Philipp Stanner [Tue, 29 Oct 2024 13:38:20 +0000 (14:38 +0100)]
drm/sched: Document purpose of drm_sched_{start,stop}
drm_sched_start()'s and drm_sched_stop()'s names suggest that those
functions might be intended for actively starting and stopping the
scheduler on initialization and teardown.
They are, however, only used on timeout handling (reset recovery). The
docstrings should reflect that to prevent confusion.
The way of implementing a flip option follows the existing
panel-samsung-s6e8aa0.c [1][2][3].
The value to flip the screen is taken from a downstream kernel file of
a similar but older panel [4]. The mipi clock [5] for the new panel
samsung-s6e88a0-ams427ap24 matches 461 MHz and a hardware read-out of the
0xcb values corresponds to revision R01 of that older panel [6]. Although
for samsung-s6e88a0-ams427ap24 that's in non-flipped state while in this
older driver it seems to be the other way around. Further up there is a
hint [7] basically saying for revision R01 to change the first word of the
0xcb command from 0x06 to 0x0e, which is actually setting BIT(3) of that
word. This causes a horizontal flip.
Jakob Hauser [Sun, 27 Oct 2024 23:42:05 +0000 (00:42 +0100)]
drm/panel: samsung-s6e88a0-ams427ap24: Add brightness control
The tables for brightness to candela, aid and elvss were taken from downstream
kernel file "dsi_panel_S6E88A0_AMS427AP24_qhd_octa_video.dtsi" [1][2][3].
The gamma table gets generated in "ss_dsi_smart_dimming_S6E88A0_AMS427AP24.c" [4]
with hard-coded starting values. The function smart_dimming_init() [5] goes
through the v{*}_adjustments, generate_gray_scale and gamma_init procedure.
Instead of calculating it manually, it's easier to compile a custom downstream
kernel with SMART_DIMMING_DEBUG enabled and read out dmesg early at boot.
Selection of the values for aid and elvss are again according to downstream
file "dsi_panel_S6E88A0_AMS427AP24_qhd_octa_video.dtsi" [6][7].
The set of write commands is guided by downstream file "ss_dsi_panel_common.c" [8]
followed by "ss_dsi_panel_S6E88A0_AMS427AP24.c" [9].
The dsi mode flag MIPI_DSI_MODE_VIDEO_NO_HFP prevents screen flickering while
changing the brightness.
This initial part of the panel driver was mostly generated by the
"linux-mdss-dsi-panel-driver-generator" tool [1], reading downstream
Android kernel file "dsi_panel_S6E88A0_AMS427AP24_qhd_octa_video.dtsi" [2].
On top of the generic output of the tool, there were a couple of changes
applied:
- Added mipi_dsi_dcs_set_display_on() to function s6e88a0_ams427ap24_on(),
otherwise the display does not show up.
- In functions s6e88a0_ams427ap24_on() and s6e88a0_ams427ap24_off()
changed DSI commands to multi context and used "accum_err" returns.
- In functions s6e88a0_ams427ap24_on() and s6e88a0_ams427ap24_off() replaced
msleep() by mipi_dsi_msleep().
- The function s6e88a0_ams427ap24_get_modes() was changed to make use of
drm_connector_helper_get_modes_fixed(). This also required to include
drm/drm_probe_helper.h.
- In function s6e88a0_ams427ap24_probe() registring the regulators was changed
to devm_regulator_bulk_get_const(). This required to change supplies in struct
s6e88a0_ams427ap24 to a pointer.
- Removed bool "prepared" from struct s6e88a0_ams427ap24 and according parts in
functions s6e88a0_ams427ap24_prepare() and s6e88a0_ams427ap24_unprepare().
- Removed include <linux/of.h>, it's not needed.
- Added comments to the mipi_dsi_dcs_write_seq_multi() lines in function
s6e88a0_ams427ap24_on().
drm_kms_helper_poll_init needs to be called after zynqmp_dpsub_kms_init.
zynqmp_dpsub_kms_init creates the connector and without it we don't
enable hotplug detection.
