| Message ID | 1449218769-16577-22-git-send-email-daniel.vetter@ffwll.ch (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, Dec 04, 2015 at 09:46:02AM +0100, Daniel Vetter wrote: [...] > diff --git a/include/drm/drm_crtc.h b/include/drm/drm_crtc.h [...] > struct drm_mode_config_funcs { > + /** > + * @fb_create: > + * > + * Create a new framebuffer object. The core does basic checks on the > + * requested metadata, but most of that is left to the driver. See > + * struct &drm_mode_fb_cmd2 for details. > + * > + * RETURNS: > + * > + * A new framebuffer with an initial refernce count of 1 or a negative "reference" > + * error code encoded with ERR_PTR(). > + */ > struct drm_framebuffer *(*fb_create)(struct drm_device *dev, > struct drm_file *file_priv, > const struct drm_mode_fb_cmd2 *mode_cmd); > + > + /** > + * @output_poll_changed: > + * > + * Callback used by helpers to inform the driver of output configuration > + * changes. > + * > + * Drivers implementing fbdev emulation with the helpers can call > + * drm_fb_helper_hotplug_changed from this hook to inform the fbdev > + * helper of output changes. > + * > + * FIXME: > + * > + * Except that there's no vtable for device-level helper callbacks > + * there's no reason this is a core function. > + */ > void (*output_poll_changed)(struct drm_device *dev); > > + /** > + * @atomic_check: > + * > + * This is the only hook to validate an atomic modeset update. This > + * function must reject any modeset and state changes which the hardware > + * or driver doesn't support. This includes but is of course not limited > + * to: > + * > + * - Checking that the modes, framebuffers, scaling and placement > + * requirements and so on are within the limits of the hardware. > + * > + * - Checking that any hidden shared resources are not oversubscribed. > + * This can be shared PLLs, shared lanes, overall memory bandwidth, > + * display fifo space (where shared between planes or maybe even > + * CRTCs). > + * > + * - Checking that virtualized resources exported to userspace are not > + * oversubscribed. For various reasons it can make sense to expose > + * more planes, crtcs or encoders than which are physically there. One > + * example is dual-pipe operations (which generally should be hidden > + * from userspace if when lockstepped in hardware, otherwise exposed), "hidden from userspace when lockstepped in hardware, exposed otherwise"? > + * where a plane might need 1 hardware plane (if it's just on one > + * pipe), 2 hardware planes (when it spans both pipes) or maybe even > + * shared a hardware plane with a 2nd plane (if there's a compatible > + * plane requested on the area handled by the other pipe). > + * > + * - Check that any transitional state is possible and that if > + * requested, the update can indeed be done in the vblank period > + * without temporarily disabling some functions. > + * > + * - Check any other constraints the driver or hardware might have. > + * > + * - This callback also needs to correctly fill out the &drm_crtc_state > + * in this update to make sure that drm_atomic_crtc_needs_modeset() > + * reflects the nature of the possible update and returns true if and > + * only if the update cannot be applied without tearing within one > + * vblank on that CRTC. The core uses that information to reject > + * updates which require a full modeset (i.e. blanking the screen, or > + * at least pausing updates for a substantial amount of time) if > + * userspace has disallowed that in its request. > + * > + * - The driver also does not need to repeat basic input validation > + * like done for the corresponding legacy entry points. The core does > + * that before calling this hook. > + * > + * See the documentation of @atomic_commit for an exhaustive list of > + * error conditions which are allowed to not be checked in this > + * callback. I think this is confusing. Perhaps "list of error conditions which don't have to be checked at the ->atomic_check() stage."? > + * > + * See the documentation for struct &drm_atomic_state for how exactly > + * an atomic modeset update is described. > + * > + * Drivers using the atomic helpers can implement this hook using > + * drm_atomic_helper_check(), or one of the exported sub-functions of > + * it. > + * > + * RETURNS: > + * > + * 0 on success or one of the below negative error codes: > + * > + * - -EINVAL, if any of the above constraints are violated. > + * > + * - -EDEADLK, when returned from an attempt to acquire an additional > + * &drm_modeset_lock through drm_modeset_lock(). > + * > + * - -ENOMEM, if allocating additional state sub-structures failed due > + * to lack of memory. > + * > + * - -EINTR, -EAGAIN or -ERESTARTSYS, if the ioctl should be restarted. > + * This can either be due to a pending signal, or because the driver > + * needs to completely bail out to recover from an exceptional > + * situation like a gpu hang. From a userspace point all errors are > + * treated equally. > + */ > int (*atomic_check)(struct drm_device *dev, > struct drm_atomic_state *a); > + > + /** > + * @atomic_commit: > + * > + * This is the only hook to commit an atomic modeset update. The core > + * guarantees that @atomic_check has been called successfully before > + * calling this function, and that nothing has been changed in the > + * interim. > + * > + * See the documentation for struct &drm_atomic_state for how exactly > + * an atomic modeset update is described. > + * > + * Drivers using the atomic helpers can implement this hook using > + * drm_atomic_helper_commit(), or one of the exported sub-functions of > + * it. > + * > + * Asynchronous commits (as indicated with the async parameter) must > + * do any preparatory work which might result in an unsuccessful commit > + * in the context of this callback. The only exception is hardware "The only exception are hardware errors" > + * errors resulting in -EIO. But even in that case the driver must > + * ensure that the display pipe is at least running, to avoid > + * compositors crashing when pageflips don't work. Anything else, > + * specifically committing the update to the hardware, should be done > + * without blocking the caller. For updates which do not require a > + * modeset this must be guaranteed. > + * > + * The driver must wait for any pending rendering to the new > + * framebuffers to complete before executing the flip. It should also > + * wait for any pending rendering from other drivers if the underlying > + * buffer is a shared dma-buf. Asynchronous commits must not wait for > + * rendering in the context of this callback. > + * > + * An application can request to be notified when the atomic commit has > + * completed. These events are per-CRTC and can be distinguished by the > + * CRTC index supplied in &drm_event to userspace. > + * > + * The drm core will supply a struct &drm_event in the event > + * member of each CRTC's &drm_crtc_state structure. This can be handled by the > + * drm_crtc_send_vblank_event() function, which the driver should call on > + * the provided event upon completion of the atomic commit. Note that if > + * the driver supports vblank signalling and timestamping the vblank > + * counters and timestamps must agree with the ones returned from page > + * flip events. With the current vblank helper infrastructure this can > + * be achieved by holding a vblank reference while the page flip is > + * pending, acquired through drm_crtc_vblank_get() and released with > + * drm_crtc_vblank_put(). Drivers are free to implement their own vblank > + * counter and timestamp tracking though, e.g. if they have accurate > + * timestamp registers in hardware. > + * > + * NOTE: > + * > + * Drivers are not allowed to shut down any display pipe successfully > + * enabled through an atomic commit on their own. Doing so can result in > + * compositors crashing if a page flip is suddenly reject because the "rejected" > + * pipe is off. > + * > + * RETURNS: > + * > + * 0 on success or one of the below negative error codes: > + * > + * - -EBUSY, if an asynchronous updated is requested and there is "update" > + * currently an earlier updated still pending. Drivers are allowed to "update", also I'd drop the "currently" and "still" because they are somewhat redundant: "... and an earlier update is pending." > + * support a queue of outstanding updates, but currently no driver > + * supports that. Note that drivers must wait for preceding updates > + * to complete if a synchronous update is requested, they are not > + * allowed to fail the commit in that case. > + * > + * - -ENOMEM, if the driver failed to allocate memory. Specifically > + * this can happen when trying to pin framebuffers, which must only > + * be done when committing the state. > + * > + * - -ENOSPC, as a refinement of the more generic -ENOMEM to indicate > + * that the driver has run out of vram, iommu space or similar gpu > + * address space needed for framebuffer. > + * > + * - -EIO, if the hardware completely died. > + * > + * - -EINTR, -EAGAIN or -ERESTARTSYS, if the ioctl should be restarted. s/ioctl/IOCTL/ > + * This can either be due to a pending signal, or because the driver > + * needs to completely bail out to recover from an exceptional > + * situation like a gpu hang. From a userspace point of view all errors are s/gpu/GPU/ > + * treated equally. > + * > + * This list is exhaustive. Specifically this hook is not allowed to > + * return -EINVAL (any invalid requests should be caught in > + * @atomic_check) or -EDEADLK (this function must not acquire > + * additional modeset locks). The core will also reject any async > + * atomic flips with -EINVAL already (for matching semantics in this > + * case with legacy page flips). Can you elaborate on this last