Instead check for master status, in case we've raced.

This is the last exception to the general rule that we restore fbcon only when there's no master active. Compositors are supposed to drop their master status before they switch to a different console back to text mode (or just switch to text mode directly, without a vt switch).

This is known to break some subtests of kms_fbcon_fbt in igt, but they're just wrong - it does a graphics/text mode switch for the vt without updating the master status.

Also add a comment to the drm_client->restore hook that this is expected going forward from all clients (there's currently just one).

v2: Also drop the force in pan_display

v3: Restore the _force to pan_display, this actually means _locked in that path. Spotted by Noralf.

Cc: Noralf Trønnes noralf@tronnes.org Reviewed-by: Noralf Trønnes noralf@tronnes.org Signed-off-by: Daniel Vetter daniel.vetter@intel.com --- drivers/gpu/drm/drm_fb_helper.c | 12 +----------- include/drm/drm_client.h | 5 +++++ 2 files changed, 6 insertions(+), 11 deletions(-)

diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c index 4c7cbce7bae7..672934e0eeed 100644 --- a/drivers/gpu/drm/drm_fb_helper.c +++ b/drivers/gpu/drm/drm_fb_helper.c @@ -250,17 +250,7 @@ int drm_fb_helper_restore_fbdev_mode_unlocked(struct drm_fb_helper *fb_helper) return 0;

mutex_lock(&fb_helper->lock); - /* - * TODO: - * We should bail out here if there is a master by dropping _force. - * Currently these igt tests fail if we do that: - * - kms_fbcon_fbt@psr - * - kms_fbcon_fbt@psr-suspend - * - * So first these tests need to be fixed so they drop master or don't - * have an fd open. - */ - ret = drm_client_modeset_commit_force(&fb_helper->client); + ret = drm_client_modeset_commit(&fb_helper->client);

do_delayed = fb_helper->delayed_hotplug; if (do_delayed) diff --git a/include/drm/drm_client.h b/include/drm/drm_client.h index 5cf2c5dd8b1e..d01d311023ac 100644 --- a/include/drm/drm_client.h +++ b/include/drm/drm_client.h @@ -44,6 +44,11 @@ struct drm_client_funcs { * returns zero gets the privilege to restore and no more clients are * called. This callback is not called after @unregister has been called. * + * Note that the core does not guarantee exclusion against concurrent + * drm_open(). Clients need to ensure this themselves, for example by + * using drm_master_internal_acquire() and + * drm_master_internal_release(). + * * This callback is optional. */ int (*restore)(struct drm_client_dev *client);

We want to only take the BKL on crap drivers, but to know whether we have a crap driver we first need to look it up. Split this shuffle out from the main BKL-disabling patch, for more clarity. Historical aside: When the kernel-wide BKL was removed, it was replaced by drm_global_mutex within the scope of the drm subsystem hence why these two things are (almost) interchangeable as concepts here.

Since the minors are refcounted drm_minor_acquire is purely internal and this does not have a driver visible effect.

v2: Push the locking even further into drm_open(), suggested by Chris. This gives us more symmetry with drm_release(), and maybe a futuer avenue where we make drm_global_mutex locking (partially) opt-in like with drm_release_noglobal().

v3: - Actually push this stuff correctly, don't unlock twice (Chris) - Fix typo on commit message, plus explain why BKL = drm_global_mutex (Sam)

Cc: Sam Ravnborg sam@ravnborg.org Cc: Chris Wilson chris@chris-wilson.co.uk Signed-off-by: Daniel Vetter daniel.vetter@intel.com --- drivers/gpu/drm/drm_drv.c | 14 +++++--------- drivers/gpu/drm/drm_file.c | 6 ++++++ 2 files changed, 11 insertions(+), 9 deletions(-)

diff --git a/drivers/gpu/drm/drm_drv.c b/drivers/gpu/drm/drm_drv.c index 8deff75b484c..05bdf0b9d2b3 100644 --- a/drivers/gpu/drm/drm_drv.c +++ b/drivers/gpu/drm/drm_drv.c @@ -1085,17 +1085,14 @@ static int drm_stub_open(struct inode *inode, struct file *filp)

