From 5e07724c28f4e06fe42dd5b58bb6f9dd56510567 Mon Sep 17 00:00:00 2001 From: Ben Skeggs Date: Thu, 11 May 2017 17:03:05 +1000 Subject: drm/nouveau/tmr: fix corruption of the pending list when rescheduling an alarm commit 9fc64667ee48c9a25e7dca1a6bcb6906fec5bcc5 upstream. At least therm/fantog "attempts" to work around this issue, which could lead to corruption of the pending alarm list. Fix it properly by not updating the timestamp without the lock held, or trying to add an already pending alarm to the pending alarm list.... Signed-off-by: Ben Skeggs Signed-off-by: Greg Kroah-Hartman --- drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c | 17 ++++++++++------- 1 file changed, 10 insertions(+), 7 deletions(-) (limited to 'drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c') diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c index d4dae1f12d62..4e958db23ca0 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c @@ -65,14 +65,17 @@ nvkm_timer_alarm(struct nvkm_timer *tmr, u32 nsec, struct nvkm_alarm *alarm) struct nvkm_alarm *list; unsigned long flags; - alarm->timestamp = nvkm_timer_read(tmr) + nsec; - - /* append new alarm to list, in soonest-alarm-first order */ + /* Remove alarm from pending list. + * + * This both protects against the corruption of the list, + * and implements alarm rescheduling/cancellation. + */ spin_lock_irqsave(&tmr->lock, flags); - if (!nsec) { - if (!list_empty(&alarm->head)) - list_del(&alarm->head); - } else { + list_del_init(&alarm->head); + + if (nsec) { + /* Insert into pending list, ordered earliest to latest. */ + alarm->timestamp = nvkm_timer_read(tmr) + nsec; list_for_each_entry(list, &tmr->alarms, head) { if (list->timestamp > alarm->timestamp) break; -- cgit v1.2.3 From 9d78e40f5f41ad1db1849f8d15acbda99d0871b4 Mon Sep 17 00:00:00 2001 From: Ben Skeggs Date: Thu, 11 May 2017 17:13:29 +1000 Subject: drm/nouveau/tmr: avoid processing completed alarms when adding a new one commit 330bdf62fe6a6c5b99a647f7bf7157107c9348b3 upstream. The idea here was to avoid having to "manually" program the HW if there's a new earliest alarm. This was lazy and bad, as it leads to loads of fun races between inter-related callers (ie. therm). Turns out, it's not so difficult after all. Go figure ;) Signed-off-by: Ben Skeggs Signed-off-by: Greg Kroah-Hartman --- drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c') diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c index 4e958db23ca0..74e40f1d8f52 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c @@ -80,12 +80,22 @@ nvkm_timer_alarm(struct nvkm_timer *tmr, u32 nsec, struct nvkm_alarm *alarm) if (list->timestamp > alarm->timestamp) break; } + list_add_tail(&alarm->head, &list->head); + + /* Update HW if this is now the earliest alarm. */ + list = list_first_entry(&tmr->alarms, typeof(*list), head); + if (list == alarm) { + tmr->func->alarm_init(tmr, alarm->timestamp); + /* This shouldn't happen if callers aren't stupid. + * + * Worst case scenario is that it'll take roughly + * 4 seconds for the next alarm to trigger. + */ + WARN_ON(alarm->timestamp <= nvkm_timer_read(tmr)); + } } spin_unlock_irqrestore(&tmr->lock, flags); - - /* process pending alarms */ - nvkm_timer_alarm_trigger(tmr); } void -- cgit v1.2.3 From e4add1cf6b4154804350c3385c6d447cff3570de Mon Sep 17 00:00:00 2001 From: Ben Skeggs Date: Thu, 11 May 2017 17:19:48 +1000 Subject: drm/nouveau/tmr: handle races with hw when updating the next alarm time commit 1b0f84380b10ee97f7d2dd191294de9017e94d1d upstream. If the time to the next alarm is short enough, we could race with HW and end up with an ~4 second delay until it triggers. Fix this by checking again after we update HW. Signed-off-by: Ben Skeggs Signed-off-by: Greg Kroah-Hartman --- drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c | 26 +++++++++++++++--------- 1 file changed, 16 insertions(+), 10 deletions(-) (limited to 'drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c') diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c index 74e40f1d8f52..79fcdb43e174 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c @@ -36,23 +36,29 @@ nvkm_timer_alarm_trigger(struct nvkm_timer *tmr) unsigned long flags; LIST_HEAD(exec); - /* move any due alarms off the pending list */ + /* Process pending alarms. */ spin_lock_irqsave(&tmr->lock, flags); list_for_each_entry_safe(alarm, atemp, &tmr->alarms, head) { - if (alarm->timestamp <= nvkm_timer_read(tmr)) - list_move_tail(&alarm->head, &exec); + /* Have we hit the earliest alarm that hasn't gone off? */ + if (alarm->timestamp > nvkm_timer_read(tmr)) { + /* Schedule it. If we didn't race, we're done. */ + tmr->func->alarm_init(tmr, alarm->timestamp); + if (alarm->timestamp > nvkm_timer_read(tmr)) + break; + } + + /* Move to completed list. We'll drop the lock before + * executing the callback so it can reschedule itself. + */ + list_move_tail(&alarm->head, &exec); } - /* reschedule interrupt for next alarm time */ - if (!list_empty(&tmr->alarms)) { - alarm = list_first_entry(&tmr->alarms, typeof(*alarm), head); - tmr->func->alarm_init(tmr, alarm->timestamp); - } else { + /* Shut down interrupt if no more pending alarms. */ + if (list_empty(&tmr->alarms)) tmr->func->alarm_fini(tmr); - } spin_unlock_irqrestore(&tmr->lock, flags); - /* execute any pending alarm handlers */ + /* Execute completed callbacks. */ list_for_each_entry_safe(alarm, atemp, &exec, head) { list_del_init(&alarm->head); alarm->func(alarm); -- cgit v1.2.3 From 5dffc1be6552bff51a99ef78820fc85fc84277e2 Mon Sep 17 00:00:00 2001 From: Ben Skeggs Date: Mon, 5 Jun 2017 17:23:32 +1000 Subject: drm/nouveau/tmr: fully separate alarm execution/pending lists commit b4e382ca7586a63b6c1e5221ce0863ff867c2df6 upstream. Reusing the list_head for both is a bad idea. Callback execution is done with the lock dropped so that alarms can be rescheduled from the callback, which means that with some unfortunate timing, lists can get corrupted. The execution list should not require its own locking, the single function that uses it can only be called from a single context. Signed-off-by: Ben Skeggs Signed-off-by: Greg Kroah-Hartman --- drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c') diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c index 79fcdb43e174..46033909d950 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/base.c @@ -50,7 +50,8 @@ nvkm_timer_alarm_trigger(struct nvkm_timer *tmr) /* Move to completed list. We'll drop the lock before * executing the callback so it can reschedule itself. */ - list_move_tail(&alarm->head, &exec); + list_del_init(&alarm->head); + list_add(&alarm->exec, &exec); } /* Shut down interrupt if no more pending alarms. */ @@ -59,8 +60,8 @@ nvkm_timer_alarm_trigger(struct nvkm_timer *tmr) spin_unlock_irqrestore(&tmr->lock, flags); /* Execute completed callbacks. */ - list_for_each_entry_safe(alarm, atemp, &exec, head) { - list_del_init(&alarm->head); + list_for_each_entry_safe(alarm, atemp, &exec, exec) { + list_del(&alarm->exec); alarm->func(alarm); } } -- cgit v1.2.3