Merge branch 'for-3.18' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu

Pull percpu updates from Tejun Heo:
 "A lot of activities on percpu front.  Notable changes are...

   - percpu allocator now can take @gfp.  If @gfp doesn't contain
     GFP_KERNEL, it tries to allocate from what's already available to
     the allocator and a work item tries to keep the reserve around
     certain level so that these atomic allocations usually succeed.

     This will replace the ad-hoc percpu memory pool used by
     blk-throttle and also be used by the planned blkcg support for
     writeback IOs.

     Please note that I noticed a bug in how @gfp is interpreted while
     preparing this pull request and applied the fix 6ae833c7fe0c
     ("percpu: fix how @gfp is interpreted by the percpu allocator")
     just now.

   - percpu_ref now uses longs for percpu and global counters instead of
     ints.  It leads to more sparse packing of the percpu counters on
     64bit machines but the overhead should be negligible and this
     allows using percpu_ref for refcnting pages and in-memory objects
     directly.

   - The switching between percpu and single counter modes of a
     percpu_ref is made independent of putting the base ref and a
     percpu_ref can now optionally be initialized in single or killed
     mode.  This allows avoiding percpu shutdown latency for cases where
     the refcounted objects may be synchronously created and destroyed
     in rapid succession with only a fraction of them reaching fully
     operational status (SCSI probing does this when combined with
     blk-mq support).  It's also planned to be used to implement forced
     single mode to detect underflow more timely for debugging.

  There's a separate branch percpu/for-3.18-consistent-ops which cleans
  up the duplicate percpu accessors.  That branch causes a number of
  conflicts with s390 and other trees.  I'll send a separate pull
  request w/ resolutions once other branches are merged"

* 'for-3.18' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu: (33 commits)
  percpu: fix how @gfp is interpreted by the percpu allocator
  blk-mq, percpu_ref: start q->mq_usage_counter in atomic mode
  percpu_ref: make INIT_ATOMIC and switch_to_atomic() sticky
  percpu_ref: add PERCPU_REF_INIT_* flags
  percpu_ref: decouple switching to percpu mode and reinit
  percpu_ref: decouple switching to atomic mode and killing
  percpu_ref: add PCPU_REF_DEAD
  percpu_ref: rename things to prepare for decoupling percpu/atomic mode switch
  percpu_ref: replace pcpu_ prefix with percpu_
  percpu_ref: minor code and comment updates
  percpu_ref: relocate percpu_ref_reinit()
  Revert "blk-mq, percpu_ref: implement a kludge for SCSI blk-mq stall during probe"
  Revert "percpu: free percpu allocation info for uniprocessor system"
  percpu-refcount: make percpu_ref based on longs instead of ints
  percpu-refcount: improve WARN messages
  percpu: fix locking regression in the failure path of pcpu_alloc()
  percpu-refcount: add @gfp to percpu_ref_init()
  proportions: add @gfp to init functions
  percpu_counter: add @gfp to percpu_counter_init()
  percpu_counter: make percpu_counters_lock irq-safe
  ...

39 files changed, 879 insertions, 444 deletions

diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
index 3201e93ebd07..ac1c4de3a484 100644
--- a/arch/x86/kvm/mmu.c
+++ b/arch/x86/kvm/mmu.c
@@ -4549,7 +4549,7 @@ int kvm_mmu_module_init(void)
 	if (!mmu_page_header_cache)
 		goto nomem;
 
-	if (percpu_counter_init(&kvm_total_used_mmu_pages, 0))
+	if (percpu_counter_init(&kvm_total_used_mmu_pages, 0, GFP_KERNEL))
 		goto nomem;
 
 	register_shrinker(&mmu_shrinker);
diff --git a/block/blk-mq-sysfs.c b/block/blk-mq-sysfs.c
index ed5217867555..371d8800b48a 100644
--- a/block/blk-mq-sysfs.c
+++ b/block/blk-mq-sysfs.c
@@ -402,6 +402,12 @@ static void blk_mq_sysfs_init(struct request_queue *q)
 	}
 }
 
+/* see blk_register_queue() */
+void blk_mq_finish_init(struct request_queue *q)
+{
+	percpu_ref_switch_to_percpu(&q->mq_usage_counter);
+}
+
 int blk_mq_register_disk(struct gendisk *disk)
 {
 	struct device *dev = disk_to_dev(disk);
diff --git a/block/blk-mq.c b/block/blk-mq.c
index df8e1e09dd17..38f4a165640d 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -119,16 +119,7 @@ void blk_mq_freeze_queue(struct request_queue *q)
 	spin_unlock_irq(q->queue_lock);
 
 	if (freeze) {
-		/*
-		 * XXX: Temporary kludge to work around SCSI blk-mq stall.
-		 * SCSI synchronously creates and destroys many queues
-		 * back-to-back during probe leading to lengthy stalls.
-		 * This will be fixed by keeping ->mq_usage_counter in
-		 * atomic mode until genhd registration, but, for now,
-		 * let's work around using expedited synchronization.
-		 */
-		__percpu_ref_kill_expedited(&q->mq_usage_counter);
-
+		percpu_ref_kill(&q->mq_usage_counter);
 		blk_mq_run_queues(q, false);
 	}
 	wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
@@ -1804,7 +1795,12 @@ struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
 	if (!q)
 		goto err_hctxs;
 
-	if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release))
+	/*
+	 * Init percpu_ref in atomic mode so that it's faster to shutdown.
+	 * See blk_register_queue() for details.
+	 */
+	if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
+			    PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
 		goto err_map;
 
 	setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 17f5c84ce7bf..521ae9089c50 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -551,12 +551,19 @@ int blk_register_queue(struct gendisk *disk)
 		return -ENXIO;
 
