It is necessary to wait for the full frame to finish streaming
through the DMA engine before we can safely disable it by removing
the DISP_PARA_DISP_ON bit. Disabling it in-flight can leave the
hardware confused and unable to resume streaming for the next frame.
This causes the FIFO underrun and empty status bits to be set and
a single solid color to be shown on the display, coming from one of
the pixels of the previous frame. The issue occurs sporadically when
a new mode is set, which triggers the crtc disable and enable paths.
Setting the shadow load bit and waiting for it to be cleared by the
DMA engine allows waiting for completion.
The NXP BSP driver addresses this issue with a hardcoded 25 ms sleep.
Fixes: 9db35bb349 ("drm: lcdif: Add support for i.MX8MP LCDIF variant")
Signed-off-by: Paul Kocialkowski <paulk@sys-base.io>
Co-developed-by: Lucas Stach <l.stach@pengutronix.de>
Reviewed-by: Frieder Schrempf <frieder.schrempf@kontron.de>
Tested-by: Frieder Schrempf <frieder.schrempf@kontron.de>
Acked-by: Liu Ying <victor.liu@nxp.com>
Link: https://patch.msgid.link/20260402183351.3281123-3-paulk@sys-base.io
Signed-off-by: Lucas Stach <l.stach@pengutronix.de>
Original commit b22fd0b963 ("drm/tegra: dsi: Clear enable register if
powered by bootloader") was added to address the issue of DSI being in an
unknown state after the bootloader, ensuring correct panel configuration.
This worked fairly well under the assumption that the bootloader had set
up DSI; however, in cases where it did not, the device would hang because
a DSI read was called before the DSI hardware was ready.
Removing this workaround results in the issue described in the original
fix: the panel initialization sequence fails and the panel gets stuck in
an undefined state. This is especially noticeable with command mode panels
In order to properly address this issue, the original workaround is
restored and placed after the DSI hardware is prepared for R/W operations.
This fixes behavior for both cases: where DSI is set by the bootloader and
where DSI is untouched.
I have tested this change on Tegra20 (Motorola Atrix 4G),
Tegra114 (NVIDIA Tegra Note 7 and ASUS Transformer Pad TF701T), and
Tegra124 (Xiaomi Mi Pad) with U-Boot, using both bootloader-initialized
DSI and untouched DSI.
Fixes: b22fd0b963 ("drm/tegra: dsi: Clear enable register if powered by bootloader")
Fixes: 660b299bed ("Revert "drm/tegra: dsi: Clear enable register if powered by bootloader"")
Signed-off-by: Svyatoslav Ryhel <clamor95@gmail.com>
Reviewed-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Link: https://patch.msgid.link/20260511074538.24563-3-clamor95@gmail.com
Extend the scoreboard approach from the previous commit to used blocks,
so drm_buddy_print() can report per-order allocation pressure in O(1).
Unlike free blocks, an allocated block can leave the allocated state
through mark_free() (normal free and gpu_buddy_block_trim()) or be
consumed directly by gpu_block_free() during coalescing. Both sites are
guarded by gpu_buddy_block_is_allocated() and paired with the increment
in mark_allocated().
v3:
- Assert scoreboard is empty at fini(), as sanity check (Matthew Auld)
v2:
- Update after fix for use-after-free in split_block() call sites
- Change goto label to out_free_used_scoreboard for clarity
- Make drm_buddy_print() and gpu_buddy_print() symmetric for used and
free
Assisted-by: GitHub Copilot:claude-sonnet-4.6
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://lore.kernel.org/r/20260522092600.32818-6-francois.dugast@intel.com
Signed-off-by: Francois Dugast <francois.dugast@intel.com>
Reporting per-order free block counts in drm_buddy_print() currently
requires walking all rbtrees, which is O(n) over the total number of
free blocks and holds the allocator lock for the duration. This becomes
expensive on large VRAM heaps with many small free fragments.
Maintain a free_scoreboard[] array indexed by order instead, so that
the count for any order is always available in O(1). The scoreboard is
kept accurate by hooking into the four places where a block's free state
changes: mark_free(), mark_allocated(), mark_split(), and the sites in
__gpu_buddy_free(), __force_merge(), and the four err_undo paths that
call rbtree_remove() directly on free blocks without going through
mark_*().