Sean Anderson [Fri, 9 Aug 2024 19:36:00 +0000 (15:36 -0400)]
drm: zynqmp_dp: Add debugfs interface for compliance testing
Add a debugfs interface for exercising the various test modes supported
by the DisplayPort controller. This allows performing compliance
testing, or performing signal integrity measurements on a failing link.
At the moment, we do not support sink-driven link quality testing,
although such support would be fairly easy to add.
Additionally, add some debugfs files for ignoring AUX errors and HPD
events, as this can allow testing with equipment that cannot emulate a
DPRX.
Sean Anderson [Fri, 9 Aug 2024 19:35:59 +0000 (15:35 -0400)]
drm: zynqmp_dp: Take dp->lock in zynqmp_dp_hpd_work_func
Add a non-locking version of zynqmp_dp_bridge_detect and use it in
zynqmp_dp_hpd_work_func so we can take the lock explicitly. This will
make it easier to check for hpd_ignore when we add debugfs support.
Sean Anderson [Fri, 9 Aug 2024 19:35:56 +0000 (15:35 -0400)]
drm: zynqmp_dp: Convert to a hard IRQ
Now that all of the sleeping work is done outside of the IRQ, we can
convert it to a hard IRQ. Shared IRQs may be triggered even after
calling disable_irq, so use free_irq instead which removes our callback
altogether.
Sean Anderson [Fri, 9 Aug 2024 19:35:55 +0000 (15:35 -0400)]
drm: zynqmp_dp: Don't retrain the link in our IRQ
Retraining the link can take a while, and might involve waiting for
DPCD reads/writes to complete. In preparation for unthreading the IRQ
handler, move this into its own work function.
Sean Anderson [Fri, 9 Aug 2024 19:35:54 +0000 (15:35 -0400)]
drm: zynqmp_dp: Add locking
Add some locking to prevent the IRQ/workers/bridge API calls from stepping
on each other's toes. This lock protects:
- Non-atomic registers configuring the link. That is, everything but the
IRQ registers (since these are accessed in an atomic fashion), and the DP
AUX registers (since these don't affect the link). We also access AUX
while holding this lock, so it would be very tricky to support.
- Link configuration. This is effectively everything in zynqmp_dp which
isn't read-only after probe time. So from next_bridge onward.
This lock is designed to protect configuration changes so we don't have to
do anything tricky. Configuration should never be in the hot path, so I'm
not worried about performance.
Sean Anderson [Fri, 9 Aug 2024 19:35:53 +0000 (15:35 -0400)]
drm: zynqmp_kms: Unplug DRM device before removal
Prevent userspace accesses to the DRM device from causing
use-after-frees by unplugging the device before we remove it. This
causes any further userspace accesses to result in an error without
further calls into this driver's internals.
Karol Wachowski [Thu, 17 Oct 2024 14:58:17 +0000 (16:58 +0200)]
accel/ivpu: Move secondary preemption buffer allocation to DMA range
Secondary preemption buffer is accessible by NPU's DMA and can be
allocated with addresses above 4 GB. Move secondary preemption buffer
allocation from SHAVE range which is much smaller (2GB) to DMA range.
This allows to allocate more command queues with corresponding
preemption buffers without running out of address range.
Karol Wachowski [Thu, 17 Oct 2024 14:58:16 +0000 (16:58 +0200)]
accel/ivpu: Increase DMA address range
Increase DMA address range to:
* 128 GB on 37xx (due to MMU limitations)
* 256 GB on other generations
Merge User and DMA ranges on 40xx and above as it is possible
to access whole 256 GBs from both FW and DMA.
Increase User range on 37xx from 255MB to 511MB
to allow loading very large models.
Do not set global_alias_pio_base/size on other generations than 37xx
as it's only used on 37xx anyway.
Maciej Falkowski [Thu, 17 Oct 2024 14:58:15 +0000 (16:58 +0200)]
accel/ivpu: Add debug Kconfig option
Add CONFIG_DRM_ACCEL_IVPU_DEBUG option that:
- Adds -DDEBUG that enables printk regardless of the kernel config
- Enables unsafe module params (that are now disabled by default)
Karol Wachowski [Thu, 17 Oct 2024 14:58:14 +0000 (16:58 +0200)]
accel/ivpu: Don't allocate preemption buffers when MIP is disabled
Do not allocate preemption buffers when Mid Inference Preemption (MIP)
is disabled through test mode.