 	/*
-	 * Initialization must be complete by now.  Finish the initial
-	 * bypass from queue allocation.
+	 * SCSI probing may synchronously create and destroy a lot of
+	 * request_queues for non-existent devices.  Shutting down a fully
+	 * functional queue takes measureable wallclock time as RCU grace
+	 * periods are involved.  To avoid excessive latency in these
+	 * cases, a request_queue starts out in a degraded mode which is
+	 * faster to shut down and is made fully functional here as
+	 * request_queues for non-existent devices never get registered.
 	 */
 	if (!blk_queue_init_done(q)) {
 		queue_flag_set_unlocked(QUEUE_FLAG_INIT_DONE, q);
 		blk_queue_bypass_end(q);
+		if (q->mq_ops)
+			blk_mq_finish_init(q);
 	}
 
 	ret = blk_trace_init_sysfs(dev);
diff --git a/drivers/target/target_core_tpg.c b/drivers/target/target_core_tpg.c
index fddfae61222f..be783f717f19 100644
--- a/drivers/target/target_core_tpg.c
+++ b/drivers/target/target_core_tpg.c
@@ -819,7 +819,8 @@ int core_tpg_add_lun(
 {
 	int ret;
 
-	ret = percpu_ref_init(&lun->lun_ref, core_tpg_lun_ref_release);
+	ret = percpu_ref_init(&lun->lun_ref, core_tpg_lun_ref_release, 0,
+			      GFP_KERNEL);
 	if (ret < 0)
 		return ret;
 
diff --git a/fs/aio.c b/fs/aio.c
index 733750096b71..84a751005f5b 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -661,10 +661,10 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 
 	INIT_LIST_HEAD(&ctx->active_reqs);
 
-	if (percpu_ref_init(&ctx->users, free_ioctx_users))
+	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
 		goto err;
 
-	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
+	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
 		goto err;
 
 	ctx->cpu = alloc_percpu(struct kioctx_cpu);
diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index a1d36e62179c..d0d78dc07792 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -1183,7 +1183,7 @@ static struct btrfs_subvolume_writers *btrfs_alloc_subvolume_writers(void)
 	if (!writers)
 		return ERR_PTR(-ENOMEM);
 
-	ret = percpu_counter_init(&writers->counter, 0);
+	ret = percpu_counter_init(&writers->counter, 0, GFP_KERNEL);
 	if (ret < 0) {
 		kfree(writers);
 		return ERR_PTR(ret);
@@ -2188,7 +2188,7 @@ int open_ctree(struct super_block *sb,
 		goto fail_srcu;
 	}
 
-	ret = percpu_counter_init(&fs_info->dirty_metadata_bytes, 0);
+	ret = percpu_counter_init(&fs_info->dirty_metadata_bytes, 0, GFP_KERNEL);
 	if (ret) {
 		err = ret;
 		goto fail_bdi;
@@ -2196,13 +2196,13 @@ int open_ctree(struct super_block *sb,
 	fs_info->dirty_metadata_batch = PAGE_CACHE_SIZE *
 					(1 + ilog2(nr_cpu_ids));
 
-	ret = percpu_counter_init(&fs_info->delalloc_bytes, 0);
+	ret = percpu_counter_init(&fs_info->delalloc_bytes, 0, GFP_KERNEL);
 	if (ret) {
 		err = ret;
 		goto fail_dirty_metadata_bytes;
 	}
 
-	ret = percpu_counter_init(&fs_info->bio_counter, 0);
+	ret = percpu_counter_init(&fs_info->bio_counter, 0, GFP_KERNEL);
 	if (ret) {
 		err = ret;
 		goto fail_delalloc_bytes;
diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
index 3efe1c3877bf..caaf015d6e4b 100644
--- a/fs/btrfs/extent-tree.c
+++ b/fs/btrfs/extent-tree.c
@@ -3494,7 +3494,7 @@ static int update_space_info(struct btrfs_fs_info *info, u64 flags,
 	if (!found)
 		return -ENOMEM;
 
-	ret = percpu_counter_init(&found->total_bytes_pinned, 0);
+	ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
 	if (ret) {
 		kfree(found);
 		return ret;
diff --git a/fs/ext2/super.c b/fs/ext2/super.c
index b88edc05c230..170dc41e8bf4 100644
--- a/fs/ext2/super.c
+++ b/fs/ext2/super.c
@@ -1067,14 +1067,14 @@ static int ext2_fill_super(struct super_block *sb, void *data, int silent)
 	ext2_rsv_window_add(sb, &sbi->s_rsv_window_head);
 
 	err = percpu_counter_init(&sbi->s_freeblocks_counter,
-				ext2_count_free_blocks(sb));
+				ext2_count_free_blocks(sb), GFP_KERNEL);
 	if (!err) {
 		err = percpu_counter_init(&sbi->s_freeinodes_counter,
-				ext2_count_free_inodes(sb));
+				ext2_count_free_inodes(sb), GFP_KERNEL);
 	}
 	if (!err) {
 		err = percpu_counter_init(&sbi->s_dirs_counter,
-				ext2_count_dirs(sb));
+				ext2_count_dirs(sb), GFP_KERNEL);
 	}
 	if (err) {
 		ext2_msg(sb, KERN_ERR, "error: insufficient memory");
diff --git a/fs/ext3/super.c b/fs/ext3/super.c
index 622e88249024..bb0fdacad058 100644
--- a/fs/ext3/super.c
+++ b/fs/ext3/super.c
@@ -2039,14 +2039,14 @@ static int ext3_fill_super (struct super_block *sb, void *data, int silent)
 		goto failed_mount2;
 	}
 	err = percpu_counter_init(&sbi->s_freeblocks_counter,
-			ext3_count_free_blocks(sb));
+			ext3_count_free_blocks(sb), GFP_KERNEL);
 	if (!err) {
 		err = percpu_counter_init(&sbi->s_freeinodes_counter,
-				ext3_count_free_inodes(sb));
+				ext3_count_free_inodes(sb), GFP_KERNEL);
 	}
 	if (!err) {
 		err = percpu_counter_init(&sbi->s_dirs_counter,
-				ext3_count_dirs(sb));
+				ext3_count_dirs(sb), GFP_KERNEL);
 	}
 	if (err) {
 		ext3_msg(sb, KERN_ERR, "error: insufficient memory");
diff --git a/fs/ext4/super.c b/fs/ext4/super.c
index 0b28b36e7915..05c159218bc2 100644
--- a/fs/ext4/super.c
+++ b/fs/ext4/super.c
@@ -3892,7 +3892,8 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent)
 	/* Register extent status tree shrinker */
 	ext4_es_register_shrinker(sbi);
 
