Re: [PATCHSET] mempool, percpu, blkcg: fix percpu stat allocation and remove stats_lock

[Vivek Goyal <vgoyal@redhat.com>]

On Tue, Mar 06, 2012 at 01:55:31PM -0800, Andrew Morton wrote:

[..]
> > > hoo boy that looks like an infinite loop.  What's going on here?
> > 
> > If allocation fails, I am trying to allocate it again in infinite loop.
> > What should I do? Try it after sleeping a bit? Or give up after certain
> > number of tries? This is in worker thread context though, so main IO path
> > is not impacted.
> 
> On a non-preemptible unprocessor kernel it's game over, isn't it? 
> Unless someone frees some memory from interrupt context it is time for
> the Big Red Button.

Yes.  Its an issue on non-preemptible UP kernels. I changed the logic to
msleep(10) before retrying. Tested on UP non-preemptible kernel with
always failing allocation and things are fine.

> 
> I'm not sure what to suggest, really - if an allocation failed then
> there's nothing the caller can reliably do to fix that.  The best
> approach is to fail all the way back to userspace with -ENOMEM.

As user space is not waiting for this allocation, -ENOMEM is really
not an option.

> 
> In this context I suppose you could drop a warning into the logs then
> bale out and retry on the next IO attempt.

Yes, that also can be done. I found msleep(10) to be easier solution then
remvoing group from list, and trying again when new IO comes in. Is this
acceptable?
 
[..]
> 
> btw, speaking of uniprocessor: please do perform a uniprocessor build
> and see what impact the patch has upon the size(1) output for the .o
> files.  We should try to minimize the pointless bloat for the UP
> kernel.

But this logic is required both for UP and SMP kernels. So bloat on UP
is not unnecessary?

I ran size(1) on block/blk-cgroup.o with and without the patch and I can
see some bloat.

Without patch(UP kernel)
------------------------
# size block/blk-cgroup.o
   text    data     bss     dec     hex filename
  12950    5248      50   18248    4748 block/blk-cgroup.o

With patch(UP kernel)
------------------------
# size block/blk-cgroup.o
   text    data     bss     dec     hex filename
  13316    5376      58   18750    493e block/blk-cgroup.o

Here is the V3 of the patch.


blk-cgroup: Alloc per cpu stats from worker thread in a delayed manner

Current per cpu stat allocation assumes GFP_KERNEL allocation flag. But in
IO path there are times when we want GFP_NOIO semantics. As there is no
way to pass the allocation flags to alloc_percpu(), this patch delays the
allocation of stats using a worker thread.

v2-> tejun suggested following changes. Changed the patch accordingly.
	- move alloc_node location in structure
	- reduce the size of names of some of the fields
	- Reduce the scope of locking of alloc_list_lock
	- Simplified stat_alloc_fn() by allocating stats for all
	  policies in one go and then assigning these to a group.

v3 -> Andrew suggested to put some comments in the code. Also raised
      concerns about trying to allocate infinitely in case of allocation
      failure. I have changed the logic to sleep for 10ms before retrying.
      That should take care of non-preemptible UP kernels.

Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
---
 block/blk-cgroup.c |  139 +++++++++++++++++++++++++++++++++++++----------------
 block/blk-cgroup.h |    2 
 2 files changed, 101 insertions(+), 40 deletions(-)

Index: tejun-misc/block/blk-cgroup.c
===================================================================
--- tejun-misc.orig/block/blk-cgroup.c	2012-03-07 20:36:44.019949136 -0500
+++ tejun-misc/block/blk-cgroup.c	2012-03-07 20:37:51.188951195 -0500
@@ -30,6 +30,15 @@ static LIST_HEAD(blkio_list);
 static DEFINE_MUTEX(all_q_mutex);
 static LIST_HEAD(all_q_list);
 
+/* List of groups pending per cpu stats allocation */
+static DEFINE_SPINLOCK(alloc_list_lock);
+static LIST_HEAD(alloc_list);
+
+/* Array of per cpu stat pointers allocated for blk groups */
+static void *pcpu_stats[BLKIO_NR_POLICIES];
+static void blkio_stat_alloc_fn(struct work_struct *);
+static DECLARE_WORK(blkio_stat_alloc_work, blkio_stat_alloc_fn);
+
 struct blkio_cgroup blkio_root_cgroup = { .weight = 2*BLKIO_WEIGHT_DEFAULT };
 EXPORT_SYMBOL_GPL(blkio_root_cgroup);
 
