writeback: more clean-ups and fixes
Subsystems (3)
cover writeback: more clean-ups and fixes
[Sorry for a duplicating mail, I tried to make this series to be
an In-Reply-To to the previous clean-up series, but failed, so just re-sending with git send-email ]
And on top of those 4 clean-up patches I've implemented the following 3 patches which do further clean-up and fix one of the races I see. There is more work, but I am sending this early to get an early feedback.
Note, I only gave few test to my patches - booted Fedora, made sure bdi writeback threads are created/deleted/doing writeback, hotplugged external USB drive, saw writeback thread created, unplugged and saw it be removed. Did some work while running a kernel with these patches.
The patches are against your for-2.6.36 branch.
My long-term plan is to get rid of unneeded wake-ups, but I'm still just reading your code and learning, and clean-up / fix things I spot.
Artem.
PATCH 1/3 writeback: remove redundant list initialization
From: Artem Bityutskiy <[email protected]>
We do not have to call 'INIT_LIST_HEAD()' for list elements ('bdi->bdi_list') before inserting them to the 'bdi_pending_list', because 'list_add_tail' will initialize the element anyway. Thus, kill the redundant initialization.
--- mm/backing-dev.c | 1 - 1 files changed, 0 insertions(+), 1 deletions(-)
mm/backing-dev.c +0 -1
@@ -416,7 +416,6 @@ static void bdi_add_to_pending(struct rcu_head *head)struct backing_dev_info *bdi; bdi = container_of(head, struct backing_dev_info, rcu_head); -INIT_LIST_HEAD (&bdi->bdi_list);spin_lock (&bdi_lock);list_add_tail (&bdi->bdi_list, &bdi_pending_list);
Signature · 1 line
1.7.1.1
PATCH 2/3 writeback: simplify bdi code a little
In reply to Artem BityutskiyFrom: Artem Bityutskiy <[email protected]>
This patch simplifies bdi code a little by removing the 'pending_list' which is redundant. Indeed, currently the forker thread ('bdi_forker_thread()') is working like this:
1. In a loop, fetch all bdi's which have works but have no writeback thread and move them to the 'pending_list'.
2. If the list is empty, sleep for 5 s.
3. Otherwise, take one BDI from the list, fork the writeback thread for this bdi, and repeat the loop again.
IOW, it first moves everything to the 'pending_list', then process only one element, and so on. With this patch the algorithm is:
1. Find the first bdi which has work and remove it from the global list of bdi's (bdi_list).
2. If there were not such bdi, sleep 5 s.
3. Fork the writeback thread for this bdi and repeat the loop again.
IOW, now we find the first bdi to process, process it, and so on. This is simpler and involves less lists.
This patch also removes unneeded and unused complications which involve 'rcu_call()' and 'bdi->rcu_head' - now we use simple 'synchronize_rcu()' and remove the 'rcu_head' field from 'struct backing_dev_info'.
--- include/linux/backing-dev.h | 1 - mm/backing-dev.c | 85 ++++++++++++------------------------------- 2 files changed, 23 insertions(+), 63 deletions(-)
include/linux/backing-dev.h +0 -1
@@ -58,7 +58,6 @@ struct bdi_writeback {struct backing_dev_info {struct list_head bdi_list; -struct rcu_head rcu_head;unsigned long ra_pages;/* max readahead in PAGE_CACHE_SIZE units */ unsigned long state;/* Always use atomic bitops on this */ unsigned int capabilities;/* Device capabilities */
mm/backing-dev.c +23 -62
138 lines not on this page · raw message
Signature · 1 line
1.7.1.1
PATCH 3/3 writeback: fix possible race when shutting down bdi
In reply to Artem BityutskiyFrom: Artem Bityutskiy <[email protected]>
Current bdi code has the following race between 'bdi_wb_shutdown()' and 'bdi_forker_thread()':
Initial condition for the bdi we consider: BDI_pending is cleaned, bdi has no writeback thread, because it was inactive and exited, 'bdi_wb_shutdown()' and 'bdi_forker_thread()' are executed concurrently.
1. bdi_wb_shutdown() executes wait_on_bit(), tests the BDI_pending bit, it is clean, so it does not wait for anything.
2. 'bdi_forker_thread()' takes the 'bdi_lock', finds out that bdi has work to do, takes it out of the 'bdi_list', sets the BDI_pending flag, unlocks the 'bdi_lock' lock
3. 'bdi_wb_shutdown()' takes the lock, and nasty things start happening:
a) it removes the bdi from bdi->bdi_list, but the bdi is not in any list
b) it starts deleting the bdi, but 'bdi_forker_thread()' is still working with it.
Note, it is very difficult to hit this race, and I never observed it, so it is quite theoretical, but it is still a race. Also note, this race exist without my previous clean-ups as well.
This patch fixes this race by making 'bdi_wb_shutdown()' first search for the bdi in the 'bdi_list', and only if it is there, remove it from 'bdi_list' and destroy. But if it is not there, assume it is in transit and re-try waiting on the BDI_pending bit.
--- mm/backing-dev.c | 41 ++++++++++++++++++++++++++++------------- 1 files changed, 28 insertions(+), 13 deletions(-)
mm/backing-dev.c +28 -13
@@ -418,15 +418,26 @@ static int bdi_forker_thread(void *ptr) }/* -* Remove bdi from bdi_list, and ensure that it is no longer visible +* Look up for bdi in the bdi_list. If found, remove it, ensure that it is +* no longer visible, and return 0. If not found, return 1. */ -static void bdi_remove_from_list (struct backing_dev_info *bdi) +static int bdi_remove_from_list (struct backing_dev_info *me) { +struct backing_dev_info *bdi; +spin_lock_bh (&bdi_lock); -list_del_rcu (&bdi->bdi_list); +list_for_each_entry (bdi, &bdi_list, bdi_list) { +if (bdi == me) { +list_del_rcu (&me->bdi_list); +spin_unlock_bh (&bdi_lock); +synchronize_rcu (); +return 0 ; + } + + }spin_unlock_bh (&bdi_lock); +return 1 ; -synchronize_rcu (); } int bdi_register(struct backing_dev_info *bdi, struct device *parent,
@@ -494,16 +505,20 @@ static void bdi_wb_shutdown(struct backing_dev_info *bdi)if (!bdi_cap_writeback_dirty (bdi))return ; -/* -* If setup is pending, wait for that to complete first -*/ -wait_on_bit (&bdi->state, BDI_pending, bdi_sched_wait, -TASK_UNINTERRUPTIBLE ); +do { +/* +* If setup is pending, wait for that to complete first +*/ +wait_on_bit (&bdi->state, BDI_pending, bdi_sched_wait, +TASK_UNINTERRUPTIBLE ); -/* -* Make sure nobody finds us on the bdi_list anymore -*/ -bdi_remove_from_list (bdi); +/* +* Make sure nobody finds us on the bdi_list anymore. However, +* bdi may be temporary be not in the bdi_list but be in transit +* in bdi_forker_thread. Namely, this may happen if we race +* with the forker thread. +*/ + }while (bdi_remove_from_list (bdi));/* * Finally, kill the kernel thread. We don't need to be RCU
Signature · 1 line
1.7.1.1