-	if ((err = percpu_counter_init(&sbi->s_extent_cache_cnt, 0)) != 0) {
+	err = percpu_counter_init(&sbi->s_extent_cache_cnt, 0, GFP_KERNEL);
+	if (err) {
 		ext4_msg(sb, KERN_ERR, "insufficient memory");
 		goto failed_mount3;
 	}
@@ -4106,17 +4107,20 @@ no_journal:
 	block = ext4_count_free_clusters(sb);
 	ext4_free_blocks_count_set(sbi->s_es, 
 				   EXT4_C2B(sbi, block));
-	err = percpu_counter_init(&sbi->s_freeclusters_counter, block);
+	err = percpu_counter_init(&sbi->s_freeclusters_counter, block,
+				  GFP_KERNEL);
 	if (!err) {
 		unsigned long freei = ext4_count_free_inodes(sb);
 		sbi->s_es->s_free_inodes_count = cpu_to_le32(freei);
-		err = percpu_counter_init(&sbi->s_freeinodes_counter, freei);
+		err = percpu_counter_init(&sbi->s_freeinodes_counter, freei,
+					  GFP_KERNEL);
 	}
 	if (!err)
 		err = percpu_counter_init(&sbi->s_dirs_counter,
-					  ext4_count_dirs(sb));
+					  ext4_count_dirs(sb), GFP_KERNEL);
 	if (!err)
-		err = percpu_counter_init(&sbi->s_dirtyclusters_counter, 0);
+		err = percpu_counter_init(&sbi->s_dirtyclusters_counter, 0,
+					  GFP_KERNEL);
 	if (err) {
 		ext4_msg(sb, KERN_ERR, "insufficient memory");
 		goto failed_mount6;
diff --git a/fs/file_table.c b/fs/file_table.c
index 385bfd31512a..0bab12b20460 100644
--- a/fs/file_table.c
+++ b/fs/file_table.c
@@ -331,5 +331,5 @@ void __init files_init(unsigned long mempages)
 
 	n = (mempages * (PAGE_SIZE / 1024)) / 10;
 	files_stat.max_files = max_t(unsigned long, n, NR_FILE);
-	percpu_counter_init(&nr_files, 0);
+	percpu_counter_init(&nr_files, 0, GFP_KERNEL);
 } 
diff --git a/fs/quota/dquot.c b/fs/quota/dquot.c
index f2d0eee9d1f1..8b663b2d9562 100644
--- a/fs/quota/dquot.c
+++ b/fs/quota/dquot.c
@@ -2725,7 +2725,7 @@ static int __init dquot_init(void)
 		panic("Cannot create dquot hash table");
 
 	for (i = 0; i < _DQST_DQSTAT_LAST; i++) {
-		ret = percpu_counter_init(&dqstats.counter[i], 0);
+		ret = percpu_counter_init(&dqstats.counter[i], 0, GFP_KERNEL);
 		if (ret)
 			panic("Cannot create dquot stat counters");
 	}
diff --git a/fs/super.c b/fs/super.c
index b9a214d2fe98..1b836107acee 100644
--- a/fs/super.c
+++ b/fs/super.c
@@ -175,7 +175,8 @@ static struct super_block *alloc_super(struct file_system_type *type, int flags)
 		goto fail;
 
 	for (i = 0; i < SB_FREEZE_LEVELS; i++) {
-		if (percpu_counter_init(&s->s_writers.counter[i], 0) < 0)
+		if (percpu_counter_init(&s->s_writers.counter[i], 0,
+					GFP_KERNEL) < 0)
 			goto fail;
 		lockdep_init_map(&s->s_writers.lock_map[i], sb_writers_name[i],
 				 &type->s_writers_key[i], 0);
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index a1e31f274fcd..c13a0c09faea 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -140,6 +140,7 @@ enum {
 };
 
 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *);
+void blk_mq_finish_init(struct request_queue *q);
 int blk_mq_register_disk(struct gendisk *);
 void blk_mq_unregister_disk(struct gendisk *);
 
diff --git a/include/linux/flex_proportions.h b/include/linux/flex_proportions.h
index 4ebc49fae391..0d348e011a6e 100644
--- a/include/linux/flex_proportions.h
+++ b/include/linux/flex_proportions.h
@@ -10,6 +10,7 @@
 #include <linux/percpu_counter.h>
 #include <linux/spinlock.h>
 #include <linux/seqlock.h>
+#include <linux/gfp.h>
 
 /*
  * When maximum proportion of some event type is specified, this is the
@@ -32,7 +33,7 @@ struct fprop_global {
 	seqcount_t sequence;
 };
 
-int fprop_global_init(struct fprop_global *p);
+int fprop_global_init(struct fprop_global *p, gfp_t gfp);
 void fprop_global_destroy(struct fprop_global *p);
 bool fprop_new_period(struct fprop_global *p, int periods);
 
@@ -79,7 +80,7 @@ struct fprop_local_percpu {
 	raw_spinlock_t lock;	/* Protect period and numerator */
 };
 
-int fprop_local_init_percpu(struct fprop_local_percpu *pl);
+int fprop_local_init_percpu(struct fprop_local_percpu *pl, gfp_t gfp);
 void fprop_local_destroy_percpu(struct fprop_local_percpu *pl);
 void __fprop_inc_percpu(struct fprop_global *p, struct fprop_local_percpu *pl);
 void __fprop_inc_percpu_max(struct fprop_global *p, struct fprop_local_percpu *pl,
diff --git a/include/linux/percpu-refcount.h b/include/linux/percpu-refcount.h
index 68a64f11ce02..d5c89e0dd0e6 100644
--- a/include/linux/percpu-refcount.h
+++ b/include/linux/percpu-refcount.h
@@ -13,7 +13,7 @@
  *
  * The refcount will have a range of 0 to ((1U << 31) - 1), i.e. one bit less
  * than an atomic_t - this is because of the way shutdown works, see
- * percpu_ref_kill()/PCPU_COUNT_BIAS.
+ * percpu_ref_kill()/PERCPU_COUNT_BIAS.
  *
  * Before you call percpu_ref_kill(), percpu_ref_put() does not check for the
  * refcount hitting 0 - it can't, if it was in percpu mode. percpu_ref_kill()
@@ -49,29 +49,60 @@
 #include <linux/kernel.h>
 #include <linux/percpu.h>
 #include <linux/rcupdate.h>
+#include <linux/gfp.h>
 
 struct percpu_ref;
 typedef void (percpu_ref_func_t)(struct percpu_ref *);
 
