Hi,

this series looks quite good already. However it is very long and I'll need more time to review the connector-orientation code. I cannot say much about the DT bindings.

So shall we merge patches 1 to 6 and 8 first? You'd have something in upstream already and not chase a moving target. From there, we can merge the other features in the coming weeks.

Please see my few comments on the patch 1 to 8.

Best regards
Thomas


Am 27.10.26 um 14:53 schrieb Màxim Pedraza Padilla:
This is the sixth version of the splash client, which I took over from
Francesco with v4 [1]. It works in Thomas's review of v5 [13] and Rob's
of the binding [18], fixes what sashiko found in v5 [11][12], and adds
three patches: the BMP headers move to <video/bmp.h> to be shared with
efifb (patch 3), the choice of the image source gets a patch of its own
(patch 8), and the splash follows the connector's orientation
(patch 14). The command line parameters are gone, as Thomas asked [20].

Why this is in the kernel
-------------------------

Maxime asked in v1 why this should not be left to userspace or to the
bootloader. The cases it is for are the ones neither covers:

   - Nothing initialises the display before the kernel does. U-Boot's
     Falcon mode, meant to cut boot time, starts the kernel from the SPL
     and never runs U-Boot proper, where display support lives: there is
     no bootloader splash to hand over, and the device tree is the only
     thing that reaches the kernel. Userspace comes seconds later.

   - One kernel image for several products. An image built into the
     kernel, with CONFIG_EXTRA_FIRMWARE, means one kernel per logo, or
     every logo in every product's kernel, all of them staying in memory
     for good. Francesco described exactly this case in v2: one board
     support package for a family of devices, each with its own logo and
     background, picked by the bootloader.

   - A splash handed over from the bootloader does not always survive the
     native driver taking over the display.

Where the image comes from
--------------------------

In this order: the device tree, the EFI BGRT, and a BMP loaded as
firmware, named at build time ("drm_splash.bmp" by default); with none
of them, only the background is drawn, in a colour set at build time
or, for a device tree image, by its node.

The device tree node lives under /chosen (patch 7) and carries the BMP
either in its "logo-image" property, which dtc fills in with /incbin/,
or in a reserved memory region the bootloader loaded it into. The
region keeps the image out of the device tree, so that it can be
changed from userspace by updating a partition, as Francesco suggested.
A node is only there if someone put it there for that board, so it goes
before the BGRT.

Placement uses a position with -1 centring an axis, plus an offset, in
screen pixels. A rotation turns the image and not the screen, counter
clockwise as for panels and as DRM_MODE_ROTATE_* count. A device tree
image is placed as its node says, a BGRT image at the table offsets, and
anything else is centred.

BGRT images with the ACPI 6.2 orientation bits set are no longer skipped
but turned (patch 13). I took the meaning of the bits and of the offsets
from Plymouth, the main user of the BGRT, so that the kernel and
Plymouth put the image in the same place and the handover does not
jump. Plymouth takes orientation 1 (90 degrees) for a panel mounted
left side up and 3 (270) for right side up, which DRM counts as
DRM_MODE_ROTATE_90 and DRM_MODE_ROTATE_270; with the bits set it draws
the image upright at offsets given on the turned screen, and without
them it takes the image as already turned for a turned panel, at
offsets on the native screen. Patches 13 and 14 do the same. I have
not found the specification text to check against; corrections are
welcome.

Changes since v5
----------------

 From Thomas's review [13][14][15]:

