is unlike the previous macros
which take a void *, and work for anything derived from nvkm_object (by
way of some awful heuristics).
The output will be a bit confused until everything has been transitioned,
as the logging format used is a more standard style that previously.
In addition, usage of pr_cont(), which doesn't work correctly with the
dev_*() printk functions (and was potentially racy to begin with), will
be replaced.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
These require an explicit struct nvkm_device pointer, unlike the previous
macros which take a void *, and work for (almost) anything derived from
nvkm_object by using some heuristics.
These macros are more general than the previous ones, and can be used to
handle PTIMER-based busy-waits (will be used in later devinit fixes) as
well as more complicated wait conditions.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Pretty much every subdev/engine is going to need access to nvkm_device
shortly to touch registers and/or output messages.
The odd placement of the includes is necessary to work around some
inter-dependencies that currently exist. This will be fixed later.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
A future commit will hide the platform/pci specifics from nvkm_device,
but it's still very useful in a lot of places to have access to the
Linux device struct.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Will be used in upcoming commits to remove the need for lookup/runtime
type-checking functions when accessing foreign subdevs.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Will be utilised in upcoming commits to remove the need for heuristics
to lookup the device a subdev belongs to.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Only a handful of machines have this enabled by default, where it's been
proven to work. The workaround can be explicitly enabled with a module
option also.
Still waiting on feedback from NVIDIA for a proper idea of exactly what
this fix is doing, and how to implement it properly.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
We previously assumed that the values "2" and "4" were new in DCB 4.1,
however, there's at least one GM107 DCB 4.0 board (Quadro K620) that
uses the newer values.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
MSI interrupts appear to not work for nv46 based cards. Change the mc
subdev oclass for these cards from nv44 to nv4c, the nv4c mc code is
identical to the nv44 mc code except that it does not use msi
(it does not define a msi_rearm callback).
BugLink: https://bugs.freedesktop.org/show_bug.cgi?id=90435
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
High level hardware events related to PBFB will monitor all partitions.
While we are at it, fix bitfield for this mux.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This mux only exists on GF108+ (except for GF110 one), but since it is
not used by the userspace we can drop it for now.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Hardware signals index 0x00 are defined for some domains and they have
to be allowed to enable sources like the others.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
I thought that using TPC[0x0] like for G84:GT215 was sufficient on G80,
but it's actually not the case. According to NVIDIA PerfKit on Windows,
we have to configure PGRAPH related muxs on TPC[0x3] for this chipset.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
No known VBIOSes use these, but they are present in the actual VBIOS
table parsing logic. No harm in adding these too.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Recognize GM20B and assign the right engines and subdevs.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Add support for GM20B's graphics engine, based on GK20A. Note that this
code alone will not allow the engine to initialize on released devices
which require PMU-assisted secure boot.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
GK20A's initialization was based on GK104, but differences exist in the
way the initial context is built and the initialization process itself.
This patch follows the same initialization sequence as nvgpu performs
to avoid bad surprises. Since the register bundles initialization also
differ considerably from GK104, the register packs are now loaded from
firmware files, again similarly to what is done with nvgpu.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
NVIDIA will officially start providing GR firmwares through
linux-firmware for GPUs that require it. Change the GR firmware lookup
function to use these files.
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
These signals and sources have been reverse engineered from CUPTI
(Linux). Graphics signals exposed by PerfKit (Windows only) will be
added later. I need to reverse engineer them and it's a bit painful.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This trivial patch makes thing more consistent since hardware signals
names are prefixed by 'pcXX'.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This is going to be very useful for GF100+ because each GPC can
have its own domain of counters.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
According to the tstate calculation in nvkm_clk_tstate(),
the range of tstate is from -(clk->state_nr - 1) to 0,
it mean the tstate is negative value. But in nvkm_pstate_work(),
it use (clk->state_nr - 1 - clk->tstate) to limit pstate,
it's not correct.