The print functions are simplified as a result: the rbtree traversal
is replaced by a direct array lookup.
v3: Update after introducing __gpu_buddy_undo_splits() helper
v2: Update after fix for use-after-free in split_block() call sites
Assisted-by: GitHub Copilot:claude-sonnet-4.6
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://lore.kernel.org/r/20260522092600.32818-5-francois.dugast@intel.com
Signed-off-by: Francois Dugast <francois.dugast@intel.com>
The err_undo label in alloc_from_freetree() is only reachable via a
goto from inside the `while (tmp != order)` loop, which means tmp is
guaranteed to differ from order at that point. The surrounding
`if (tmp != order)` guard was therefore always true and can be dropped
without any behavioral change.
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://lore.kernel.org/r/20260522092600.32818-3-francois.dugast@intel.com
Signed-off-by: Francois Dugast <francois.dugast@intel.com>
When split_block() fails it returns before calling mark_split(), leaving
the block in the FREE state and still linked in the rbtree. The four
err_undo paths then call __gpu_buddy_free() without first removing the
block from the tree, which leads to two distinct bugs:
- If the buddy is also free, __gpu_buddy_free() merges the two siblings
by calling gpu_block_free(mm, block) while block->rb is still linked
in the tree. Any subsequent rbtree traversal will follow the now-
dangling pointer, causing a use-after-free.
- In alloc_from_freetree(), where there is no buddy guard,
__gpu_buddy_free() always reaches mark_free() -> rbtree_insert() with
block still in the tree, corrupting the rbtree.
The same pattern is already used correctly in __force_merge(): call
rbtree_remove() to unlink the block before handing it to
__gpu_buddy_free(). Apply the same fix to all four err_undo sites.
Reported-by: Sashiko <sashiko-bot@kernel.org>
Assisted-by: GitHub Copilot:claude-sonnet-4.6
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://lore.kernel.org/r/20260522092600.32818-2-francois.dugast@intel.com
Signed-off-by: Francois Dugast <francois.dugast@intel.com>
Function and structure names in the Hyper-V DRM driver currently
use "hyperv_" as the prefix. This conflicts with usage in core Hyper-V
and VMBus code, and incorrectly implies that functions and structures
in this driver apply generically to Hyper-V. A specific conflict arises
for "hyperv_init", which is an initcall for generic Hyper-V
initialization on arm64. The conflict prevents the use of
initcall_blacklist on the kernel boot line to skip loading this driver.
Fix this by substituting "hv_drm_" as the prefix for all functions and
structures in this driver. In most places, this is replacing "hyperv_"
with "hv_drm_". In a few places, the substitution results in
"hv_drm_drm_", which has been collapsed to just "hv_drm_". In two
cases, the existing prefix is a bare "hv" (including in the to_hv()
macro), which has been replaced with "hv_drm" for consistency.
The changes are all mechanical text substitution in symbol names.
There are no other code or functional changes.
Signed-off-by: Michael Kelley <mhklinux@outlook.com>
Reviewed-by: Dexuan Cui <decui@microsoft.com>
Signed-off-by: Hamza Mahfooz <someguy@effective-light.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20260529014826.41256-1-mhklkml@zohomail.com
Register a PM notifier in virtio-gpu to handle suspend/hibernate
events. On PM_POST_HIBERNATION, restore all GPU objects so that the
driver can properly recover after resume.
Suggested-by: Dmitry Osipenko <dmitry.osipenko@collabora.com>
Cc: Vivek Kasireddy <vivek.kasireddy@intel.com>
Signed-off-by: Dongwon Kim <dongwon.kim@intel.com>
Signed-off-by: Dmitry Osipenko <dmitry.osipenko@collabora.com>
Link: https://patch.msgid.link/20260526192814.179673-4-dongwon.kim@intel.com
[dmitry.osipenko@collabora.com: rename vgdev->hibernation -> hibernated]
[dmitry.osipenko@collabora.com: restore objs on suspend error]
[dmitry.osipenko@collabora.com: no_cb int -> bool]
[dmitry.osipenko@collabora.com: don't include pm_runtime.h]
[dmitry.osipenko@collabora.com: proeprly handle failing register_pm_notifier]
When the host KVM/QEMU resumes from hibernation, it loses all graphics
resources previously submitted by the guest OS, as the QEMU process is
terminated during the suspend-resume cycle. This leads to invalid resource
errors when the guest OS attempts to interact with the host using those
resources after resumption.
To resolve this, the virtio-gpu driver now tracks all active virtio_gpu_objects
and provides a mechanism to restore them by re-submitting the objects to QEMU
when needed (e.g., during resume from hibernation).