+/* flags set in the lower bits of percpu_ref->percpu_count_ptr */
+enum {
+	__PERCPU_REF_ATOMIC	= 1LU << 0,	/* operating in atomic mode */
+	__PERCPU_REF_DEAD	= 1LU << 1,	/* (being) killed */
+	__PERCPU_REF_ATOMIC_DEAD = __PERCPU_REF_ATOMIC | __PERCPU_REF_DEAD,
+
+	__PERCPU_REF_FLAG_BITS	= 2,
+};
+
+/* @flags for percpu_ref_init() */
+enum {
+	/*
+	 * Start w/ ref == 1 in atomic mode.  Can be switched to percpu
+	 * operation using percpu_ref_switch_to_percpu().  If initialized
+	 * with this flag, the ref will stay in atomic mode until
+	 * percpu_ref_switch_to_percpu() is invoked on it.
+	 */
+	PERCPU_REF_INIT_ATOMIC	= 1 << 0,
+
+	/*
+	 * Start dead w/ ref == 0 in atomic mode.  Must be revived with
+	 * percpu_ref_reinit() before used.  Implies INIT_ATOMIC.
+	 */
+	PERCPU_REF_INIT_DEAD	= 1 << 1,
+};
+
 struct percpu_ref {
-	atomic_t		count;
+	atomic_long_t		count;
 	/*
 	 * The low bit of the pointer indicates whether the ref is in percpu
 	 * mode; if set, then get/put will manipulate the atomic_t.
 	 */
-	unsigned long		pcpu_count_ptr;
+	unsigned long		percpu_count_ptr;
 	percpu_ref_func_t	*release;
-	percpu_ref_func_t	*confirm_kill;
+	percpu_ref_func_t	*confirm_switch;
+	bool			force_atomic:1;
 	struct rcu_head		rcu;
 };
 
 int __must_check percpu_ref_init(struct percpu_ref *ref,
-				 percpu_ref_func_t *release);
-void percpu_ref_reinit(struct percpu_ref *ref);
+				 percpu_ref_func_t *release, unsigned int flags,
+				 gfp_t gfp);
 void percpu_ref_exit(struct percpu_ref *ref);
+void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
+				 percpu_ref_func_t *confirm_switch);
+void percpu_ref_switch_to_percpu(struct percpu_ref *ref);
 void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
 				 percpu_ref_func_t *confirm_kill);
-void __percpu_ref_kill_expedited(struct percpu_ref *ref);
+void percpu_ref_reinit(struct percpu_ref *ref);
 
 /**
  * percpu_ref_kill - drop the initial ref
@@ -88,26 +119,24 @@ static inline void percpu_ref_kill(struct percpu_ref *ref)
 	return percpu_ref_kill_and_confirm(ref, NULL);
 }
 
-#define PCPU_REF_DEAD		1
-
 /*
  * Internal helper.  Don't use outside percpu-refcount proper.  The
  * function doesn't return the pointer and let the caller test it for NULL
  * because doing so forces the compiler to generate two conditional
- * branches as it can't assume that @ref->pcpu_count is not NULL.
+ * branches as it can't assume that @ref->percpu_count is not NULL.
  */
-static inline bool __pcpu_ref_alive(struct percpu_ref *ref,
-				    unsigned __percpu **pcpu_countp)
+static inline bool __ref_is_percpu(struct percpu_ref *ref,
+					  unsigned long __percpu **percpu_countp)
 {
-	unsigned long pcpu_ptr = ACCESS_ONCE(ref->pcpu_count_ptr);
+	unsigned long percpu_ptr = ACCESS_ONCE(ref->percpu_count_ptr);
 
 	/* paired with smp_store_release() in percpu_ref_reinit() */
 	smp_read_barrier_depends();
 
-	if (unlikely(pcpu_ptr & PCPU_REF_DEAD))
+	if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC))
 		return false;
 
-	*pcpu_countp = (unsigned __percpu *)pcpu_ptr;
+	*percpu_countp = (unsigned long __percpu *)percpu_ptr;
 	return true;
 }
 
@@ -115,18 +144,20 @@ static inline bool __pcpu_ref_alive(struct percpu_ref *ref,
  * percpu_ref_get - increment a percpu refcount
  * @ref: percpu_ref to get
  *
- * Analagous to atomic_inc().
-  */
+ * Analagous to atomic_long_inc().
+ *
+ * This function is safe to call as long as @ref is between init and exit.
+ */
 static inline void percpu_ref_get(struct percpu_ref *ref)
 {
-	unsigned __percpu *pcpu_count;
+	unsigned long __percpu *percpu_count;
 
 	rcu_read_lock_sched();
 
-	if (__pcpu_ref_alive(ref, &pcpu_count))
-		this_cpu_inc(*pcpu_count);
+	if (__ref_is_percpu(ref, &percpu_count))
+		this_cpu_inc(*percpu_count);
 	else
-		atomic_inc(&ref->count);
+		atomic_long_inc(&ref->count);
 
 	rcu_read_unlock_sched();
 }
@@ -138,20 +169,20 @@ static inline void percpu_ref_get(struct percpu_ref *ref)
  * Increment a percpu refcount unless its count already reached zero.
  * Returns %true on success; %false on failure.
  *
- * The caller is responsible for ensuring that @ref stays accessible.
+ * This function is safe to call as long as @ref is between init and exit.
  */
 static inline bool percpu_ref_tryget(struct percpu_ref *ref)
 {
-	unsigned __percpu *pcpu_count;
-	int ret = false;
+	unsigned long __percpu *percpu_count;
+	int ret;
 
 	rcu_read_lock_sched();
 
