Finally all the core gem and a lot of drivers are entirely free of
dev->struct_mutex depencies, and we can start to have an entirely
lockless unref path.
To make sure that no one who touches the core code accidentally breaks
existing drivers which still require dev->struct_mutex I've made the
might_lock check unconditional.
While at it de-inline the ref/unref functions, they've become a bit
too big.
v2: Make it not leak like a sieve.
v3: Review from Lucas:
- drop != NULL in pointer checks.
- fixup copypasted kerneldoc to actually match the functions.
v4:
Add __drm_gem_object_unreference as a fastpath helper for drivers who
abolished dev->struct_mutex, requested by Chris.
v5: Fix silly mistake in drm_gem_object_unreference_unlocked caught by
intel-gfx CI - I checked for gem_free_object instead of
gem_free_object_unlocked ...
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Alex Deucher <alexdeucher@gmail.com>
Cc: Lucas Stach <l.stach@pengutronix.de>
Reviewed-by: Lucas Stach <l.stach@pengutronix.de> (v3)
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v4)
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1462178451-1765-1-git-send-email-daniel.vetter@ffwll.ch
Pretty much all of these hasn't seen any action singe 2008 at the very
least. Barring the occasional buildfix and DRM-wide refactoring of
course.
Note: some distributions have stopped shipping their userspace
counterparts for a while. Although some still do.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
When I was writing an atomic wrapper for rmfb, I ran into the
following backtrace from lockdep:
=============================================
[ INFO: possible recursive locking detected ]
4.5.0-patser+ #4696 Tainted: G U
---------------------------------------------
kworker/2:2/2608 is trying to acquire lock:
(crtc_ww_class_mutex){+.+.+.}, at: [<ffffffffc00c9ddc>] drm_modeset_lock+0x7c/0x120 [drm]
but task is already holding lock:
(crtc_ww_class_mutex){+.+.+.}, at: [<ffffffffc00c98cd>] modeset_backoff+0x8d/0x220 [drm]
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0
----
lock(crtc_ww_class_mutex);
lock(crtc_ww_class_mutex);
*** DEADLOCK ***
May be due to missing lock nesting notation
4 locks held by kworker/2:2/2608:
#0: ("events"){.+.+.+}, at: [<ffffffff810a5eea>] process_one_work+0x15a/0x6c0
#1: ((&arg.work)){+.+.+.}, at: [<ffffffff810a5eea>] process_one_work+0x15a/0x6c0
#2: (crtc_ww_class_acquire){+.+.+.}, at: [<ffffffffc004532a>] drm_atomic_helper_remove_fb+0x4a/0x1d0 [drm_kms_helper]
#3: (crtc_ww_class_mutex){+.+.+.}, at: [<ffffffffc00c98cd>] modeset_backoff+0x8d/0x220 [drm]
While lockdep probably catches this bug when it happens, it's better
to explicitly warn when state->acquire_ctx is not set.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1462266751-29123-1-git-send-email-maarten.lankhorst@linux.intel.com
The fsl dcu now uses the clk-provider interfaces, which are not available
when CONFIG_COMMON_CLK is disabled:
drivers/gpu/drm/fsl-dcu/fsl_dcu_drm_drv.c: In function 'fsl_dcu_drm_probe':
drivers/gpu/drm/fsl-dcu/fsl_dcu_drm_drv.c:362:20: error: implicit declaration of function '__clk_get_name' [-Werror=implicit-function-declaration]
pix_clk_in_name = __clk_get_name(pix_clk_in);
This adds a Kconfig dependency to prevent the driver from being enabled
in this case.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Fixes: 2d701449bc ("drm/fsl-dcu: use common clock framework for pixel clock divider")
Acked-by: Stefan Agner <stefan@agner.ch>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1462186839-2224021-1-git-send-email-arnd@arndb.de
This is the first step of renaming async commit to nonblocking commit.
The flag passed by userspace is NONBLOCKING, and async has a different
meaning for page flips, where it means as soon as possible.
Fixing up comments in drm core is done manually, to make sure I didn't
miss anything.
For drivers, the following cocci script is used to rename bool async to bool
nonblock:
@@
identifier I =~ "^async";
identifier func;
@@
func(..., bool
- I
+ nonblock
, ...)
{
<...
- I
+ nonblock
...>
}
@@
identifier func;
type T;
identifier I =~ "^async";
@@
T func(..., bool
- I
+ nonblock
, ...);
Thanks to Tvrtko Ursulin for the cocci script.
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1461679905-30177-2-git-send-email-maarten.lankhorst@linux.intel.com
This adds fbdev deferred io support if CONFIG_FB_DEFERRED_IO is enabled.
The driver has to provide a (struct drm_framebuffer_funcs *)->dirty()
callback to get notification of fbdev framebuffer changes.
If the dirty() hook is set, then fb_deferred_io is set up automatically
by the helper.
Two functions have been added so that the driver can provide a dirty()
function:
- drm_fbdev_cma_init_with_funcs()
This makes it possible for the driver to provided a custom
(struct drm_fb_helper_funcs *)->fb_probe() function.
- drm_fbdev_cma_create_with_funcs()
This is used by the .fb_probe hook to set a driver provided
(struct drm_framebuffer_funcs *)->dirty() function.
Cc: laurent.pinchart@ideasonboard.com
Signed-off-by: Noralf Trønnes <noralf@tronnes.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1461856717-6476-6-git-send-email-noralf@tronnes.org
Export fb_deferred_io_mmap so drivers can change vma->vm_page_prot.
When the framebuffer memory is allocated using dma_alloc_writecombine()
instead of vmalloc(), I get cache syncing problems on ARM.
This solves it:
static int drm_fbdev_cma_deferred_io_mmap(struct fb_info *info,
struct vm_area_struct *vma)
{
fb_deferred_io_mmap(info, vma);
vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
return 0;
}
Could this have been done in the core?
Drivers that don't set (struct fb_ops *)->fb_mmap, gets a call to
fb_pgprotect() at the end of the default fb_mmap implementation
(drivers/video/fbdev/core/fbmem.c). This is an architecture specific
function that on many platforms uses pgprot_writecombine(), but not on
all. And looking at some of the fb_mmap implementations, some of them
sets vm_page_prot to nocache for instance, so I think the safest bet is
to do this in the driver and not in the fbdev core. And we can't call
fb_pgprotect() from fb_deferred_io_mmap() either because we don't have
access to the file pointer that powerpc needs.
Signed-off-by: Noralf Trønnes <noralf@tronnes.org>
Acked-by: Tomi Valkeinen <tomi.valkeinen@ti.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1461856717-6476-5-git-send-email-noralf@tronnes.org
This adds deferred io support to drm_fb_helper.
The fbdev framebuffer changes are flushed using the callback
(struct drm_framebuffer *)->funcs->dirty() by a dedicated worker
ensuring that it always runs in process context.
For those wondering why we need to be able to handle atomic calling
contexts: Both panic paths and cursor code and fbcon blanking can run
from atomic. See
commit bcb39af448
Author: Dave Airlie <airlied@redhat.com>
Date: Thu Feb 7 11:19:15 2013 +1000
drm/udl: make usage as a console safer
for where this was originally discovered.
Signed-off-by: Noralf Trønnes <noralf@tronnes.org>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
[danvet: Augment commit message with why we need to handle atomic
contexts.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1461856717-6476-4-git-send-email-noralf@tronnes.org