Rename IVPU_TEST_MODE_PREEMPTION_DISABLE to IVPU_TEST_MODE_MIP_DISABLE
to better describe that this test mode only disables MIP - job level
preemption will still occur.
Karol Wachowski [Thu, 17 Oct 2024 14:58:13 +0000 (16:58 +0200)]
accel/ivpu: Make command queue ID allocated on XArray
Use XArray for dynamic command queue ID allocations instead of fixed
ones. This is required by upcoming changes to UAPI that will allow to
manage command queues by user space instead of having predefined number
of queues in a context.
Karol Wachowski [Thu, 17 Oct 2024 14:58:12 +0000 (16:58 +0200)]
accel/ivpu: Use xa_alloc_cyclic() instead of custom function
Remove custom ivpu_id_alloc() wrapper used for ID allocations
and replace it with standard xa_alloc_cyclic() API.
The idea behind ivpu_id_alloc() was to have monotonic IDs, so the driver
is easier to debug because same IDs are not reused all over. The same
can be achieved just by using appropriate Linux API.
Karol Wachowski [Thu, 17 Oct 2024 14:58:10 +0000 (16:58 +0200)]
accel/ivpu: Clear CDTAB entry in case of failure
Don't leave a context descriptor in case CFGI_ALL flush fails.
Mark it as invalid (by clearing valid bit) so nothing is left in
partially-initialized state.
Karol Wachowski [Thu, 17 Oct 2024 14:58:08 +0000 (16:58 +0200)]
accel/ivpu: Defer MMU root page table allocation
Defer root page table allocation and unify context init/fini functions.
Move allocation of the root page table from the file_priv_open function to
perform a lazy allocation approach during ivpu_bo_pin().
By doing so, we avoid the overhead of allocating page tables for simple
operations like GET_PARAM that do not require them.
Additionally, the MMU context descriptor table initialization has been
moved to the ivpu_mmu_context_map_page function.
This change streamlines the process and ensures that the descriptor table
is only initialized when it is actually needed.
Refactor init/fini functions to remove redundant code and make the context
management more straightforward.
Overall, these changes lead to a reduction in the time taken by the file
descriptor open operation, as the costly root page table allocation is now
avoided for operations that do not require it.
Marek Vasut [Sat, 26 Oct 2024 04:10:42 +0000 (06:10 +0200)]
drm/bridge: tc358767: Fix use of unadjusted mode in the driver
The driver configures mostly Pixel PLL from the clock cached in
local copy of the mode. Make sure the driver uses adjusted mode
which contains the updated Pixel PLL settings negotiated in
tc_dpi_atomic_check()/tc_edp_atomic_check().
dispc_enable_fifomerge() last use was removed by 2012's
commit 85099f11bd03 ("Revert "OMAPDSS: APPLY: add fifo merge support
funcs"")
dispc_has_writeback(), dispc_wb_get_framedone_irq(), dispc_wb_go(),
dispc_wb_go_busy() and dispc_wb_setup() were changed from statics
to public symbols and unwired from a structure by 2020's
commit dac62bcafeaa ("drm/omap: remove dispc_ops")
but didn't have any users.
dispc_mgr_get_clock_div() got renamed from dispc_get_clock_div()
and it's last use was removed in 2011 by commit 42c9dee82129 ("OMAP: DSS2: Remove FB_OMAP_BOOTLOADER_INIT support")
Jeffrey Hugo [Mon, 21 Oct 2024 20:03:55 +0000 (14:03 -0600)]
accel/qaic: Add crashdump to Sahara
The Sahara protocol has a crashdump functionality. In the hello
exchange, the device can advertise it has a memory dump available for
the host to collect. Instead of the device making requests of the host,
the host requests data from the device which can be later analyzed.
Implement this functionality and utilize the devcoredump framework for
handing the dump over to userspace.