-	if (__pcpu_ref_alive(ref, &pcpu_count)) {
-		this_cpu_inc(*pcpu_count);
+	if (__ref_is_percpu(ref, &percpu_count)) {
+		this_cpu_inc(*percpu_count);
 		ret = true;
 	} else {
-		ret = atomic_inc_not_zero(&ref->count);
+		ret = atomic_long_inc_not_zero(&ref->count);
 	}
 
 	rcu_read_unlock_sched();
@@ -166,23 +197,26 @@ static inline bool percpu_ref_tryget(struct percpu_ref *ref)
  * Increment a percpu refcount unless it has already been killed.  Returns
  * %true on success; %false on failure.
  *
- * Completion of percpu_ref_kill() in itself doesn't guarantee that tryget
- * will fail.  For such guarantee, percpu_ref_kill_and_confirm() should be
- * used.  After the confirm_kill callback is invoked, it's guaranteed that
- * no new reference will be given out by percpu_ref_tryget().
+ * Completion of percpu_ref_kill() in itself doesn't guarantee that this
+ * function will fail.  For such guarantee, percpu_ref_kill_and_confirm()
+ * should be used.  After the confirm_kill callback is invoked, it's
+ * guaranteed that no new reference will be given out by
+ * percpu_ref_tryget_live().
  *
- * The caller is responsible for ensuring that @ref stays accessible.
+ * This function is safe to call as long as @ref is between init and exit.
  */
 static inline bool percpu_ref_tryget_live(struct percpu_ref *ref)
 {
-	unsigned __percpu *pcpu_count;
+	unsigned long __percpu *percpu_count;
 	int ret = false;
 
 	rcu_read_lock_sched();
 
-	if (__pcpu_ref_alive(ref, &pcpu_count)) {
-		this_cpu_inc(*pcpu_count);
+	if (__ref_is_percpu(ref, &percpu_count)) {
+		this_cpu_inc(*percpu_count);
 		ret = true;
+	} else if (!(ACCESS_ONCE(ref->percpu_count_ptr) & __PERCPU_REF_DEAD)) {
+		ret = atomic_long_inc_not_zero(&ref->count);
 	}
 
 	rcu_read_unlock_sched();
@@ -196,16 +230,18 @@ static inline bool percpu_ref_tryget_live(struct percpu_ref *ref)
  *
  * Decrement the refcount, and if 0, call the release function (which was passed
  * to percpu_ref_init())
+ *
+ * This function is safe to call as long as @ref is between init and exit.
  */
 static inline void percpu_ref_put(struct percpu_ref *ref)
 {
-	unsigned __percpu *pcpu_count;
+	unsigned long __percpu *percpu_count;
 
 	rcu_read_lock_sched();
 
-	if (__pcpu_ref_alive(ref, &pcpu_count))
-		this_cpu_dec(*pcpu_count);
-	else if (unlikely(atomic_dec_and_test(&ref->count)))
+	if (__ref_is_percpu(ref, &percpu_count))
+		this_cpu_dec(*percpu_count);
+	else if (unlikely(atomic_long_dec_and_test(&ref->count)))
 		ref->release(ref);
 
 	rcu_read_unlock_sched();
@@ -216,14 +252,16 @@ static inline void percpu_ref_put(struct percpu_ref *ref)
  * @ref: percpu_ref to test
  *
  * Returns %true if @ref reached zero.
+ *
+ * This function is safe to call as long as @ref is between init and exit.
  */
 static inline bool percpu_ref_is_zero(struct percpu_ref *ref)
 {
-	unsigned __percpu *pcpu_count;
+	unsigned long __percpu *percpu_count;
 
-	if (__pcpu_ref_alive(ref, &pcpu_count))
+	if (__ref_is_percpu(ref, &percpu_count))
 		return false;
-	return !atomic_read(&ref->count);
+	return !atomic_long_read(&ref->count);
 }
 
 #endif
diff --git a/include/linux/percpu.h b/include/linux/percpu.h
index 6f61b61b7996..a3aa63e47637 100644
--- a/include/linux/percpu.h
+++ b/include/linux/percpu.h
@@ -48,9 +48,9 @@
  * intelligent way to determine this would be nice.
  */
 #if BITS_PER_LONG > 32
-#define PERCPU_DYNAMIC_RESERVE		(20 << 10)
+#define PERCPU_DYNAMIC_RESERVE		(28 << 10)
 #else
-#define PERCPU_DYNAMIC_RESERVE		(12 << 10)
+#define PERCPU_DYNAMIC_RESERVE		(20 << 10)
 #endif
 
 extern void *pcpu_base_addr;
@@ -122,11 +122,16 @@ extern void __init setup_per_cpu_areas(void);
 #endif
 extern void __init percpu_init_late(void);
 
+extern void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp);
 extern void __percpu *__alloc_percpu(size_t size, size_t align);
 extern void free_percpu(void __percpu *__pdata);
 extern phys_addr_t per_cpu_ptr_to_phys(void *addr);
 
-#define alloc_percpu(type)	\
-	(typeof(type) __percpu *)__alloc_percpu(sizeof(type), __alignof__(type))
+#define alloc_percpu_gfp(type, gfp)					\
+	(typeof(type) __percpu *)__alloc_percpu_gfp(sizeof(type),	\
+						__alignof__(type), gfp)
+#define alloc_percpu(type)						\
+	(typeof(type) __percpu *)__alloc_percpu(sizeof(type),		\
+						__alignof__(type))
 
 #endif /* __LINUX_PERCPU_H */
diff --git a/include/linux/percpu_counter.h b/include/linux/percpu_counter.h
index d5dd4657c8d6..50e50095c8d1 100644
--- a/include/linux/percpu_counter.h
+++ b/include/linux/percpu_counter.h
@@ -12,6 +12,7 @@
 #include <linux/threads.h>
 #include <linux/percpu.h>
 #include <linux/types.h>
+#include <linux/gfp.h>
 
 #ifdef CONFIG_SMP
 
@@ -26,14 +27,14 @@ struct percpu_counter {
 
 extern int percpu_counter_batch;
 
-int __percpu_counter_init(struct percpu_counter *fbc, s64 amount,
+int __percpu_counter_init(struct percpu_counter *fbc, s64 amount, gfp_t gfp,
 			  struct lock_class_key *key);
 
-#define percpu_counter_init(fbc, value)					\
+#define percpu_counter_init(fbc, value, gfp)				\
 	({								\
 		static struct lock_class_key __key;			\
 									\
-		__percpu_counter_init(fbc, value, &__key);		\
+		__percpu_counter_init(fbc, value, gfp, &__key);		\
 	})
 
 void percpu_counter_destroy(struct percpu_counter *fbc);
@@ -89,7 +90,8 @@ struct percpu_counter {
 	s64 count;
 };
 