   - Patch 4 draws from a work item on system_power_efficient_wq instead
     of a render thread. One mutex covers the scanouts, the image and
     the firmware handed over by its callback, and the image is kept
     until unregister, so that the work can run again.
   - Every hotplug sets the outputs up again, so an output plugged in
     later gets the splash, and one unplugged is no longer driven. The
     buffers on screen stay there until the new ones are committed, by
     the work or by a restore, so the outputs do not go dark in between,
     nor when a commit or the hotplug fails; and a hotplug that fails no
     longer stops the later ones.
   - Restore commits whatever the modesets hold: an output that has no
     buffer has no mode either, and is switched off rather than left
     showing something else.
   - New patch 3: the BMP headers move from efifb to include/video/bmp.h,
     as __le16/__le32 fields; efifb reads them with get_unaligned_le*(),
     with no change on the little endian systems EFI runs on.
   - The image is drawn in two steps: the BMP is read, flipped and turned
     into an XRGB8888 buffer, which is then converted into the
     framebuffer's format; filling and blitting go by the format's
     drm_format_info rather than by bpp. Rotation (patch 11) only
     touches the first step.
   - No command line parameters [13][16][17][20]: patch 4 drops
     splash_color and splash_bmp, the colour and the BMP's file name are
     set at build time (CONFIG_DRM_CLIENT_SPLASH_BMP_NAME is new), and
     the v5 patch that added splash_bmp=none, splash_pos, splash_offset
     and splash_rotation is gone.
   - New patch 8: the image source is an enum, picked in one place, so
     that the device tree patch only adds its entry; the switch on it
     has no default, so that -Wswitch catches a source left out.
   - Messages are drm_dbg(), selects in alphabetical order, the mutex is
     destroyed at unregister, and the commit message of patch 4 carries
     the changes of each version.

 From Rob's review [18]:

   - Patch 7: "image" is now "logo-image", and "rotation" is the one of
     panel-common.yaml, keeping its own description, as it turns the
     image rather than say how a panel is mounted.

 From sashiko's review of v5 [11][12], and further review:

   - The 24 bpp source pixels are read with get_unaligned_le24() rather
     than a u32 load, which read a byte past the last pixel of the image
     and was unaligned [11], and the BGRT offsets are checked without the
     addition that could overflow [11]. Later patches already did; the
     fix now comes with the code it fixes.
   - The device tree source holds the node for as long as it draws the
     image in place, and drops it with the image [12].
   - Patch 4: a tiled framebuffer is sized from the tile offsets, in
     whatever order the CRTCs come, and a tile whose plane cannot take
     its group's format is left out; a preferred format that cannot be
     drawn to, such as C8, is not picked; a modeset left without a buffer
     is switched off instead of failing the commit for every output, and
     never keeps the framebuffer it had; nothing is drawn or committed
     while the device is suspended, a restore included, and the work is
     queued again on resume; only formats the plane takes
     with the linear modifier are picked; a background drawn again gets
     the image drawn again on top; the V4 and V5 BMP headers, which GIMP
     and ImageMagick write, are accepted; the build time background
     colour is limited to RGB888; the Kconfig help says that the BMP is
     requested only once.
   - Patch 2: efi_bgrt_has_image() says no when the image address, a u64
     in the table, does not fit in phys_addr_t, so that it cannot
     disagree with efi_bgrt_image_address() on 32-bit kernels without
     PAE. Thomas, I have kept your Reviewed-by, as said in [19].
   - Patch 7: the node is a "logo" node under /chosen with compatible
     "boot-logo"; logo-position and logo-offset give their defaults, and
     "logo-image" says that some architectures limit the size of the
     device tree (arm64 to 2 MB), so a large image belongs in a reserved
     memory region.
   - Patch 9: the image is found when the source is picked, and a node
     whose image cannot be had (a region that does not resolve or cannot
     be mapped, or "memory-region" on a kernel without
     CONFIG_OF_RESERVED_MEM) is warned about once and skipped, so the
     other sources are tried, rather than giving no image.
   - Patch 10: logo-position and logo-offset must hold two cells each,
     within the binding's range; otherwise the kernel warns and keeps the
     default.
   - Patch 12: a background-color that is not one cell of RGB888 is
     warned about and left out.
   - New patch 14: the splash follows the connector's orientation, the
     panel's or video=...,rotate=, as fbdev does: the plane turns the
     buffer when it can, and the splash turns it in software otherwise.
     A BGRT image stays as the firmware drew it, and tiles that need
     different turns get none in software, as with fbdev; so does a
     reflection, which is not drawn, and the legacy commit, which leaves
     the plane upright, gets the whole turn in software. It can be
     dropped; see the open questions.
   - Patch 15: the documentation follows all of the above.

