The current submission opcode sequence first takes the engine MLOCK,
and then switches to HOST1X class to wait prefences. This is fine
while we only use a single channel per engine and there is no
virtualization, since jobs are serialized on that one channel anyway.
However, when that assumption doesn't hold, we are keeping the
engine locked while not running anything on it while waiting for
prefences to complete.
To resolve this, execute wait commands in the beginning of the job
outside the engine MLOCK. We still take the HOST1X MLOCK because
recent hardware requires register opcodes to be executed within some
MLOCK, but the hardware also allows unlimited channels to take the
HOST1X MLOCK at the same time.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Link: https://lore.kernel.org/r/20250708-host1x-wait-prefences-outside-mlock-v1-1-13e98044e35a@nvidia.com
Use the newly implemented tegra_dev_iommu_get_stream_id() helper to
encapsulate and centralize the IOMMU stream ID access.
Signed-off-by: Thierry Reding <treding@nvidia.com>
Currently all fences have a 30 second timeout to ensure they are
cleaned up if the fence never completes otherwise. However, this
one size fits all solution doesn't actually fit in every case,
such as syncpoint waiting where we want to be able to have timeouts
longer than 30 seconds. As such, we want to be able to give control
over fence cancellation to the caller (and maybe eventually get rid
of the internal timeout altogether).
Here we add this cancellation mechanism by essentially adding a
function for entering the timeout path by function call, and changing
the syncpoint wait function to use it.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Move from the old, complex intr handling code to a new implementation
based on dma_fences. While there is a fair bit of churn to get there,
the new implementation is much simpler and likely faster as well due
to allowing signaling directly from interrupt context.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
In anticipation of removal of the intr API, implement job tracking
using DMA fences instead. The main two things about this are
making cdma_update schedule the work since fence completion can
now be called from interrupt context, and some complication in
ensuring the callback is not running when we free the fence.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
The code to write the syncpoint channel assignment register
incorrectly skips the write if hypervisor registers are not available.
The register, however, is within the guest aperture so remove the
check and assign syncpoints properly even on virtualized systems.
Fixes: c3f52220f2 ("gpu: host1x: Enable Tegra186 syncpoint protection")
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
On Tegra186+, the syncpoint ID has 10 bits of space. To allow
using more than 256 syncpoints, fix the mask.
Fixes: 9abdd497cd ("gpu: host1x: Tegra234 device data and headers")
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
During the refactoring of channel_submit(), assignment of syncval was
moved but it is also used in channel_submit(). Add this assignment back
to channel_submit() as well.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
With the full-featured opcode sequence using MLOCKs, we need to also
unlock those MLOCKs in the event of a timeout. However, it turns out
that on Tegra186/Tegra194, by default, we don't need to do this;
furthermore, on Tegra234 it is much simpler to do; so only implement
this on Tegra234 for the time being.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
For new (Tegra186+) SoCs, use a new ('full-featured') job opcode
sequence that is compatible with virtualization. In particular,
the Host1x hardware in Tegra234 is more strict regarding the sequence,
requiring ACQUIRE_MLOCK-SETCLASS-SETSTREAMID opcodes to occur in
that sequence without gaps (except for SETPAYLOAD), so let's do it
properly in one go now.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
On Tegra234, each Host1x VM has 8 interrupt lines. Each syncpoint
can be configured with which interrupt line should be used for
threshold interrupt, allowing for load balancing.
For now, to keep backwards compatibility, just set all syncpoints
to the first interrupt.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Host1x class information and opcodes are unchanged or backwards
compatible across SoCs so let's not duplicate them for each one
but have them in a shared header file.
At the same time, add opcode functions for acquire/release_mlock.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Add code to do stream ID switching at the beginning of a job. The
stream ID is switched to the stream ID specified by the context
passed in the job structure.
Before switching the stream ID, an OP_DONE wait is done on the
channel's engine to ensure that there is no residual ongoing
work that might do DMA using the new stream ID.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Add runtime PM and OPP support to the Host1x driver. For the starter we
will keep host1x always-on because dynamic power management require a major
refactoring of the driver code since lot's of code paths are missing the
RPM handling and we're going to remove some of these paths in the future.
Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Tested-by: Peter Geis <pgwipeout@gmail.com> # Ouya T30
Tested-by: Paul Fertser <fercerpav@gmail.com> # PAZ00 T20
Tested-by: Nicolas Chauvet <kwizart@gmail.com> # PAZ00 T20 and TK1 T124
Tested-by: Matt Merhar <mattmerhar@protonmail.com> # Ouya T30
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Show the values of the DMASTART and DMAEND registers when dumping status
to help with failure analysis.
Signed-off-by: Thierry Reding <treding@nvidia.com>
Dumping the full CDMA push buffer takes a long time and isn't very
useful since most of the contents are not relevant. Instead only show
the CDMA push buffer entries associated with current jobs.
While at it, tweak the indentation a bit to make the output more
readable.
Signed-off-by: Thierry Reding <treding@nvidia.com>
The host1x debug code uses a mix of phys_addr_t, dma_addr_t and u32 to
represent addresses. However, these addresses are always DMA addresses
so use the appropriate type.
This fixes some issues with how these addresses are displayed, because
they could be truncated in some cases and not show the full address.
Signed-off-by: Thierry Reding <treding@nvidia.com>
Add support for inserting syncpoint waits in the CDMA pushbuffer.
These waits need to be done in HOST1X class, while gather submitted
by the application execute in engine class.
Support is added by converting the gather list of job into a command
list that can include both gathers and waits. When the job is
submitted, these commands are pushed as the appropriate opcodes
on the CDMA pushbuffer.
Also supported are waits relative to the start of the job,
which are useful for jobs doing multiple things with an engine
that doesn't natively support pipelining.
While at it, use 32-bit waits on chips that support them.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Add a new property for jobs to enable or disable recovery i.e.
CPU increments of syncpoints to max value on job timeout. This
allows for a more solid model for hanged jobs, where userspace
doesn't need to guess if a syncpoint increment happened because
the job completed, or because job timeout was triggered.
On job timeout, we stop the channel, NOP all future jobs on the
channel using the same syncpoint, mark the syncpoint as locked
and resume the channel from the next job, if any.
The future jobs are NOPed, since because we don't do the CPU
increments, the value of the syncpoint is no longer synchronized,
and any waiters would become confused if a future job incremented
the syncpoint. The syncpoint is marked locked to ensure that any
future jobs cannot increment the syncpoint either, until the
application has recognized the situation and reallocated the
syncpoint.
Signed-off-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