@@ -391,6 +400,10 @@ void blkiocg_update_dispatch_stats(struc
 	struct blkio_group_stats_cpu *stats_cpu;
 	unsigned long flags;
 
+	/* If per cpu stats are not allocated yet, don't do any accounting. */
+	if (pd->stats_cpu == NULL)
+		return;
+
 	/*
 	 * Disabling interrupts to provide mutual exclusion between two
 	 * writes on same cpu. It probably is not needed for 64bit. Not
@@ -443,6 +456,10 @@ void blkiocg_update_io_merged_stats(stru
 	struct blkio_group_stats_cpu *stats_cpu;
 	unsigned long flags;
 
+	/* If per cpu stats are not allocated yet, don't do any accounting. */
+	if (pd->stats_cpu == NULL)
+		return;
+
 	/*
 	 * Disabling interrupts to provide mutual exclusion between two
 	 * writes on same cpu. It probably is not needed for 64bit. Not
@@ -460,6 +477,68 @@ void blkiocg_update_io_merged_stats(stru
 }
 EXPORT_SYMBOL_GPL(blkiocg_update_io_merged_stats);
 
+/*
+ * Worker for allocating per cpu stat for blk groups. This is scheduled
+ * once there are some groups on the alloc_list waiting for allocation
+ */
+static void blkio_stat_alloc_fn(struct work_struct *work)
+{
+
+	struct blkio_group *blkg, *n;
+	int i;
+
+alloc_stats:
+	spin_lock_irq(&alloc_list_lock);
+		if (list_empty(&alloc_list)) {
+			/* No more groups needing per cpu stat allocation */
+			spin_unlock_irq(&alloc_list_lock);
+			return;
+		}
+	spin_unlock_irq(&alloc_list_lock);
+
+	for (i = 0; i < BLKIO_NR_POLICIES; i++) {
+		if (pcpu_stats[i] != NULL)
+			continue;
+
+		pcpu_stats[i] = alloc_percpu(struct blkio_group_stats_cpu);
+		/* Allocatoin failed. Try again after some time. */
+		if (pcpu_stats[i] == NULL) {
+			msleep(10);
+			goto alloc_stats;
+		}
+	}
+
+	spin_lock_irq(&blkio_list_lock);
+	spin_lock(&alloc_list_lock);
+
+	list_for_each_entry_safe(blkg, n, &alloc_list, alloc_node) {
+		for (i = 0; i < BLKIO_NR_POLICIES; i++) {
+			struct blkio_policy_type *pol = blkio_policy[i];
+			struct blkg_policy_data *pd;
+
+			if (!pol)
+				continue;
+
+			if (!blkg->pd[i])
+				continue;
+
+			pd = blkg->pd[i];
+			if (pd->stats_cpu)
+				continue;
+
+			pd->stats_cpu = pcpu_stats[i];
+			pcpu_stats[i] = NULL;
+		}
+		list_del_init(&blkg->alloc_node);
+		break;
+	}
+	spin_unlock(&alloc_list_lock);
+	spin_unlock_irq(&blkio_list_lock);
+
+	/* Check if there are more groups needing per cpu stat allocation. */
+	goto alloc_stats;
+}
+
 /**
  * blkg_free - free a blkg
  * @blkg: blkg to free
@@ -491,9 +570,6 @@ static void blkg_free(struct blkio_group
  * @q: request_queue the new blkg is associated with
  *
  * Allocate a new blkg assocating @blkcg and @q.
- *
- * FIXME: Should be called with queue locked but currently isn't due to
- *        percpu stat breakage.
  */
 static struct blkio_group *blkg_alloc(struct blkio_cgroup *blkcg,
 				      struct request_queue *q)
@@ -509,6 +585,7 @@ static struct blkio_group *blkg_alloc(st
 	spin_lock_init(&blkg->stats_lock);
 	blkg->q = q;
 	INIT_LIST_HEAD(&blkg->q_node);
+	INIT_LIST_HEAD(&blkg->alloc_node);
 	blkg->blkcg = blkcg;
 	blkg->refcnt = 1;
 	cgroup_path(blkcg->css.cgroup, blkg->path, sizeof(blkg->path));
@@ -530,13 +607,6 @@ static struct blkio_group *blkg_alloc(st
 
 		blkg->pd[i] = pd;
 		pd->blkg = blkg;
-
-		/* broken, read comment in the callsite */
-		pd->stats_cpu = alloc_percpu(struct blkio_group_stats_cpu);
-		if (!pd->stats_cpu) {
-			blkg_free(blkg);
-			return NULL;
-		}
 	}
 
