From 774a1221e862b343388347bac9b318767336b20b Mon Sep 17 00:00:00 2001 From: Tejun Heo Date: Tue, 15 Jan 2013 18:52:51 -0800 Subject: module, async: async_synchronize_full() on module init iff async is used If the default iosched is built as module, the kernel may deadlock while trying to load the iosched module on device probe if the probing was running off async. This is because async_synchronize_full() at the end of module init ends up waiting for the async job which initiated the module loading. async A modprobe 1. finds a device 2. registers the block device 3. request_module(default iosched) 4. modprobe in userland 5. load and init module 6. async_synchronize_full() Async A waits for modprobe to finish in request_module() and modprobe waits for async A to finish in async_synchronize_full(). Because there's no easy to track dependency once control goes out to userland, implementing properly nested flushing is difficult. For now, make module init perform async_synchronize_full() iff module init has queued async jobs as suggested by Linus. This avoids the described deadlock because iosched module doesn't use async and thus wouldn't invoke async_synchronize_full(). This is hacky and incomplete. It will deadlock if async module loading nests; however, this works around the known problem case and seems to be the best of bad options. For more details, please refer to the following thread. http://thread.gmane.org/gmane.linux.kernel/1420814 Signed-off-by: Tejun Heo Reported-by: Alex Riesen Tested-by: Ming Lei Tested-by: Alex Riesen Cc: Arjan van de Ven Cc: Jens Axboe Signed-off-by: Linus Torvalds --- kernel/async.c | 3 +++ 1 file changed, 3 insertions(+) (limited to 'kernel/async.c') diff --git a/kernel/async.c b/kernel/async.c index 9d3118384858..a1d585c351d6 100644 --- a/kernel/async.c +++ b/kernel/async.c @@ -196,6 +196,9 @@ static async_cookie_t __async_schedule(async_func_ptr *ptr, void *data, struct a atomic_inc(&entry_count); spin_unlock_irqrestore(&async_lock, flags); + /* mark that this task has queued an async job, used by module init */ + current->flags |= PF_USED_ASYNC; + /* schedule for execution */ queue_work(system_unbound_wq, &entry->work); -- cgit From f56c3196f251012de9b3ebaff55732a9074fdaae Mon Sep 17 00:00:00 2001 From: Tejun Heo Date: Tue, 22 Jan 2013 16:15:15 -0800 Subject: async: fix __lowest_in_progress() Commit 083b804c4d3e ("async: use workqueue for worker pool") made it possible that async jobs are moved from pending to running out-of-order. While pending async jobs will be queued and dispatched for execution in the same order, nothing guarantees they'll enter "1) move self to the running queue" of async_run_entry_fn() in the same order. Before the conversion, async implemented its own worker pool. An async worker, upon being woken up, fetches the first item from the pending list, which kept the executing lists sorted. The conversion to workqueue was done by adding work_struct to each async_entry and async just schedules the work item. The queueing and dispatching of such work items are still in order but now each worker thread is associated with a specific async_entry and moves that specific async_entry to the executing list. So, depending on which worker reaches that point earlier, which is non-deterministic, we may end up moving an async_entry with larger cookie before one with smaller one. This broke __lowest_in_progress(). running->domain may not be properly sorted and is not guaranteed to contain lower cookies than pending list when not empty. Fix it by ensuring sort-inserting to the running list and always looking at both pending and running when trying to determine the lowest cookie. Over time, the async synchronization implementation became quite messy. We better restructure it such that each async_entry is linked to two lists - one global and one per domain - and not move it when execution starts. There's no reason to distinguish pending and running. They behave the same for synchronization purposes. Signed-off-by: Tejun Heo Cc: Arjan van de Ven Cc: stable@vger.kernel.org Signed-off-by: Linus Torvalds --- kernel/async.c | 27 ++++++++++++++++++++------- 1 file changed, 20 insertions(+), 7 deletions(-) (limited to 'kernel/async.c') diff --git a/kernel/async.c b/kernel/async.c index a1d585c351d6..6f34904a0b53 100644 --- a/kernel/async.c +++ b/kernel/async.c @@ -86,18 +86,27 @@ static atomic_t entry_count; */ static async_cookie_t __lowest_in_progress(struct async_domain *running) { + async_cookie_t first_running = next_cookie; /* infinity value */ + async_cookie_t first_pending = next_cookie; /* ditto */ struct async_entry *entry; + /* + * Both running and pending lists are sorted but not disjoint. + * Take the first cookies from both and return the min. + */ if (!list_empty(&running->domain)) { entry = list_first_entry(&running->domain, typeof(*entry), list); - return entry->cookie; + first_running = entry->cookie; } - list_for_each_entry(entry, &async_pending, list) - if (entry->running == running) - return entry->cookie; + list_for_each_entry(entry, &async_pending, list) { + if (entry->running == running) { + first_pending = entry->cookie; + break; + } + } - return next_cookie; /* "infinity" value */ + return min(first_running, first_pending); } static async_cookie_t lowest_in_progress(struct async_domain *running) @@ -118,13 +127,17 @@ static void async_run_entry_fn(struct work_struct *work) { struct async_entry *entry = container_of(work, struct async_entry, work); + struct async_entry *pos; unsigned long flags; ktime_t uninitialized_var(calltime), delta, rettime; struct async_domain *running = entry->running; - /* 1) move self to the running queue */ + /* 1) move self to the running queue, make sure it stays sorted */ spin_lock_irqsave(&async_lock, flags); - list_move_tail(&entry->list, &running->domain); + list_for_each_entry_reverse(pos, &running->domain, list) + if (entry->cookie < pos->cookie) + break; + list_move_tail(&entry->list, &pos->list); spin_unlock_irqrestore(&async_lock, flags); /* 2) run (and print duration) */ -- cgit