assertion? Why does the core reject async atomic flips? I don't see where it does that. > + */ > int (*atomic_commit)(struct drm_device *dev, > struct drm_atomic_state *a, Why is the state variable called "a" here? Why not "state"? Same for ->atomic_check() above. > + /** > + * @atomic_state_alloc: > + * > + * This optional hook can be used by drivers who want to subclass struct "... drivers that want ..." > + * &drm_atomic_state to be able to track their own driver-private global > + * state easily. If this hook is implemented, drivers must also > + * implement @atomic_state_clear and @atomic_state_free. > + * > + * RETURNS: > + * > + * A new &drm_atomic_state on success or NULL on failure. > + */ > struct drm_atomic_state *(*atomic_state_alloc)(struct drm_device *dev); > + > + /** > + * @atomic_state_clear: > + * > + * This hook must clear any driver private state duplicated into the > + * passed-in &drm_atomic_state. This hook is called when the caller > + * encountered a &drm_modeset_lock deadlock and needs to drop all > + * already acquired locks as part of the deadlock avoidance dance > + * implemented in drm_modeset_lock_backoff(). > + * > + * Any duplicated state must be invalidated since a concurrent atomic > + * update might change it, and the drm atomic interfaces always apply > + * updates as relative changes to the current state. > + * > + * Drivers who implement this must call drm_atomic_state_default_clear() "Drivers that implement ..." > + * to clear common state. > + */ > void (*atomic_state_clear)(struct drm_atomic_state *state); > + > + /** > + * @atomic_state_free: > + * > + * This hook needs driver private resources and the &drm_atomic_state Did you mean "This hook frees ..."? > + * itself. Note that the core first calls drm_atomic_state_clear to Parentheses after drm_atomic_state_clear? Thierry
On Mon, Dec 07, 2015 at 02:14:48PM +0100, Thierry Reding wrote: > On Fri, Dec 04, 2015 at 09:46:02AM +0100, Daniel Vetter wrote: > > + * This list is exhaustive. Specifically this hook is not allowed to > > + * return -EINVAL (any invalid requests should be caught in > > + * @atomic_check) or -EDEADLK (this function must not acquire > > + * additional modeset locks). The core will also reject any async > > + * atomic flips with -EINVAL already (for matching semantics in this > > + * case with legacy page flips). > > Can you elaborate on this last assertion? Why does the core reject async > atomic flips? I don't see where it does that. Was an ongoing discussion with Daniels Stone about what exactly we should do here. Separate patch, not yet resolved so I dropped that sentence for now - currently core code does indeed not do that. All other suggestions applied, thanks. -Daniel
diff --git a/include/drm/drm_crtc.h b/include/drm/drm_crtc.h index 551484fb55e5..37cff64a26ef 100644 --- a/include/drm/drm_crtc.h +++ b/include/drm/drm_crtc.h @@ -1715,31 +1715,256 @@ struct drm_mode_set { /** * struct drm_mode_config_funcs - basic driver provided mode setting functions - * @fb_create: create a new framebuffer object - * @output_poll_changed: function to handle output configuration changes - * @atomic_check: check whether a given atomic state update is possible - * @atomic_commit: commit an atomic state update previously verified with - * atomic_check() - * @atomic_state_alloc: allocate a new atomic state - * @atomic_state_clear: clear the atomic state - * @atomic_state_free: free the atomic state * * Some global (i.e. not per-CRTC, connector, etc) mode setting functions that * involve drivers. */ struct drm_mode_config_funcs { + /** + * @fb_create: + * + * Create a new framebuffer object. The core does basic checks on the + * requested metadata, but most of that is left to the driver. See + * struct &drm_mode_fb_cmd2 for details. + * + * RETURNS: + * + * A new framebuffer with an initial refernce count of 1 or a negative + * error code encoded with ERR_PTR(). + */ struct drm_framebuffer *(*fb_create)(struct drm_device *dev, struct drm_file *file_priv, const struct drm_mode_fb_cmd2 *mode_cmd); + + /** + * @output_poll_changed: + * + * Callback used by helpers to inform the driver of output configuration + * changes. + * + * Drivers implementing fbdev emulation with the helpers can call + * drm_fb_helper_hotplug_changed from this hook to inform the fbdev + * helper of output changes. + * + * FIXME: + * + * Except that there's no vtable for device-level helper callbacks + * there's no reason this is a core function. + */ void (*output_poll_changed)(struct drm_device *dev); + /** + * @atomic_check: + * + * This is the only hook to validate an atomic modeset update. This + * function must reject any modeset and state changes which the hardware + * or driver doesn't support. This includes but is of course not limited + * to: + * + * - Checking that the modes, framebuffers, scaling and placement + * requirements