Changes in v5
-------------

   - efi: bgrt: patch 2 adds accessors for the boot graphics instead of
     exporting bgrt_tab and bgrt_image_size, with stubs in efi-bgrt.h for
     CONFIG_ACPI_BGRT=n (Jani [2]); both variables are __ro_after_init
     (Ard [3]). Ard, this is the shape we discussed [4]: it is no longer
     the patch you acked, so I have not carried your Acked-by.
   - The client reads the BGRT through those accessors only, and drops
     its own wrappers, its stubs and the #ifdef around them; no "static
     inline" in drm_splash.c any more (Jani [5]).
   - Fixes from sashiko's review [6]:
     - when the render thread could not be created, the framebuffers
       were freed but the modesets still pointed at them, and the next
       restore committed them: a use-after-free that closing the last
       DRM file is enough to trigger;
     - when no buffer could be created, the scanouts were not freed, and
       a later restore committed a mode without a framebuffer;
     - a BMP loaded as firmware that arrived after the render thread had
       left (another master took over) was kept until unbind; it is now
       released when it arrives.
   - The binding and the documentation now ask a bootloader that loads no
     image into the reserved region to disable or remove the node, rather
     than to clear the region, which left the kernel an invalid image to
     complain about on every boot (sashiko [7]).
   - Dropped the RFC tag.

Changes in v4
-------------

Fixes to Francesco's patches, folded in and listed in the commit notes;
most of them were also in sashiko's review of v3 [9]:

   - the render thread stayed in TASK_UNINTERRUPTIBLE for good when the
     BMP firmware was missing (hung task warning), could miss a wake up,
     and kthread_stop() was called on a thread that had already exited
     (use-after-free), never existed or failed to start;
   - a use-after-free on unbind: the image was released after
     drm_client_release(), which frees the client;
   - a use-after-free in the firmware callback when the device goes away
     first, as the request cannot be cancelled; unregister now waits;
   - the BMP bounds check could overflow and let a crafted bitmap offset
     read 4 GiB away;
   - the 24 bit blitters read a byte past the image;
   - buffers went to the wrong modesets when one was skipped, tiled
     groups dereferenced a buffer before creating it and swapped width
     and height, a failed dumb buffer left an ERR_PTR behind, and the
     modeset list was walked without modeset_mutex;
   - the BGRT symbols are exported (Mario's report [10]), GPL only;
   - the DRM_CLIENT_DEFAULT entry is indented with tabs (patch 1).

Additions: the device tree source and its binding, placement, rotation,
the background colour following the image, BGRT orientation, the
command line parameters, and the documentation.