-static inline int percpu_counter_init(struct percpu_counter *fbc, s64 amount)
+static inline int percpu_counter_init(struct percpu_counter *fbc, s64 amount,
+				      gfp_t gfp)
 {
 	fbc->count = amount;
 	return 0;
diff --git a/include/linux/proportions.h b/include/linux/proportions.h
index 26a8a4ed9b07..00e8e8fa7358 100644
--- a/include/linux/proportions.h
+++ b/include/linux/proportions.h
@@ -12,6 +12,7 @@
 #include <linux/percpu_counter.h>
 #include <linux/spinlock.h>
 #include <linux/mutex.h>
+#include <linux/gfp.h>
 
 struct prop_global {
 	/*
@@ -40,7 +41,7 @@ struct prop_descriptor {
 	struct mutex mutex;		/* serialize the prop_global switch */
 };
 
-int prop_descriptor_init(struct prop_descriptor *pd, int shift);
+int prop_descriptor_init(struct prop_descriptor *pd, int shift, gfp_t gfp);
 void prop_change_shift(struct prop_descriptor *pd, int new_shift);
 
 /*
@@ -61,7 +62,7 @@ struct prop_local_percpu {
 	raw_spinlock_t lock;		/* protect the snapshot state */
 };
 
-int prop_local_init_percpu(struct prop_local_percpu *pl);
+int prop_local_init_percpu(struct prop_local_percpu *pl, gfp_t gfp);
 void prop_local_destroy_percpu(struct prop_local_percpu *pl);
 void __prop_inc_percpu(struct prop_descriptor *pd, struct prop_local_percpu *pl);
 void prop_fraction_percpu(struct prop_descriptor *pd, struct prop_local_percpu *pl,
diff --git a/include/net/dst_ops.h b/include/net/dst_ops.h
index 2f26dfb8450e..1f99a1de0e4f 100644
--- a/include/net/dst_ops.h
+++ b/include/net/dst_ops.h
@@ -63,7 +63,7 @@ static inline void dst_entries_add(struct dst_ops *dst, int val)
 
 static inline int dst_entries_init(struct dst_ops *dst)
 {
-	return percpu_counter_init(&dst->pcpuc_entries, 0);
+	return percpu_counter_init(&dst->pcpuc_entries, 0, GFP_KERNEL);
 }
 
 static inline void dst_entries_destroy(struct dst_ops *dst)
diff --git a/include/net/inet_frag.h b/include/net/inet_frag.h
index 65a8855e99fe..8d1765577acc 100644
--- a/include/net/inet_frag.h
+++ b/include/net/inet_frag.h
@@ -151,7 +151,7 @@ static inline void add_frag_mem_limit(struct inet_frag_queue *q, int i)
 
 static inline void init_frag_mem_limit(struct netns_frags *nf)
 {
-	percpu_counter_init(&nf->mem, 0);
+	percpu_counter_init(&nf->mem, 0, GFP_KERNEL);
 }
 
 static inline unsigned int sum_frag_mem_limit(struct netns_frags *nf)
diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index cab7dc4284dc..136eceadeed1 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c
@@ -1607,7 +1607,8 @@ static int cgroup_setup_root(struct cgroup_root *root, unsigned int ss_mask)
 		goto out;
 	root_cgrp->id = ret;
 
-	ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release);
+	ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 0,
+			      GFP_KERNEL);
 	if (ret)
 		goto out;
 
@@ -4482,7 +4483,7 @@ static int create_css(struct cgroup *cgrp, struct cgroup_subsys *ss,
 
 	init_and_link_css(css, ss, cgrp);
 
-	err = percpu_ref_init(&css->refcnt, css_release);
+	err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
 	if (err)
 		goto err_free_css;
 
@@ -4555,7 +4556,7 @@ static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
 		goto out_unlock;
 	}
 
-	ret = percpu_ref_init(&cgrp->self.refcnt, css_release);
+	ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
 	if (ret)
 		goto out_free_cgrp;
 
diff --git a/lib/flex_proportions.c b/lib/flex_proportions.c
index ebf3bac460b0..8f25652f40d4 100644
--- a/lib/flex_proportions.c
+++ b/lib/flex_proportions.c
@@ -34,13 +34,13 @@
  */
 #include <linux/flex_proportions.h>
 
-int fprop_global_init(struct fprop_global *p)
+int fprop_global_init(struct fprop_global *p, gfp_t gfp)
 {
 	int err;
 
 	p->period = 0;
 	/* Use 1 to avoid dealing with periods with 0 events... */
-	err = percpu_counter_init(&p->events, 1);
+	err = percpu_counter_init(&p->events, 1, gfp);
 	if (err)
 		return err;
 	seqcount_init(&p->sequence);
@@ -168,11 +168,11 @@ void fprop_fraction_single(struct fprop_global *p,
  */
 #define PROP_BATCH (8*(1+ilog2(nr_cpu_ids)))
 
-int fprop_local_init_percpu(struct fprop_local_percpu *pl)
+int fprop_local_init_percpu(struct fprop_local_percpu *pl, gfp_t gfp)
 {
 	int err;
 
-	err = percpu_counter_init(&pl->events, 0);
+	err = percpu_counter_init(&pl->events, 0, gfp);
 	if (err)
 		return err;
 	pl->period = 0;
diff --git a/lib/percpu-refcount.c b/lib/percpu-refcount.c
index a89cf09a8268..6111bcb28376 100644
--- a/lib/percpu-refcount.c
+++ b/lib/percpu-refcount.c
@@ -1,6 +1,8 @@
 #define pr_fmt(fmt) "%s: " fmt "\n", __func__
 
 #include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/wait.h>
 #include <linux/percpu-refcount.h>
 