and so on are within the limits of the hardware. + * + * - Checking that any hidden shared resources are not oversubscribed. + * This can be shared PLLs, shared lanes, overall memory bandwidth, + * display fifo space (where shared between planes or maybe even + * CRTCs). + * + * - Checking that virtualized resources exported to userspace are not + * oversubscribed. For various reasons it can make sense to expose + * more planes, crtcs or encoders than which are physically there. One + * example is dual-pipe operations (which generally should be hidden + * from userspace if when lockstepped in hardware, otherwise exposed), + * where a plane might need 1 hardware plane (if it's just on one + * pipe), 2 hardware planes (when it spans both pipes) or maybe even + * shared a hardware plane with a 2nd plane (if there's a compatible + * plane requested on the area handled by the other pipe). + * + * - Check that any transitional state is possible and that if + * requested, the update can indeed be done in the vblank period + * without temporarily disabling some functions. + * + * - Check any other constraints the driver or hardware might have. + * + * - This callback also needs to correctly fill out the &drm_crtc_state + * in this update to make sure that drm_atomic_crtc_needs_modeset() + * reflects the nature of the possible update and returns true if and + * only if the update cannot be applied without tearing within one + * vblank on that CRTC. The core uses that information to reject + * updates which require a full modeset (i.e. blanking the screen, or + * at least pausing updates for a substantial amount of time) if + * userspace has disallowed that in its request. + * + * - The driver also does not need to repeat basic input validation + * like done for the corresponding legacy entry points. The core does + * that before calling this hook. + * + * See the documentation of @atomic_commit for an exhaustive list of + * error conditions which are allowed to not be checked in this + * callback. + * + * See the documentation for struct &drm_atomic_state for how exactly + * an atomic modeset update is described. + * + * Drivers using the atomic helpers can implement this hook using + * drm_atomic_helper_check(), or one of the exported sub-functions of + * it. + * + * RETURNS: + * + * 0 on success or one of the below negative error codes: + * + * - -EINVAL, if any of the above constraints are violated. + * + * - -EDEADLK, when returned from an attempt to acquire an additional + * &drm_modeset_lock through drm_modeset_lock(). + * + * - -ENOMEM, if allocating additional state sub-structures failed due + * to lack of memory. + * + * - -EINTR, -EAGAIN or -ERESTARTSYS, if the ioctl should be restarted. + * This can either be due to a pending signal, or because the driver + * needs to completely bail out to recover from an exceptional + * situation like a gpu hang. From a userspace point all errors are + * treated equally. + */ int (*atomic_check)(struct drm_device *dev, struct drm_atomic_state *a); + + /** + * @atomic_commit: + * + * This is the only hook to commit an atomic modeset update. The core + * guarantees that @atomic_check has been called successfully before + * calling this function, and that nothing has been changed in the + * interim. + * + * See the documentation for struct &drm_atomic_state for how exactly + * an atomic modeset update is described. + * + * Drivers using the atomic helpers can implement this hook using + * drm_atomic_helper_commit(), or one of the exported sub-functions of + * it. + * + * Asynchronous commits (as indicated with the async parameter) must + * do any preparatory work which might result in an unsuccessful commit + * in the context of this callback. The only exception is hardware + * errors resulting in -EIO. But even in that case the driver must + * ensure that the display pipe is at least running, to avoid + * compositors crashing when pageflips don't work. Anything else, + * specifically committing the update to the hardware, should be done + * without blocking the caller. For updates which do not require a + * modeset this must be guaranteed. + * + * The driver must wait for any pending rendering to the new + * framebuffers to complete before executing the flip. It should also + * wait for any pending rendering from other drivers if the underlying + * buffer is a shared dma-buf. Asynchronous commits must not wait for + * rendering in the context of this callback. + * + * An application can request to be notified when the atomic commit has + * completed. These events are per-CRTC and can be distinguished by the + * CRTC index supplied in &drm_event to userspace. + * + * The drm core will supply a struct &drm_event in the event + * member of each CRTC's &drm_crtc_state structure. This can be handled by the + * drm_crtc_send_vblank_event() function, which the