Testing
-------

   - QEMU versatilepb (pl111, 16 bpp): 45 cases for the device tree
     source (inline and region, invalid nodes and values, placement,
     clamping, rotations, rows with and without padding, backgrounds, a
     V5 BMP header), plus 3 for the order of the sources, among them a
     region that cannot be read falling back to the BMP loaded as
     firmware, each compared pixel by pixel.
   - QEMU virt arm64 (virtio-gpu, 32 bpp) with KASAN and kmemleak, with
     the BMP loaded as firmware from an initramfs and the driver bound
     and unbound repeatedly: no reports, also with a crafted bitmap
     offset and with the firmware arriving only after the unbind (sysfs
     fallback).
   - Hotplug on vkms: when the connection moves from one connector to
     another after the splash is up, a restore drives the new one, and
     the old buffers go only once the new ones are committed; with the
     first hotplug forced to fail, the next one brings the splash up.
   - Several outputs: with the first output off and the second on, v3 put
     the buffer on the dead output and left the live one blank; now the
     splash is on the live one.
   - A tiled display, simulated by forcing on two virtio-gpu heads EDIDs
     with a DisplayID tiled topology block (2x1 tiles of 1024x768): one
     2048x768 buffer, the image centred across both tiles; v3 oopsed on
     such a display.
   - The error paths, with a debug patch forcing them: a buffer or a
     format that fails on one output of two leaves that one off and the
     splash on the other; work queued while suspended runs on resume; a
     background that fails to draw is drawn on the next try; no WARN or
     KASAN report in any of them.
   - x86-64 with OVMF for the BGRT through the new accessors, including
     forced orientation bits, a connector turned with video=...,rotate=,
     and drm_client_lib built as a module.
   - efifb before and after patch 3, with the BGRT replaced from the
     initrd to move the image: efifb draws it at the new place, the same
     pixel for pixel with and without the patch.
   - i386 without PAE on 32-bit OVMF, with the BGRT replaced from the
     initrd so that it points above 4 GiB: without the patch 2 check the
     splash draws the image at the truncated address; with it, the
     background alone.
   - An AM335x board (tilcdc, 800x480 panel of which only the bottom 297
     lines are visible, 16 bpp), booting over TFTP and NFS: 6 cases,
     plus 6 with the BMP loaded as firmware from an initramfs, and 5
     unbind and bind cycles of tilcdc, checked by reading the scanout
     buffer back through /dev/mem.
   - The handover on the same board, with the product's Qt5 application
     drawing through EGLFS and the SGX GPU: it takes the display straight
     from the splash, and when it exits the splash is restored, pixel for
     pixel.
   - Falcon mode, on an AM335x board booting from NAND: the SPL starts
     the kernel directly, with the device tree appended to it and the
     image inline in the node, and U-Boot proper never runs. The splash
     comes up as the node asks, checked pixel by pixel.
   - W=1 and sparse clean; every patch builds on its own on arm and
     x86-64; checkpatch --strict clean apart from a quoted modpost line,
     the MAINTAINERS reminder on patch 4, answered by patch 5, and the
     one on patch 3, whose new header FRAMEBUFFER LAYER already covers;
     dt_binding_check clean for the binding.

Not tested: a BGRT with orientation bits from real firmware, nor one
on a plane that turns in hardware.

Open questions
--------------

   - The second drawing step, from XRGB8888 into the framebuffer's
     format, converts each pixel itself for now. It is kept apart so that
     it can move onto a shared drawing interface for clients once there
     is one [15].
   - Rob: how the image should be described, a BMP in a property or raw
     pixels described like simple-framebuffer [8]. The binding here uses
     the BMP; the client would need a second drawing path for the other.
   - Once the binding settles I will send the dt-schema change that lets
     the node live under /chosen; until then dtbs_check flags it.
   - The splash is a DRM client, so a device using it has no fbdev
     emulation and no /dev/fb0. That rules it out where userspace still
     draws through fbdev.
   - Unbinding waits for a pending firmware request, as the request cannot
     be cancelled and its callback uses the splash, and it waits under
     the device's clientlist_mutex. Normally the request is long over; with
     the sysfs fallback enabled it can take up to the loading timeout
     (60 s by default). Not waiting would mean splitting what the callback
     uses into its own refcounted object; I left it as is, but can do it
     if preferred.
   - Patch 14 makes the splash follow the connector's orientation the way
     fbcon does: what the plane does not turn in hardware is turned in
     software, so the logo looks the same on every driver, and the logo's
     own rotation is relative to what is seen. Without it the logo
     follows the plane, which matches userspace that ignores the panel
     orientation, but the result depends on the driver (and can change
     when a native driver takes over from simpledrm). If following the
     plane is preferred, patch 14 can simply be dropped.
   - The BGRT offsets are for the mode the firmware drew in. When KMS
     picks another one (a non-atomic driver, a secondary output, or
     video=), the x86 sanity check taken from efifb leaves the screen
     without a logo, and elsewhere the logo lands off centre. The offsets
     could be used only when the scanout matches the firmware's
     framebuffer, as screen_info describes it, and the image centred
     otherwise; I have left it as efifb has it for now.

Based on drm-misc-next-2026-10-23 (cf5d8e390e2a).