DRM_DEBUG("\n");

- mutex_lock(&drm_global_mutex); minor = drm_minor_acquire(iminor(inode)); - if (IS_ERR(minor)) { - err = PTR_ERR(minor); - goto out_unlock; - } + if (IS_ERR(minor)) + return PTR_ERR(minor);

new_fops = fops_get(minor->dev->driver->fops); if (!new_fops) { err = -ENODEV; - goto out_release; + goto out; }

replace_fops(filp, new_fops); @@ -1104,10 +1101,9 @@ static int drm_stub_open(struct inode *inode, struct file *filp) else err = 0;

-out_release: +out: drm_minor_release(minor); -out_unlock: - mutex_unlock(&drm_global_mutex); + return err; }

diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c index 1075b3a8b5b1..80d556402ab4 100644 --- a/drivers/gpu/drm/drm_file.c +++ b/drivers/gpu/drm/drm_file.c @@ -378,6 +378,8 @@ int drm_open(struct inode *inode, struct file *filp) if (IS_ERR(minor)) return PTR_ERR(minor);

+ mutex_lock(&drm_global_mutex); + dev = minor->dev; if (!atomic_fetch_inc(&dev->open_count)) need_setup = 1; @@ -395,10 +397,14 @@ int drm_open(struct inode *inode, struct file *filp) goto err_undo; } } + + mutex_unlock(&drm_global_mutex); + return 0;

err_undo: atomic_dec(&dev->open_count); + mutex_unlock(&drm_global_mutex); drm_minor_release(minor); return retcode; }

This catches the majority of drivers (unfortunately not if we take users into account, because all the big drivers have at least a lastclose hook).

With the prep patches out of the way all drm state is fully protected and either prevents or can deal with the races from dropping the BKL around open/close. The only thing left to audit are the various driver hooks - by keeping the BKL around if any of them are set we have a very simple cop-out!

Note that one of the biggest prep pieces to get here was making dev->open_count atomic, which was done in

commit 7e13ad896484a0165a68197a2e64091ea28c9602 Author: Chris Wilson chris@chris-wilson.co.uk Date: Fri Jan 24 13:01:07 2020 +0000

drm: Avoid drm_global_mutex for simple inc/dec of dev->open_count

v2: - Rebase and fix locking in drm_open() (Chris) - Indentation fix in drm_release - Typo fix in the commit message (Sam)

Cc: Chris Wilson chris@chris-wilson.co.uk Cc: Sam Ravnborg sam@ravnborg.org Signed-off-by: Daniel Vetter daniel.vetter@intel.com --- drivers/gpu/drm/drm_drv.c | 6 +++-- drivers/gpu/drm/drm_file.c | 48 +++++++++++++++++++++++++++++----- drivers/gpu/drm/drm_internal.h | 1 + 3 files changed, 47 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/drm_drv.c b/drivers/gpu/drm/drm_drv.c index 05bdf0b9d2b3..9fcd6ab3c154 100644 --- a/drivers/gpu/drm/drm_drv.c +++ b/drivers/gpu/drm/drm_drv.c @@ -946,7 +946,8 @@ int drm_dev_register(struct drm_device *dev, unsigned long flags) struct drm_driver *driver = dev->driver; int ret;