 	/* invoke per-policy init */
@@ -556,7 +626,7 @@ struct blkio_group *blkg_lookup_create(s
 				       bool for_root)
 	__releases(q->queue_lock) __acquires(q->queue_lock)
 {
-	struct blkio_group *blkg, *new_blkg;
+	struct blkio_group *blkg;
 
 	WARN_ON_ONCE(!rcu_read_lock_held());
 	lockdep_assert_held(q->queue_lock);
@@ -580,48 +650,27 @@ struct blkio_group *blkg_lookup_create(s
 
 	/*
 	 * Allocate and initialize.
-	 *
-	 * FIXME: The following is broken.  Percpu memory allocation
-	 * requires %GFP_KERNEL context and can't be performed from IO
-	 * path.  Allocation here should inherently be atomic and the
-	 * following lock dancing can be removed once the broken percpu
-	 * allocation is fixed.
 	 */
-	spin_unlock_irq(q->queue_lock);
-	rcu_read_unlock();
-
-	new_blkg = blkg_alloc(blkcg, q);
-
-	rcu_read_lock();
-	spin_lock_irq(q->queue_lock);
-
-	/* did bypass get turned on inbetween? */
-	if (unlikely(blk_queue_bypass(q)) && !for_root) {
-		blkg = ERR_PTR(blk_queue_dead(q) ? -EINVAL : -EBUSY);
-		goto out;
-	}
-
-	/* did someone beat us to it? */
-	blkg = blkg_lookup(blkcg, q);
-	if (unlikely(blkg))
-		goto out;
+	blkg = blkg_alloc(blkcg, q);
 
 	/* did alloc fail? */
-	if (unlikely(!new_blkg)) {
+	if (unlikely(!blkg)) {
 		blkg = ERR_PTR(-ENOMEM);
 		goto out;
 	}
 
 	/* insert */
 	spin_lock(&blkcg->lock);
-	swap(blkg, new_blkg);
-
 	hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
 	list_add(&blkg->q_node, &q->blkg_list);
-
 	spin_unlock(&blkcg->lock);
+
+	spin_lock(&alloc_list_lock);
+	list_add(&blkg->alloc_node, &alloc_list);
+	/* Queue per cpu stat allocation from worker thread. */
+	queue_work(system_nrt_wq, &blkio_stat_alloc_work);
+	spin_unlock(&alloc_list_lock);
 out:
-	blkg_free(new_blkg);
 	return blkg;
 }
 EXPORT_SYMBOL_GPL(blkg_lookup_create);
@@ -654,6 +703,10 @@ static void blkg_destroy(struct blkio_gr
 	list_del_init(&blkg->q_node);
 	hlist_del_init_rcu(&blkg->blkcg_node);
 
+	spin_lock(&alloc_list_lock);
+	list_del_init(&blkg->alloc_node);
+	spin_unlock(&alloc_list_lock);
+
 	/*
 	 * Put the reference taken at the time of creation so that when all
 	 * queues are gone, group can be destroyed.
@@ -752,6 +805,9 @@ static void blkio_reset_stats_cpu(struct
 	struct blkg_policy_data *pd = blkg->pd[plid];
 	struct blkio_group_stats_cpu *stats_cpu;
 	int i, j, k;
+
+	if (pd->stats_cpu == NULL)
+		return;
 	/*
 	 * Note: On 64 bit arch this should not be an issue. This has the
 	 * possibility of returning some inconsistent value on 32bit arch
@@ -883,6 +939,9 @@ static uint64_t blkio_read_stat_cpu(stru
 	struct blkio_group_stats_cpu *stats_cpu;
 	u64 val = 0, tval;
 
+	if (pd->stats_cpu == NULL)
+		return val;
+
 	for_each_possible_cpu(cpu) {
 		unsigned int start;
 		stats_cpu = per_cpu_ptr(pd->stats_cpu, cpu);
Index: tejun-misc/block/blk-cgroup.h
===================================================================
--- tejun-misc.orig/block/blk-cgroup.h	2012-03-07 20:36:44.007949131 -0500
+++ tejun-misc/block/blk-cgroup.h	2012-03-07 20:37:51.189951195 -0500
@@ -190,6 +190,8 @@ struct blkio_group {
 	spinlock_t stats_lock;
 	struct blkg_policy_data *pd[BLKIO_NR_POLICIES];
 
+	/* List of blkg waiting for per cpu stats memory to be allocated */
+	struct list_head alloc_node;
 	struct rcu_head rcu_head;
 };