Cc: Dmitry Osipenko <dmitry.osipenko@collabora.com>
Cc: Vivek Kasireddy <vivek.kasireddy@intel.com>
Cc: Nirmoy Das <nirmoyd@nvidia.com>
Signed-off-by: Dongwon Kim <dongwon.kim@intel.com>
Signed-off-by: Dmitry Osipenko <dmitry.osipenko@collabora.com>
Link: https://patch.msgid.link/20260526192814.179673-3-dongwon.kim@intel.com
[dmitry.osipenko@collabora.com: remove extra blank line]
[dmitry.osipenko@collabora.com: move err msg to virtio_gpu_object_restore_all]
drm_mode_config_reset() can be used to create the initial state, but
also to return to the initial state, when doing a suspend/resume cycle
for example.
It also affects both the software and the hardware, and drivers can
choose to reset the hardware as well. Most will just create an empty
state and the synchronisation between hardware and software states will
effectively be done when the first commit is done.
That dual role can be harmful, since some objects do need to be
initialized but also need to be preserved across a suspend/resume cycle.
drm_private_obj are such objects for example.
Thus, create another helper for drivers to call to initialize their
state when the driver is loaded, so we can make
drm_mode_config_reset() only about handling suspend/resume and similar.
Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-16-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
Commit 47b5ac7daa ("drm/atomic: Add new atomic_create_state callback
to drm_private_obj") introduced a new pattern for allocating drm object
states.
Instead of relying on the reset() callback, it created a new
atomic_create_state hook. This is helpful because reset is a bit
overloaded: it's used to create the initial software state, reset it,
but also reset the hardware.
It can also be used either at probe time, to create the initial state
and possibly reset the hardware to an expected default, but also during
suspend/resume.
Both these cases come with different expectations too: during the
initialization, we want to initialize all states, but during
suspend/resume, drm_private_states for example are expected to be kept
around.
reset() also isn't fallible, which makes it harder to handle
initialization errors properly. This is only really relevant for some
drivers though, since all the helpers for reset only create a new
state, and don't touch the hardware at all.
It was thus decided to create a new hook that would allocate and
initialize a pristine state without any side effect:
atomic_create_state to untangle a bit some of it, and to separate the
initialization with the actual reset one might need during a
suspend/resume.
Continue the transition to the new pattern with connectors.
Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-15-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
__drm_atomic_helper_connector_hdmi_reset() is typically used to
initialize a newly allocated drm_connector_state when the connector is
using the HDMI helpers, and is being called by the
drm_connector_funcs.reset implementation.
Since we want to consolidate DRM objects state allocation around the
atomic_create_state callback that will only allocate and initialize a
new drm_connector_state instance, we will need to call
__drm_atomic_helper_connector_hdmi_reset() from both the reset and
atomic_create hooks.
To avoid any confusion, we can thus rename
__drm_atomic_helper_connector_hdmi_reset() to
__drm_atomic_helper_connector_hdmi_state_init().
Suggested-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-14-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
__drm_atomic_helper_connector_state_reset() is used to initialize a
newly allocated drm_connector_state, and is being typically called by
the drm_connector_funcs.reset implementation.
Since we want to consolidate DRM objects state allocation around the
atomic_create_state callback that will only allocate and initialize a
new drm_connector_state instance, we will need to call
__drm_atomic_helper_connector_state_reset() from both the reset and
atomic_create hooks.
To avoid any confusion, we can thus rename
__drm_atomic_helper_connector_state_reset() to
__drm_atomic_helper_connector_state_init().
Suggested-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-13-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
Commit 47b5ac7daa ("drm/atomic: Add new atomic_create_state callback
to drm_private_obj") introduced a new pattern for allocating drm object
states.
Instead of relying on the reset() callback, it created a new
atomic_create_state hook. This is helpful because reset is a bit
overloaded: it's used to create the initial software state, reset it,
but also reset the hardware.
It can also be used either at probe time, to create the initial state
and possibly reset the hardware to an expected default, but also during
suspend/resume.
Both these cases come with different expectations too: during the
initialization, we want to initialize all states, but during
suspend/resume, drm_private_states for example are expected to be kept
around.
reset() also isn't fallible, which makes it harder to handle
initialization errors properly. This is only really relevant for some
drivers though, since all the helpers for reset only create a new
state, and don't touch the hardware at all.
It was thus decided to create a new hook that would allocate and
initialize a pristine state without any side effect:
atomic_create_state to untangle a bit some of it, and to separate the
initialization with the actual reset one might need during a
suspend/resume.