This patch fix it to use (clk->state_nr - 1 + clk->tstate) to
limit pstate.
Signed-off-by: Wei Ni <wni@nvidia.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
These signals and sources have been reverse engineered from CUPTI
(Linux). Graphics signals exposed by PerfKit (Windows only) will be
added later. I need to reverse engineer them and it's a bit painful.
This commit also adds a new class for GF108 and GF117.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
These signals and sources have been reverse engineered from NVIDIA
PerfKit (Windows) and CUPTI (Linux), they will be used to build complex
hardware events from the userspace.
This commit also adds a new class for GT200.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Configuring counters from the userspace require the kernel to handle some
logic related to performance counters. Basically, it has to find a free
slot to assign a counter, to handle extra counting modes like B4/B6 and it
must return and error when it can't configure a counter.
In my opinion, the kernel should not handle all of that logic but it
should only write the configuration sent by the userspace without
checking anything. In other words, it should overwrite the configuration
even if it's already counting and do not return any errors.
This patch allows the userspace to configure a domain instead of
separate counters. This has the advantage to move all of the logic to
the userspace.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This adds a new method NVIF_PERFCTR_V0_INIT which starts a batch of
hardware counters for sampling. This will allow the userspace to start
a monitoring session using the INIT method and to stop it with SAMPLE,
for example before and after a frame is rendered.
This commit temporarily breaks nv_perfmon but this is going to be fixed
with the upcoming patch.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This allows to query the ID, the mask and the user-readable name of
sources for each signal.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
A source (or multiplexer) is a tuple addr+mask+shift which allows to
control a block of signals. The maximum number of sources that a signal
can define is arbitrary limited to 8 and this should be large enough.
This patch allows to define multi-level of sources for a signal.
Each different sources are stored to a global list and will be exposed
to the userspace through the nvif interface in order to avoid conflicts.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This signal index must be always allowed even if it's not clearly
defined in a domain in order to monitor a counter like 0x03020100
because it's the default value of signals.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
16 bits is large enough to store the maximum number of signals available
for one domain (i.e. 256).
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This will allow to configure performance counters with hardware signal
indexes instead of user-readable names in an upcoming patch.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This allows to query the number of available domains, including the
number of hardware counter and the number of signals per domain.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Since a new class has been introduced to query signals, we can now
return an error when the userspace wants to monitor unknown signals.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This commit introduces the NVIF_IOCTL_NEW_V0_PERFMON class which will be
used in order to query domains, signals and sources. This separates the
querying and the counting interface.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Martin Peres <martin.peres@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
PDAEMON signals don't have to be exposed by the perfmon engine.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Martin Peres <martin.peres@free.fr>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Tested on a few cards. Probably works quite well for most, given they should
all be GDDR3.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This looks surprisingly similar to scripts on earlier cards as well
but they don't seem to work just yet. That... and I don't have any, which
makes it a tough job to reverse engineer.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Some of the bits in there are similar to the bits in the gt215 rammap.
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Might need some generalisation to < GT200. For those: use at your own risk!
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
In preparation of NV50 reclocking, where there is no version
Signed-off-by: Roy Spliet <rspliet@eclipso.eu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This reverts commit 1addc12648
This commit seems to cause crashes in gk104_fifo_intr_runlist() by
returning 0xbad0da00 when register 0x2a00 is read. Since this commit was
intended for GM20B which is not completely supported yet, let's revert
it for the time being.
Reported-by: Eric Biggers <ebiggers3@gmail.com>
Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
Tested-by: Afzal Mohammed <afzal.mohd.ma@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
More analysis shows that this is identical to 0x79 except that it loads
the frequency indirectly from elsewhere in the VBIOS.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91025
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Opcode 0x5a is a register write for data looked up from another part of
the VBIOS image. 0x59 is a more complex opcode, but we may as well
recognize it. These occur on a single known instance of Riva TNT2
hardware.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91025
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