Similar to how firmware loading in Sahara involves multiple files,
crashdump can consist of multiple files for different parts of the dump.
Structure these into a single buffer that userspace can parse and
extract the original files from.
Philipp Stanner [Wed, 23 Oct 2024 14:15:31 +0000 (16:15 +0200)]
drm/sched: warn about drm_sched_job_init()'s partial init
drm_sched_job_init()'s name suggests that after the function succeeded,
parameter "job" will be fully initialized. This is not the case; some
members are only later set, notably drm_sched_job.sched by
drm_sched_job_arm().
Document that drm_sched_job_init() does not set all struct members.
Dmitry Baryshkov [Fri, 18 Oct 2024 21:49:16 +0000 (00:49 +0300)]
drm/msm/dp: migrate the ycbcr_420_allowed to drm_bridge
Instead of forcing the ycbcr_420_allowed flag to be set on the created
drm_connector, set it on the drm_bridge instance and allow
drm_bridge_connecgtor to propagate it to the drm_connector.
Dmitry Baryshkov [Fri, 18 Oct 2024 21:49:15 +0000 (00:49 +0300)]
drm/bridge: aux: allow interlaced and YCbCr 420 output
As both aux bridges are merely passthrough bridges, mark them as
supporting interlaced and YCbCr 420 data. Other bridges in the chain
still might limit interlaced and YCbCr 420 data support on the
corresponding connector.
Dmitry Baryshkov [Fri, 18 Oct 2024 21:49:14 +0000 (00:49 +0300)]
drm/bridge: display-connector: allow YCbCr 420 for HDMI and DP
Allow YCbCr 420 output for HDMI and DisplayPort connectors. Other
bridges in the chain still might limit YCbCr 420 support on the
corresponding connector.
Dmitry Baryshkov [Fri, 18 Oct 2024 21:49:13 +0000 (00:49 +0300)]
drm/atomic: add interlaced and ycbcr_420 flags to connector's state dump
Although the interlace_allowed and ycbcr_420_allowed flags are a part of
the struct drm_connector rather than struct drm_connector_state, still
include them into state dump in order to ease debugging of the setup
issues.
Follow the interlace_allowed example and calculate drm_connector's
ycbcr_420_allowed flag as AND of all drm_bridge's ycbcr_420_allowed
flags in a chain. This is one of the gaps between several
bridge-specific connector implementations and drm_bridge_connector.
Philipp Stanner [Mon, 21 Oct 2024 10:50:28 +0000 (12:50 +0200)]
drm/sched: memset() 'job' in drm_sched_job_init()
drm_sched_job_init() has no control over how users allocate struct
drm_sched_job. Unfortunately, the function can also not set some struct
members such as job->sched.
This could theoretically lead to UB by users dereferencing the struct's
pointer members too early.
It is easier to debug such issues if these pointers are initialized to
NULL, so dereferencing them causes a NULL pointer exception.
Accordingly, drm_sched_entity_init() does precisely that and initializes
its struct with memset().
Initialize parameter "job" to 0 in drm_sched_job_init().
The drm_fb_helper_deferred_io() uses struct fb_deferred_io_pageref,
which isn't available without CONFIG_FB_DEFERRED_IO. Put the function
under corresponding #ifdef to fix build failure if deferred I/O isn't
enabled.
Wadim Egorov [Thu, 17 Oct 2024 08:55:56 +0000 (10:55 +0200)]
drm/bridge: sii902x: Set input bus format based on bus-width
Introduce a bus-width property to define the number of parallel RGB
input pins connected to the transmitter. The input bus formats are updated
accordingly. If the property is not specified, default to 24-bit bus-width.
The SI9022 HDMI transmitter can be configured with a bus-width of 16,
18, or 24 bits. Introduce a bus-width property to the input endpoint,
specifying the number of parallel RGB input bits connected to the
transmitter.
Liu Ying [Thu, 10 Oct 2024 09:26:43 +0000 (17:26 +0800)]
drm/bridge: ite-it66121: Drop hdmi_avi_infoframe_init() function call
drm_hdmi_avi_infoframe_from_display_mode() called from
it66121_bridge_mode_set() already calls hdmi_avi_infoframe_init() to
initialize an HDMI AVI infoframe. So, drop the redundant
hdmi_avi_infoframe_init() function call from it66121_bridge_mode_set().