 /*
@@ -11,8 +13,8 @@
  * percpu counters will all sum to the correct value
  *
  * (More precisely: because moduler arithmatic is commutative the sum of all the
- * pcpu_count vars will be equal to what it would have been if all the gets and
- * puts were done to a single integer, even if some of the percpu integers
+ * percpu_count vars will be equal to what it would have been if all the gets
+ * and puts were done to a single integer, even if some of the percpu integers
  * overflow or underflow).
  *
  * The real trick to implementing percpu refcounts is shutdown. We can't detect
@@ -25,75 +27,64 @@
  * works.
  *
  * Converting to non percpu mode is done with some RCUish stuff in
- * percpu_ref_kill. Additionally, we need a bias value so that the atomic_t
- * can't hit 0 before we've added up all the percpu refs.
+ * percpu_ref_kill. Additionally, we need a bias value so that the
+ * atomic_long_t can't hit 0 before we've added up all the percpu refs.
  */
 
-#define PCPU_COUNT_BIAS		(1U << 31)
+#define PERCPU_COUNT_BIAS	(1LU << (BITS_PER_LONG - 1))
 
-static unsigned __percpu *pcpu_count_ptr(struct percpu_ref *ref)
+static DECLARE_WAIT_QUEUE_HEAD(percpu_ref_switch_waitq);
+
+static unsigned long __percpu *percpu_count_ptr(struct percpu_ref *ref)
 {
-	return (unsigned __percpu *)(ref->pcpu_count_ptr & ~PCPU_REF_DEAD);
+	return (unsigned long __percpu *)
+		(ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC_DEAD);
 }
 
 /**
  * percpu_ref_init - initialize a percpu refcount
  * @ref: percpu_ref to initialize
  * @release: function which will be called when refcount hits 0
+ * @flags: PERCPU_REF_INIT_* flags
+ * @gfp: allocation mask to use
  *
- * Initializes the refcount in single atomic counter mode with a refcount of 1;
- * analagous to atomic_set(ref, 1).
+ * Initializes @ref.  If @flags is zero, @ref starts in percpu mode with a
+ * refcount of 1; analagous to atomic_long_set(ref, 1).  See the
+ * definitions of PERCPU_REF_INIT_* flags for flag behaviors.
  *
  * Note that @release must not sleep - it may potentially be called from RCU
  * callback context by percpu_ref_kill().
  */
-int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release)
+int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release,
+		    unsigned int flags, gfp_t gfp)
 {
-	atomic_set(&ref->count, 1 + PCPU_COUNT_BIAS);
+	size_t align = max_t(size_t, 1 << __PERCPU_REF_FLAG_BITS,
+			     __alignof__(unsigned long));
+	unsigned long start_count = 0;
 
-	ref->pcpu_count_ptr = (unsigned long)alloc_percpu(unsigned);
-	if (!ref->pcpu_count_ptr)
+	ref->percpu_count_ptr = (unsigned long)
+		__alloc_percpu_gfp(sizeof(unsigned long), align, gfp);
+	if (!ref->percpu_count_ptr)
 		return -ENOMEM;
 
-	ref->release = release;
-	return 0;
-}
-EXPORT_SYMBOL_GPL(percpu_ref_init);
-
-/**
- * percpu_ref_reinit - re-initialize a percpu refcount
- * @ref: perpcu_ref to re-initialize
- *
- * Re-initialize @ref so that it's in the same state as when it finished
- * percpu_ref_init().  @ref must have been initialized successfully, killed
- * and reached 0 but not exited.
- *
- * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
- * this function is in progress.
- */
-void percpu_ref_reinit(struct percpu_ref *ref)
-{
-	unsigned __percpu *pcpu_count = pcpu_count_ptr(ref);
-	int cpu;
+	ref->force_atomic = flags & PERCPU_REF_INIT_ATOMIC;
 
-	BUG_ON(!pcpu_count);
-	WARN_ON(!percpu_ref_is_zero(ref));
+	if (flags & (PERCPU_REF_INIT_ATOMIC | PERCPU_REF_INIT_DEAD))
+		ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
+	else
+		start_count += PERCPU_COUNT_BIAS;
 
-	atomic_set(&ref->count, 1 + PCPU_COUNT_BIAS);
+	if (flags & PERCPU_REF_INIT_DEAD)
+		ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
+	else
+		start_count++;
 
-	/*
-	 * Restore per-cpu operation.  smp_store_release() is paired with
-	 * smp_read_barrier_depends() in __pcpu_ref_alive() and guarantees
-	 * that the zeroing is visible to all percpu accesses which can see
-	 * the following PCPU_REF_DEAD clearing.
-	 */
-	for_each_possible_cpu(cpu)
-		*per_cpu_ptr(pcpu_count, cpu) = 0;
+	atomic_long_set(&ref->count, start_count);
 
-	smp_store_release(&ref->pcpu_count_ptr,
-			  ref->pcpu_count_ptr & ~PCPU_REF_DEAD);
+	ref->release = release;
+	return 0;
 }
-EXPORT_SYMBOL_GPL(percpu_ref_reinit);
+EXPORT_SYMBOL_GPL(percpu_ref_init);
 
 /**
  * percpu_ref_exit - undo percpu_ref_init()
@@ -107,26 +98,39 @@ EXPORT_SYMBOL_GPL(percpu_ref_reinit);
  */
 void percpu_ref_exit(struct percpu_ref *ref)
 {
-	unsigned __percpu *pcpu_count = pcpu_count_ptr(ref);
+	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
 
-	if (pcpu_count) {
-		free_percpu(pcpu_count);
-		ref->pcpu_count_ptr = PCPU_REF_DEAD;
+	if (percpu_count) {
+		free_percpu(percpu_count);
+		ref->percpu_count_ptr = __PERCPU_REF_ATOMIC_DEAD;
 	}
 }
 EXPORT_SYMBOL_GPL(percpu_ref_exit);
 
-static void percpu_ref_kill_rcu(struct rcu_head *rcu)
+static void percpu_ref_call_confirm_rcu(struct rcu_head *rcu)
+{
+	struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu);
+
+	ref->confirm_switch(ref);
+	ref->confirm_switch = NULL;
+	wake_up_all(&percpu_ref_switch_waitq);
+
+	/* drop ref from percpu_ref_switch_to_atomic() */
+	percpu_ref_put(ref);
+}
+
+static void percpu_ref_switch_to_atomic_rcu(struct rcu_head *rcu)
 {
 	struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu);
-	unsigned __percpu *pcpu_count = pcpu_count_ptr(ref);
-	unsigned count = 0;
+	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
+	unsigned long count = 0;
 	int cpu;
 
 	for_each_possible_cpu(cpu)
-		count += *per_cpu_ptr(pcpu_count, cpu);
+		count += *per_cpu_ptr(percpu_count, cpu);
 
