@@ -137,8 +137,22 @@ static void nand_release_device(struct mtd_info *mtd)
/* Release the controller and the chip */
spin_lock(&chip->controller->lock);
chip->controller->active = NULL;
- chip->state = FL_READY;
- wake_up(&chip->controller->wq);
+
+ if (waitqueue_active(&chip->controller->wq)) {
+ wait_queue_head_t *q;
+ wait_queue_t *waiter;
+
+ q = &chip->controller->wq;
+ chip->state = FL_HANDOVER;
+
+ waiter = list_first_entry(&q->task_list, wait_queue_t,
+ task_list);
+
+ chip->controller->handover_waiter = waiter;
+ wake_up(q);
+ } else {
+ chip->state = FL_READY;
+ }
spin_unlock(&chip->controller->lock);
}
@@ -843,10 +857,18 @@ nand_get_device(struct mtd_info *mtd, int new_state)
if (!chip->controller->active)
chip->controller->active = chip;
- if (chip->controller->active == chip && chip->state == FL_READY) {
- chip->state = new_state;
- spin_unlock(lock);
- return 0;
+ if (chip->controller->active == chip) {
+ if (chip->state == FL_READY) {
+ chip->state = new_state;
+ spin_unlock(lock);
+ return 0;
+ }
+ if (chip->state == FL_HANDOVER &&
+ chip->controller->handover_waiter == &wait) {
+ chip->state = new_state;
+ spin_unlock(lock);
+ return 0;
+ }
}
if (new_state == FL_PM_SUSPENDED) {
if (chip->controller->active->state == FL_PM_SUSPENDED) {
@@ -856,7 +878,7 @@ nand_get_device(struct mtd_info *mtd, int new_state)
}
}
set_current_state(TASK_UNINTERRUPTIBLE);
- add_wait_queue(wq, &wait);
+ add_wait_queue_exclusive(wq, &wait);
spin_unlock(lock);
schedule();
remove_wait_queue(wq, &wait);
@@ -58,6 +58,7 @@ typedef enum {
FL_OTPING,
FL_PREPARING_ERASE,
FL_VERIFYING_ERASE,
+ FL_HANDOVER,
FL_UNKNOWN
} flstate_t;
@@ -439,6 +439,7 @@ struct nand_hw_control {
spinlock_t lock;
struct nand_chip *active;
wait_queue_head_t wq;
+ wait_queue_t *handover_waiter;
};
/**
I have here a live lock in UBI doing ensure_wear_leveling() -> wear_leveling_worker() -> ubi_eba_copy_leb() MOVE_RETRY -> schedule_erase() -> ensure_wear_leveling() on the same PEB over and over again. The reason for MOVE_RETRY is that the LEB-Lock owner is stucked in nand_get_device() and does not get the device lock. The PEB-lock owner is only scheduled on the CPU while the UBI thread is idle during erase or read while (again) owning the device-lock so the LEB-lock owner makes no progress. To fix this live lock I ensure that there FIFO processing in nand_get_device(). On release the first waiter is marked as the new owner. If someone asks for the device and is not the waiter to which nand device has been handed over then it will put itself on the waitqueue. The FIFO processing was suggested by Peter Zijlstra. As a small optimization I use add_wait_queue_exclusive() instead add_wait_queue() to make sure that only _one_ waiter is woken up and not all of them. Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> --- v1…v2: drop superfluously locking around list_first_entry() drivers/mtd/nand/nand_base.c | 36 +++++++++++++++++++++++++++++------- include/linux/mtd/flashchip.h | 1 + include/linux/mtd/nand.h | 1 + 3 files changed, 31 insertions(+), 7 deletions(-)