"data-lines" property is way too similar to "data-lanes". It is also
duplicating "bus-width" from video-interfaces.yaml schema. "data-lines"
was deprecated in the bindings and "bus-width" is preferred, so parse it
instead while keeping things backwards compatible.
dt-bindings: display: bridge: tc358768: switch to bus-width
"data-lines" property is way too similar to "data-lanes". It is also
duplicating "bus-width" from video-interfaces.yaml schema. Deprecate
"data-lines" and use the common property.
The kernel-doc comments in these 3 files was removed so remove the
references to these files to prevent kernel-doc warnings.
drivers/gpu/drm/drm_fbdev_dma.c:1: warning: no structured comments found
drivers/gpu/drm/drm_fbdev_shmem.c:1: warning: no structured comments found
drivers/gpu/drm/drm_fbdev_ttm.c:1: warning: no structured comments found
Ryan Walklin [Sun, 20 Oct 2024 08:37:41 +0000 (21:37 +1300)]
drm: panel: nv3052c: correct spi_device_id for RG35XX panel
The Anbernic RG35XX devices use an SPI LCD panel from an unknown OEM,
with an NV3052C driver chip.
As discussed previously, the integrating vendor and device name are
preferred instead of the OEM serial. A previous patch corrected the
device tree binding and of_device_id in the NV3052C driver, however the
spi_device_id also needs correction.
Chen Ni [Wed, 18 Sep 2024 07:48:41 +0000 (15:48 +0800)]
drm/fsl-dcu: Remove redundant dev_err()
There is no need to call the dev_err() function directly to print a
custom message when handling an error from platform_get_irq() function
as it is going to display an appropriate error message in case of a
failure.
The PIXCLK needs to be enabled in SCFG before accessing certain DCU
registers, or the access will hang. For simplicity, the PIXCLK is enabled
unconditionally, resulting in increased power consumption.
Simon Horman [Fri, 11 Oct 2024 09:57:12 +0000 (10:57 +0100)]
accel/qaic: Pass string literal as format argument of alloc_workqueue()
Recently I noticed that both gcc-14 and clang-18 report that passing
a non-string literal as the format argument of alloc_workqueue()
is potentially insecure.
E.g. clang-18 says:
.../qaic_drv.c:61:23: warning: format string is not a string literal (potentially insecure) [-Wformat-security]
61 | wq = alloc_workqueue(fmt, WQ_UNBOUND, 0);
| ^~~
.../qaic_drv.c:61:23: note: treat the string as an argument to avoid this
61 | wq = alloc_workqueue(fmt, WQ_UNBOUND, 0);
| ^
| "%s",
It is always the case where the contents of fmt is safe to pass as the
format argument. That is, in my understanding, it never contains any
format escape sequences.
But, it seems better to be safe than sorry. And, as a bonus, compiler
output becomes less verbose by addressing this issue as suggested by
clang-18.
Also, change the name of the parameter of qaicm_wq_init from
fmt to name to better reflect it's purpose.
Jean Delvare [Tue, 15 Oct 2024 11:46:06 +0000 (13:46 +0200)]
drm/display: Drop obsolete dependency on COMPILE_TEST
Since commit 0166dc11be91 ("of: make CONFIG_OF user selectable"), it
is possible to test-build any driver which depends on OF on any
architecture by explicitly selecting OF. Therefore depending on
COMPILE_TEST as an alternative is no longer needed.
To avoid reintroducing the randconfig bug originally fixed by commit 876271118aa4 ("drm/display: Fix build error without CONFIG_OF"),
DRM_MSM which selects DRM_DISPLAY_DP_HELPER must explicitly depend
on OF. This is consistent with what all other DRM drivers are doing.
Add drm_client_lib.ko to contain DRM's built-in client. Move the
existing client for fbdev emulation into the new module. Protect the
new module behind CONFIG_DRM_CLIENT_LIB.