[1] https://lore.kernel.org/all/aryqYhHqFk3cXh9d@bywater/
[2] 
https://lore.kernel.org/all/[email protected]/
[3] 
https://lore.kernel.org/all/[email protected]/
[4] 
https://lore.kernel.org/all/[email protected]/
[5] 
https://lore.kernel.org/all/[email protected]/
[6] https://lore.kernel.org/all/[email protected]/
[7] https://lore.kernel.org/all/[email protected]/
[8] 
https://lore.kernel.org/all/CAEUXW=Gh6wMk1i=x_KT0eSdkHeHs8Hkxkj4hfsuTRyY6=ap...@mail.gmail.com/
[9] 
https://sashiko.dev/#/patchset/20260510-drm_client_splash-v3-0-a9aee9f0b2fc%40valla.it
[10] 
https://lore.kernel.org/dri-devel/[email protected]/
[11] https://lore.kernel.org/all/[email protected]/
[12] https://lore.kernel.org/all/[email protected]/
[13] https://lore.kernel.org/all/[email protected]/
[14] https://lore.kernel.org/all/[email protected]/
[15] https://lore.kernel.org/all/[email protected]/
[16] https://lore.kernel.org/all/[email protected]/
[17] https://lore.kernel.org/all/[email protected]/
[18] https://lore.kernel.org/all/[email protected]/
[19] 
https://lore.kernel.org/all/[email protected]/
[20] https://lore.kernel.org/all/[email protected]/

v5: https://lore.kernel.org/r/[email protected]
v4: https://lore.kernel.org/r/[email protected]
v3: 
https://lore.kernel.org/r/[email protected]
v2: 
https://lore.kernel.org/r/[email protected]
v1: 
https://lore.kernel.org/r/[email protected]

Màxim

Francesco Valla (3):
   drm: client: add splash client
   MAINTAINERS: add entry for DRM splash client
   drm: docs: remove bootsplash from TODO

Màxim Pedraza Padilla (12):
   drm/clients: Kconfig: indent DRM_CLIENT_DEFAULT with tabs
   efi: bgrt: add accessors for the boot graphics
   fbdev: efifb: move the BMP headers to <video/bmp.h>
   dt-bindings: display: add a boot logo node under /chosen
   drm/client: splash: pick the image source in one place
   drm/client: splash: add a device tree image source
   drm/client: splash: place the device tree image where it asks
   drm/client: splash: turn the device tree image as it asks
   drm/client: splash: take the background colour from the image source
   drm/client: splash: turn the BGRT image on panels mounted turned
   drm/client: splash: follow the connector's orientation
   drm/client: splash: document the image sources

  .../bindings/display/boot-logo.yaml           |  156 ++
  Documentation/gpu/drm-client.rst              |    6 +
  Documentation/gpu/todo.rst                    |   17 -
  MAINTAINERS                                   |    9 +
  drivers/firmware/efi/efi-bgrt.c               |   75 +-
  drivers/gpu/drm/clients/Kconfig               |   96 +-
  drivers/gpu/drm/clients/Makefile              |    1 +
  drivers/gpu/drm/clients/drm_client_internal.h |    9 +
  drivers/gpu/drm/clients/drm_client_setup.c    |    8 +
  drivers/gpu/drm/clients/drm_splash.c          | 1482 +++++++++++++++++
  drivers/video/fbdev/efifb.c                   |   49 +-
  include/linux/efi-bgrt.h                      |   39 +-
  include/video/bmp.h                           |   40 +
  13 files changed, 1930 insertions(+), 57 deletions(-)
  create mode 100644 Documentation/devicetree/bindings/display/boot-logo.yaml
  create mode 100644 drivers/gpu/drm/clients/drm_splash.c
  create mode 100644 include/video/bmp.h


base-commit: cf5d8e390e2ac2cfa24a50645efdbb9f0eebe65f

--
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com
GF: Stefan Gaiser, Jochen Jaser, Abhinav Puri, (HRB 36809, AG Nürnberg)


Reply via email to