Continue the transition to the new pattern with CRTCs.
Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@oss.qualcomm.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-12-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
__drm_atomic_helper_crtc_state_reset() is used to initialize a newly
allocated drm_crtc_state, and is being typically called by the
drm_crtc_funcs.reset implementation.
Since we want to consolidate DRM objects state allocation around the
atomic_create_state callback that will only allocate and initialize a
new drm_crtc_state instance, we will need to call
__drm_atomic_helper_crtc_state_reset() from both the reset and
atomic_create hooks.
To avoid any confusion, we can thus rename
__drm_atomic_helper_crtc_state_reset() to
__drm_atomic_helper_crtc_state_init().
Suggested-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-11-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
Commit 47b5ac7daa ("drm/atomic: Add new atomic_create_state callback
to drm_private_obj") introduced a new pattern for allocating drm object
states.
Instead of relying on the reset() callback, it created a new
atomic_create_state hook. This is helpful because reset is a bit
overloaded: it's used to create the initial software state, reset it,
but also reset the hardware.
It can also be used either at probe time, to create the initial state
and possibly reset the hardware to an expected default, but also during
suspend/resume.
Both these cases come with different expectations too: during the
initialization, we want to initialize all states, but during
suspend/resume, drm_private_states for example are expected to be kept
around.
reset() also isn't fallible, which makes it harder to handle
initialization errors properly. This is only really relevant for some
drivers though, since all the helpers for reset only create a new
state, and don't touch the hardware at all.
It was thus decided to create a new hook that would allocate and
initialize a pristine state without any side effect:
atomic_create_state to untangle a bit some of it, and to separate the
initialization with the actual reset one might need during a
suspend/resume.
Continue the transition to the new pattern with planes.
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-10-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
__drm_atomic_helper_plane_state_reset() is used to initialize a newly
allocated drm_plane_state, and is being typically called by the
drm_plane_funcs.reset implementation.
Since we want to consolidate DRM objects state allocation around the
atomic_create_state callback that will only allocate and initialize a
new drm_plane_state instance, we will need to call
__drm_atomic_helper_plane_state_reset() from both the reset and
atomic_create hooks.
To avoid any confusion, we can thus rename
__drm_atomic_helper_plane_state_reset() to
__drm_atomic_helper_plane_state_init().
Suggested-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-9-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
Commit 47b5ac7daa ("drm/atomic: Add new atomic_create_state callback
to drm_private_obj") introduced a new pattern for allocating drm object
states.
Instead of relying on the reset() callback, it created a new
atomic_create_state hook. This is helpful because reset is a bit
overloaded: it's used to create the initial software state, reset it,
but also reset the hardware.
It can also be used either at probe time, to create the initial state
and possibly reset the hardware to an expected default, but also during
suspend/resume.
Both these cases come with different expectations too: during the
initialization, we want to initialize all states, but during
suspend/resume, drm_private_states for example are expected to be kept
around.
reset() also isn't fallible, which makes it harder to handle
initialization errors properly. This is only really relevant for some
drivers though, since all the helpers for reset only create a new
state, and don't touch the hardware at all.
It was thus decided to create a new hook that would allocate and
initialize a pristine state without any side effect:
atomic_create_state to untangle a bit some of it, and to separate the
initialization with the actual reset one might need during a
suspend/resume.
Continue the transition to the new pattern with drm_colorop.
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-7-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
__drm_colorop_state_reset() is used to initialize a newly allocated
drm_colorop_state, and is being typically called by drm_colorop_reset().
Since we want to consolidate DRM objects state allocation around the
atomic_create_state callback that will only allocate and initialize a
new drm_colorop_state instance, we will need to call
__drm_colorop_state_reset() from both the reset and atomic_create paths.
To avoid any confusion, we can thus rename __drm_colorop_state_reset()
to __drm_colorop_state_init().
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-6-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
The initial intent of the atomic_create_state hook was to simply
allocate a proper drm_private_state and return it, without any side
effect.
However, __drm_atomic_helper_private_obj_create_state(), which most
atomic_create_state implementations call, introduces a side effect by
setting drm_private_obj.state to the newly allocated state.
This assignment defeats the purpose, but is also redundant since
drm_atomic_private_obj_init(), the only call site for the
atomic_create_state hook, will also set this pointer to the newly
allocated state.
Drop the assignment in __drm_atomic_helper_private_obj_create_state().