driver should call on + * the provided event upon completion of the atomic commit. Note that if + * the driver supports vblank signalling and timestamping the vblank + * counters and timestamps must agree with the ones returned from page + * flip events. With the current vblank helper infrastructure this can + * be achieved by holding a vblank reference while the page flip is + * pending, acquired through drm_crtc_vblank_get() and released with + * drm_crtc_vblank_put(). Drivers are free to implement their own vblank + * counter and timestamp tracking though, e.g. if they have accurate + * timestamp registers in hardware. + * + * NOTE: + * + * Drivers are not allowed to shut down any display pipe successfully + * enabled through an atomic commit on their own. Doing so can result in + * compositors crashing if a page flip is suddenly reject because the + * pipe is off. + * + * RETURNS: + * + * 0 on success or one of the below negative error codes: + * + * - -EBUSY, if an asynchronous updated is requested and there is + * currently an earlier updated still pending. Drivers are allowed to + * support a queue of outstanding updates, but currently no driver + * supports that. Note that drivers must wait for preceding updates + * to complete if a synchronous update is requested, they are not + * allowed to fail the commit in that case. + * + * - -ENOMEM, if the driver failed to allocate memory. Specifically + * this can happen when trying to pin framebuffers, which must only + * be done when committing the state. + * + * - -ENOSPC, as a refinement of the more generic -ENOMEM to indicate + * that the driver has run out of vram, iommu space or similar gpu + * address space needed for framebuffer. + * + * - -EIO, if the hardware completely died. + * + * - -EINTR, -EAGAIN or -ERESTARTSYS, if the ioctl should be restarted. + * This can either be due to a pending signal, or because the driver + * needs to completely bail out to recover from an exceptional + * situation like a gpu hang. From a userspace point of view all errors are + * treated equally. + * + * This list is exhaustive. Specifically this hook is not allowed to + * return -EINVAL (any invalid requests should be caught in + * @atomic_check) or -EDEADLK (this function must not acquire + * additional modeset locks). The core will also reject any async + * atomic flips with -EINVAL already (for matching semantics in this + * case with legacy page flips). + */ int (*atomic_commit)(struct drm_device *dev, struct drm_atomic_state *a, bool async); + + /** + * @atomic_state_alloc: + * + * This optional hook can be used by drivers who want to subclass struct + * &drm_atomic_state to be able to track their own driver-private global + * state easily. If this hook is implemented, drivers must also + * implement @atomic_state_clear and @atomic_state_free. + * + * RETURNS: + * + * A new &drm_atomic_state on success or NULL on failure. + */ struct drm_atomic_state *(*atomic_state_alloc)(struct drm_device *dev); + + /** + * @atomic_state_clear: + * + * This hook must clear any driver private state duplicated into the + * passed-in &drm_atomic_state. This hook is called when the caller + * encountered a &drm_modeset_lock deadlock and needs to drop all + * already acquired locks as part of the deadlock avoidance dance + * implemented in drm_modeset_lock_backoff(). + * + * Any duplicated state must be invalidated since a concurrent atomic + * update might change it, and the drm atomic interfaces always apply + * updates as relative changes to the current state. + * + * Drivers who implement this must call drm_atomic_state_default_clear() + * to clear common state. + */ void (*atomic_state_clear)(struct drm_atomic_state *state); + + /** + * @atomic_state_free: + * + * This hook needs driver private resources and the &drm_atomic_state + * itself. Note that the core first calls drm_atomic_state_clear to + * avoid code duplicate between the clear and free hooks. + * + * Drivers who implement this must call drm_atomic_state_default_free() + * to release common resources. + */ void (*atomic_state_free)(struct drm_atomic_state *state); };
The meat here is definitely the detailed specs for what atomic_check and atomic_commit are supposed to do. And another candidate for a core vfunc that should be in a helper really (output_poll_changed this time around). v2: Feedback from Eric on irc: - spelling fixes. - spec what async should do - copy the event related paragraphs from page_flip and adjust - make it clear that a successful async commit is not allowed to leave the pipe dead or disabled. v3: Use FIXME comments to annotate functions that we should move to some helpers. Cc: Eric Anholt <eric@anholt.net> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> --- include/drm/drm_crtc.h | 241 +++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 233 insertions(+), 8 deletions(-)