The Kconfig rules separate the DRM drivers from the DRM clients. A
driver can opt into the default clients, but the user configures
each client individually. To do so, DRM drivers still select
DRM_CLIENT_SELECTION. The option is now a tristate that further
selects all dependencies of the enabled DRM clients. There's
a menu option for each client. Enabling at least one client also
selects DRM_CLIENT_SETUP, so that drivers call drm_client_setup().
New DRM clients should depend on DRM_CLIENT_SELECTION.
There are existing kernel options in drm_fb_helper.o, so leave this
file in the KMS-helper module for now.
v5:
- leave fbdev helpers in drm_kms_helper.ko for now
v3:
- fix commit changelog
v2:
- keep client code in core
- protect lib with DRM_CLIENT_LIB
- remove duplicate line from Makefile (Jocelyn)
Only build client code if DRM_CLIENT has been selected. Automatially
do so if one of the default clients has been enabled. If client support
has been disabled, the helpers for client-related events are empty and
the regular client functions are not present.
Amdgpu has an internal DRM client, so it has to select DRM_CLIENT by
itself unconditionally.
v3:
- provide empty drm_client_debugfs_init() if DRM_CLIENT=n (kernel
test robot)
drm/radeon: Suspend and resume clients with client helpers
Replace calls to radeon_fbdev_set_suspend() with calls to the client
functions drm_client_dev_suspend() and drm_client_dev_resume(). Any
registered in-kernel client will now receive suspend and resume events.
v4:
- refer to radeon_fbdev_set_suspend() in commit description (Jonathan)
drm/nouveau: Suspend and resume clients with client helpers
Replace calls to drm_fb_helper_set_suspend_unlocked() with calls
to the client functions drm_client_dev_suspend() and
drm_client_dev_resume(). Any registered in-kernel client will now
receive suspend and resume events.
drm/amdgpu: Suspend and resume internal clients with client helpers
Replace calls to drm_fb_helper_set_suspend_unlocked() with calls
to the client functions drm_client_dev_suspend() and
drm_client_dev_resume(). Any registered in-kernel client will now
receive suspend and resume events.
drm/client: Move suspend/resume into DRM client callbacks
Suspend and resume is still tied to fbdev emulation. Modeset helpers
and several drivers call drm_fb_helper_set_suspend_unlocked() to inform
the fbdev client about suspend/resume events.
To make it work with arbitrary clients, add per-client callback
functions for suspend and resume. Implement them for fbdev emulation
with the existing drm_fb_helper_set_suspend_unlocked(). Then update
DRM's modeset helpers to call the new interface.
Clients that are not fbdev can now implement suspend/resume to their
requirements.
The callback parameter holds_console_lock is a workaround for i915,
radeon and xe, which possibly call the interface while having the
console lock acquired. Even though the commit doesn't modify these
drivers, it already adds the flag to avoid churn later on. New code
should not hold the console lock.
v4:
- clarify holds_console_lock in commit description (Jonathan)
drm/client: Move client event handlers to drm_client_event.c
A number of DRM-client functions serve as entry points from device
operations to client code. Moving them info a separate file will later
allow for a more fine-grained kernel configuration. For most of the
users it is sufficient to include <drm/drm_client_event.h> instead of
the full driver-side interface in <drm/drm_client.h>
drm/fbdev: Select fbdev I/O helpers from modules that require them
Fbdev emulation for SHMEM and TTM requires helpers from the fbdev
subsystem. Select them from the modules that use them instead of the
core DRM module.