Fixes: e7be39ed17 ("drm/atomic-helper: Add private_obj atomic_create_state helper")
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patch.msgid.link/20260526-drm-mode-config-init-v6-3-852346394200@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
Currently the code enqueue work items using {queue|mod}_delayed_work(),
using system_long_wq. This workqueue should be used when long works are
expected, but it is a per-cpu workqueue.
This is important because queue_delayed_work() queue the work using:
queue_delayed_work_on(WORK_CPU_UNBOUND, ...);
Note that WORK_CPU_UNBOUND = NR_CPUS.
This would end up calling __queue_delayed_work() that does:
if (housekeeping_enabled(HK_TYPE_TIMER)) {
// [....]
} else {
if (likely(cpu == WORK_CPU_UNBOUND))
add_timer_global(timer);
else
add_timer_on(timer, cpu);
}
So when cpu == WORK_CPU_UNBOUND the timer is global and is
not using a specific CPU. Later, when __queue_work() is called:
if (req_cpu == WORK_CPU_UNBOUND) {
if (wq->flags & WQ_UNBOUND)
cpu = wq_select_unbound_cpu(raw_smp_processor_id());
else
cpu = raw_smp_processor_id();
}
Because the wq is not unbound, it takes the CPU where the timer
fired and enqueue the work on that CPU.
The consequence of all of this is that the work can run anywhere,
depending on where the timer fired.
Recently, a new unbound workqueue specific for long running work has
been added in commit c116737e97 ("workqueue: Add system_dfl_long_wq
for long unbound works").
So change system_long_wq with system_dfl_long_wq so that the work may
benefit from scheduler task placement.
Signed-off-by: Marco Crivellari <marco.crivellari@suse.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patch.msgid.link/20260430091532.115846-1-marco.crivellari@suse.com
Remove the header #include "drm_internal.h" from drm_dumb_buffers.c,
which is included but not used.
Header was introduced in commit 47f10854ca ("drm: Don't export the
drm_gem_dumb_destroy() function") when moving functions, but was not
removed in commit 96a7b60f6d ("drm: remove dumb_destroy callback")
when the drm_gem_dumb_destroy function was removed.
Compiles successfully with DRM enabled, pass kunit tests and
IGT-tests in a vng virtual machine.
Fixes: 96a7b60f6d ("drm: remove dumb_destroy callback")
Signed-off-by: Yicong Hui <yiconghui@gmail.com>
Reviewed-by: Thomas Zimmermann <tzimmermann@suse.de>
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Link: https://patch.msgid.link/20260409154826.8955-1-yiconghui@gmail.com
Nouveau supports Turing and Ampere GPUs with or without GSP-RM.
Support without GSP-RM is mostly academic, since GSP-RM is
needed to run the GPU at full clocks. It is also the default
mode for these GPUs.
GA100 is a special case, however. The current code has some support
for running GA100 without GSP-RM, but several features are missing.
More importantly, some required firmware images like ucode_ahesasc.bin
are not available and would need to be provided by Nvidia.
To prevent Nouveau from even trying to boot on GA100 without GSP-RM,
remove the non-GSP fallback option in the ga100_gsps[] array.
Signed-off-by: Timur Tabi <ttabi@nvidia.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Link: https://patch.msgid.link/20260430223838.2530778-10-ttabi@nvidia.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
The GPU's ROM may begin with an Init-from-ROM (IFR) header that precedes
the PCI Expansion ROM images (VBIOS). When present, the PROM shadow
method must parse this header to determine the offset where the PCI ROM
images actually begin, and adjust all subsequent reads accordingly.
On most GPUs this is not needed because either the PRAMIN shadow method
(which reads from VRAM via the display engine) succeeds first, or the IFR
microcode has already applied the ROM offset so that PROM reads
transparently skip the header. However, on GA100 neither of these
applies: GA100 has no display engine (so PRAMIN is unavailable), and the
IFR offset is not applied to PROM reads on this GPU.
Signed-off-by: Timur Tabi <ttabi@nvidia.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Link: https://patch.msgid.link/20260430223838.2530778-9-ttabi@nvidia.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
The NV_FUSE_STATUS_OPT_DISPLAY register is used to determine whether
the GPU has display hardware. The current code that normally reads
this register is instead hard-coded to check for GA100 vs later GPUs.
Since this function is called only on pre-Hopper GPUs, and this
if-statement applies only to GA100 and later, the check works
because GA100 is the only non-display Ampere and Ada GPU.