- mutex_lock(&drm_global_mutex); + if (drm_dev_needs_global_mutex(dev)) + mutex_lock(&drm_global_mutex);

if (dev->driver->load) { if (!drm_core_check_feature(dev, DRIVER_LEGACY)) @@ -992,7 +993,8 @@ int drm_dev_register(struct drm_device *dev, unsigned long flags) drm_minor_unregister(dev, DRM_MINOR_PRIMARY); drm_minor_unregister(dev, DRM_MINOR_RENDER); out_unlock: - mutex_unlock(&drm_global_mutex); + if (drm_dev_needs_global_mutex(dev)) + mutex_unlock(&drm_global_mutex); return ret; } EXPORT_SYMBOL(drm_dev_register); diff --git a/drivers/gpu/drm/drm_file.c b/drivers/gpu/drm/drm_file.c index 80d556402ab4..c4c704e01961 100644 --- a/drivers/gpu/drm/drm_file.c +++ b/drivers/gpu/drm/drm_file.c @@ -51,6 +51,37 @@ /* from BKL pushdown */ DEFINE_MUTEX(drm_global_mutex);

+bool drm_dev_needs_global_mutex(struct drm_device *dev) +{ + /* + * Legacy drivers rely on all kinds of BKL locking semantics, don't + * bother. They also still need BKL locking for their ioctls, so better + * safe than sorry. + */ + if (drm_core_check_feature(dev, DRIVER_LEGACY)) + return true; + + /* + * The deprecated ->load callback must be called after the driver is + * already registered. This means such drivers rely on the BKL to make + * sure an open can't proceed until the driver is actually fully set up. + * Similar hilarity holds for the unload callback. + */ + if (dev->driver->load || dev->driver->unload) + return true; + + /* + * Drivers with the lastclose callback assume that it's synchronized + * against concurrent opens, which again needs the BKL. The proper fix + * is to use the drm_client infrastructure with proper locking for each + * client. + */ + if (dev->driver->lastclose) + return true; + + return false; +} + /** * DOC: file operations * @@ -378,9 +409,10 @@ int drm_open(struct inode *inode, struct file *filp) if (IS_ERR(minor)) return PTR_ERR(minor);

- mutex_lock(&drm_global_mutex); - dev = minor->dev; + if (drm_dev_needs_global_mutex(dev)) + mutex_lock(&drm_global_mutex); + if (!atomic_fetch_inc(&dev->open_count)) need_setup = 1;

@@ -398,13 +430,15 @@ int drm_open(struct inode *inode, struct file *filp) } }

- mutex_unlock(&drm_global_mutex); + if (drm_dev_needs_global_mutex(dev)) + mutex_unlock(&drm_global_mutex);

return 0;

err_undo: atomic_dec(&dev->open_count); - mutex_unlock(&drm_global_mutex); + if (drm_dev_needs_global_mutex(dev)) + mutex_unlock(&drm_global_mutex); drm_minor_release(minor); return retcode; } @@ -444,7 +478,8 @@ int drm_release(struct inode *inode, struct file *filp) struct drm_minor *minor = file_priv->minor; struct drm_device *dev = minor->dev;

- mutex_lock(&drm_global_mutex); + if (drm_dev_needs_global_mutex(dev)) + mutex_lock(&drm_global_mutex);

DRM_DEBUG("open_count = %d\n", atomic_read(&dev->open_count));

@@ -453,7 +488,8 @@ int drm_release(struct inode *inode, struct file *filp) if (atomic_dec_and_test(&dev->open_count)) drm_lastclose(dev);

- mutex_unlock(&drm_global_mutex); + if (drm_dev_needs_global_mutex(dev)) + mutex_unlock(&drm_global_mutex);

drm_minor_release(minor);

diff --git a/drivers/gpu/drm/drm_internal.h b/drivers/gpu/drm/drm_internal.h index 6937bf923f05..aeec2e68d772 100644 --- a/drivers/gpu/drm/drm_internal.h +++ b/drivers/gpu/drm/drm_internal.h @@ -41,6 +41,7 @@ struct drm_printer;

/* drm_file.c */ extern struct mutex drm_global_mutex; +bool drm_dev_needs_global_mutex(struct drm_device *dev); struct drm_file *drm_file_alloc(struct drm_minor *minor); void drm_file_free(struct drm_file *file); void drm_lastclose(struct drm_device *dev);