Commit 808a40b69468 ("drm/fbdev-dma: Implement damage handling and
deferred I/O") added deferred I/O for fbdev-dma. Also select the
Kconfig symbol FB_DEFERRED_IO (via FB_DMAMEM_HELPERS_DEFERRED). Fixes
build errors about missing fbdefio, such as
drivers/gpu/drm/drm_fbdev_dma.c:218:26: error: 'struct drm_fb_helper' has no member named 'fbdefio'
218 | fb_helper->fbdefio.delay = HZ / 20;
| ^~
drivers/gpu/drm/drm_fbdev_dma.c:219:26: error: 'struct drm_fb_helper' has no member named 'fbdefio'
219 | fb_helper->fbdefio.deferred_io = drm_fb_helper_deferred_io;
| ^~
drivers/gpu/drm/drm_fbdev_dma.c:221:21: error: 'struct fb_info' has no member named 'fbdefio'
221 | info->fbdefio = &fb_helper->fbdefio;
| ^~
drivers/gpu/drm/drm_fbdev_dma.c:221:43: error: 'struct drm_fb_helper' has no member named 'fbdefio'
221 | info->fbdefio = &fb_helper->fbdefio;
| ^~
drm/rockchip: Add basic RK3588 HDMI output support
The RK3588 SoC family integrates the newer Synopsys DesignWare HDMI 2.1
Quad-Pixel (QP) TX controller IP and a HDMI/eDP TX Combo PHY based on a
Samsung IP block.
Add just the basic support for now, i.e. RGB output up to 4K@60Hz,
without audio, CEC or any of the HDMI 2.1 specific features.
Tvrtko Ursulin [Wed, 16 Oct 2024 12:20:13 +0000 (13:20 +0100)]
drm/sched: Further optimise drm_sched_entity_push_job
Having removed one re-lock cycle on the entity->lock in a patch titled
"drm/sched: Optimise drm_sched_entity_push_job", with only a tiny bit
larger refactoring we can do the same optimisation on the rq->lock.
(Currently both drm_sched_rq_add_entity() and
drm_sched_rq_update_fifo_locked() take and release the same lock.)
To achieve this we make drm_sched_rq_update_fifo_locked() and
drm_sched_rq_add_entity() expect the rq->lock to be held.
We also align drm_sched_rq_update_fifo_locked(),
drm_sched_rq_add_entity() and
drm_sched_rq_remove_fifo_locked() function signatures, by adding rq as a
parameter to the latter.
v2:
* Fix after rebase of the series.
* Avoid naming inconsistency between drm_sched_rq_add/remove. (Christian)
Tvrtko Ursulin [Wed, 16 Oct 2024 12:20:12 +0000 (13:20 +0100)]
drm/sched: Re-group and rename the entity run-queue lock
When writing to a drm_sched_entity's run-queue, writers are protected
through the lock drm_sched_entity.rq_lock. This naming, however,
frequently collides with the separate internal lock of struct
drm_sched_rq, resulting in uses like this:
Tvrtko Ursulin [Wed, 16 Oct 2024 12:20:10 +0000 (13:20 +0100)]
drm/sched: Stop setting current entity in FIFO mode
It does not seem there is a need to set the current entity in FIFO mode
since ot only serves as being a "cursor" in round-robin mode. Even if
scheduling mode is changed at runtime the change in behaviour is simply
to restart from the first entity, instead of continuing in RR mode from
where FIFO left it, and that sounds completely fine.
Tvrtko Ursulin [Wed, 16 Oct 2024 12:20:09 +0000 (13:20 +0100)]
drm/sched: Optimise drm_sched_entity_push_job
In FIFO mode (which is the default), both drm_sched_entity_push_job() and
drm_sched_rq_update_fifo(), where the latter calls the former, are
currently taking and releasing the same entity->rq_lock.
We can avoid that design inelegance, and also have a miniscule
efficiency improvement on the submit from idle path, by introducing a new
drm_sched_rq_update_fifo_locked() helper and pulling up the lock taking to
its callers.
The new driver needs the dsc helper code to be available:
x86_64-linux-ld: vmlinux.o: in function `s6e3ha8_amb577px01_wqhd_prepare':
panel-samsung-s6e3ha8.c:(.text+0x16b1e65): undefined reference to `drm_dsc_pps_payload_pack'
Select it from Kconfig as we do for other similar drivers.
Uros Bizjak [Mon, 7 Oct 2024 08:37:52 +0000 (10:37 +0200)]
dma-buf: Use atomic64_inc_return() in dma_buf_getfile()
Use atomic64_inc_return(&ref) instead of atomic64_add_return(1, &ref)
to use optimized implementation and ease register pressure around
the primitive for targets that implement optimized variant.
projects / linux.git / log