However, there actually is a register that can be read, so we should
use it.
Fixes: a34632482f ("drm/nouveau/bios/ga10[024]: initial support")
Signed-off-by: Timur Tabi <ttabi@nvidia.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Link: https://patch.msgid.link/20260430223838.2530778-8-ttabi@nvidia.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
When booting with GSP-RM, the FRTS data region normally needs to be
allocated. However, on GA100, this region is not used and so its
size needs to be set to zero.
The truth is that GA100 is just special, and the simplest way to
determine the proper FRTS data region size is to check for this
GPU specifically.
Signed-off-by: Timur Tabi <ttabi@nvidia.com>
Link: https://patch.msgid.link/20260430223838.2530778-7-ttabi@nvidia.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
On GA100, the row remapper hardware reserves a small amount of DRAM at
the end of framebuffer for spare rows used to repair memory errors at
runtime. When an uncorrectable ECC error is detected in a DRAM row,
the row remapper redirects accesses to a spare row, transparently
repairing the fault.
The LOCAL_MEMORY_RANGE register (0x100ce0) reports the GPU's FB address
range, but its encoding rounds to 1GB boundaries. On GA100, VBIOS
originally rounded this value down, which could lose up to ~1GB of
usable FB. As a workaround, newer VBIOS instead rounds up to the next
1GB boundary and programs MMU_LOCK (registers 0x1fa82c/0x1fa830) to
mark the gap between the actual usable FB and the rounded-up range as
reserved.
OpenRM's kgspCalculateFbLayout_TU102() handles this by reading the
MMU_LOCK registers and computing the WPR top boundary as:
vbiosReservedOffset = min(mmuLockLo, vgaWorkspaceOffset)
Without this, the WPR region is placed at the top of LOCAL_MEMORY_RANGE,
which overlaps the reserved region. The booter firmware detects this
and rejects the WPR layout.
Add ga100_gsp_mmu_lock_lo() to read the MMU_LOCK range and clamp
gsp->fb.bios.addr accordingly, mirroring OpenRM's behavior.
This is a GA100-only issue. GA102 and later add the
NV_USABLE_FB_SIZE_IN_MB register which reports the correct usable FB
size directly, eliminating the need for the MMU_LOCK workaround.
Signed-off-by: Timur Tabi <ttabi@nvidia.com>
Link: https://patch.msgid.link/20260430223838.2530778-5-ttabi@nvidia.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
In OpenRM's kgspCalculateFbLayout_TU102(), gspFwWprEnd is derived from
vbiosReservedOffset, which is computed as:
vbiosReservedOffset = min(mmuLockLo, vgaWorkspaceOffset)
The VGA workspace offset is one input into this calculation, not the
direct source of gspFwWprEnd. vbiosReservedOffset is the effective
top boundary for WPR2 placement, and it may be lower than the VGA
workspace when VBIOS has locked a region via MMU_LOCK.
In Nouveau, gsp->fb.bios.addr is the equivalent of vbiosReservedOffset,
while gsp->fb.bios.vga_workspace.addr corresponds to the raw VGA
workspace location. The original code assigned vga_workspace.addr to
gspFwWprEnd, which produced the correct result only because bios.addr
was always set equal to vga_workspace.addr and never adjusted.
Use gsp->fb.bios.addr for gspFwWprEnd to correctly mirror OpenRM's
layout logic, so that future adjustments to bios.addr (such as clamping
it to an MMU_LOCK boundary) are properly reflected in the WPR metadata
passed to the booter.
Signed-off-by: Timur Tabi <ttabi@nvidia.com>
Link: https://patch.msgid.link/20260430223838.2530778-4-ttabi@nvidia.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
The booter-load and booter-unload firmware run on the SEC2 falcon.
During tu102_gsp_oneinit(), the booter constructor needs device->sec2
to access the SEC2 falcon.
Without the .sec2 entry, device->sec2 is NULL and this dereference
crashes during GSP-RM boot.
Signed-off-by: Timur Tabi <ttabi@nvidia.com>
Link: https://patch.msgid.link/20260430223838.2530778-3-ttabi@nvidia.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
Paul Kocialkowski
Vitor Soares
Svyatoslav Ryhel
Godswill Onwusilike
Francois Dugast
Michael Kelley
Dongwon Kim
Maxime Ripard
Marco Crivellari
Yicong Hui
Lizhi Hou
Timur Tabi
Hongling Zeng