-	pr_debug("global %i pcpu %i", atomic_read(&ref->count), (int) count);
+	pr_debug("global %ld percpu %ld",
+		 atomic_long_read(&ref->count), (long)count);
 
 	/*
 	 * It's crucial that we sum the percpu counters _before_ adding the sum
@@ -140,21 +144,137 @@ static void percpu_ref_kill_rcu(struct rcu_head *rcu)
 	 * reaching 0 before we add the percpu counts. But doing it at the same
 	 * time is equivalent and saves us atomic operations:
 	 */
+	atomic_long_add((long)count - PERCPU_COUNT_BIAS, &ref->count);
+
+	WARN_ONCE(atomic_long_read(&ref->count) <= 0,
+		  "percpu ref (%pf) <= 0 (%ld) after switching to atomic",
+		  ref->release, atomic_long_read(&ref->count));
+
+	/* @ref is viewed as dead on all CPUs, send out switch confirmation */
+	percpu_ref_call_confirm_rcu(rcu);
+}
+
+static void percpu_ref_noop_confirm_switch(struct percpu_ref *ref)
+{
+}
+
+static void __percpu_ref_switch_to_atomic(struct percpu_ref *ref,
+					  percpu_ref_func_t *confirm_switch)
+{
+	if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC)) {
+		/* switching from percpu to atomic */
+		ref->percpu_count_ptr |= __PERCPU_REF_ATOMIC;
+
+		/*
+		 * Non-NULL ->confirm_switch is used to indicate that
+		 * switching is in progress.  Use noop one if unspecified.
+		 */
+		WARN_ON_ONCE(ref->confirm_switch);
+		ref->confirm_switch =
+			confirm_switch ?: percpu_ref_noop_confirm_switch;
+
+		percpu_ref_get(ref);	/* put after confirmation */
+		call_rcu_sched(&ref->rcu, percpu_ref_switch_to_atomic_rcu);
+	} else if (confirm_switch) {
+		/*
+		 * Somebody already set ATOMIC.  Switching may still be in
+		 * progress.  @confirm_switch must be invoked after the
+		 * switching is complete and a full sched RCU grace period
+		 * has passed.  Wait synchronously for the previous
+		 * switching and schedule @confirm_switch invocation.
+		 */
+		wait_event(percpu_ref_switch_waitq, !ref->confirm_switch);
+		ref->confirm_switch = confirm_switch;
 
-	atomic_add((int) count - PCPU_COUNT_BIAS, &ref->count);
+		percpu_ref_get(ref);	/* put after confirmation */
+		call_rcu_sched(&ref->rcu, percpu_ref_call_confirm_rcu);
+	}
+}
+
+/**
+ * percpu_ref_switch_to_atomic - switch a percpu_ref to atomic mode
+ * @ref: percpu_ref to switch to atomic mode
+ * @confirm_switch: optional confirmation callback
+ *
+ * There's no reason to use this function for the usual reference counting.
+ * Use percpu_ref_kill[_and_confirm]().
+ *
+ * Schedule switching of @ref to atomic mode.  All its percpu counts will
+ * be collected to the main atomic counter.  On completion, when all CPUs
+ * are guaraneed to be in atomic mode, @confirm_switch, which may not
+ * block, is invoked.  This function may be invoked concurrently with all
+ * the get/put operations and can safely be mixed with kill and reinit
+ * operations.  Note that @ref will stay in atomic mode across kill/reinit
+ * cycles until percpu_ref_switch_to_percpu() is called.
+ *
+ * This function normally doesn't block and can be called from any context
+ * but it may block if @confirm_kill is specified and @ref is already in
+ * the process of switching to atomic mode.  In such cases, @confirm_switch
+ * will be invoked after the switching is complete.
+ *
+ * Due to the way percpu_ref is implemented, @confirm_switch will be called
+ * after at least one full sched RCU grace period has passed but this is an
+ * implementation detail and must not be depended upon.
+ */
+void percpu_ref_switch_to_atomic(struct percpu_ref *ref,
+				 percpu_ref_func_t *confirm_switch)
+{
+	ref->force_atomic = true;
+	__percpu_ref_switch_to_atomic(ref, confirm_switch);
+}
 
-	WARN_ONCE(atomic_read(&ref->count) <= 0, "percpu ref <= 0 (%i)",
-		  atomic_read(&ref->count));
+static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
+{
+	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
+	int cpu;
+
+	BUG_ON(!percpu_count);
 
-	/* @ref is viewed as dead on all CPUs, send out kill confirmation */
-	if (ref->confirm_kill)
-		ref->confirm_kill(ref);
+	if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC))
+		return;
+
+	wait_event(percpu_ref_switch_waitq, !ref->confirm_switch);
+
+	atomic_long_add(PERCPU_COUNT_BIAS, &ref->count);
 
 	/*
-	 * Now we're in single atomic_t mode with a consistent refcount, so it's
-	 * safe to drop our initial ref:
+	 * Restore per-cpu operation.  smp_store_release() is paired with
+	 * smp_read_barrier_depends() in __ref_is_percpu() and guarantees
+	 * that the zeroing is visible to all percpu accesses which can see
+	 * the following __PERCPU_REF_ATOMIC clearing.
 	 */
-	percpu_ref_put(ref);
+	for_each_possible_cpu(cpu)
+		*per_cpu_ptr(percpu_count, cpu) = 0;
+
+	smp_store_release(&ref->percpu_count_ptr,
+			  ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC);
+}
+
+/**
+ * percpu_ref_switch_to_percpu - switch a percpu_ref to percpu mode
+ * @ref: percpu_ref to switch to percpu mode
+ *
+ * There's no reason to use this function for the usual reference counting.
+ * To re-use an expired ref, use percpu_ref_reinit().
+ *
+ * Switch @ref to percpu mode.  This function may be invoked concurrently
+ * with all the get/put operations and can safely be mixed with kill and
+ * reinit operations.  This function reverses the sticky atomic state set
+ * by PERCPU_REF_INIT_ATOMIC or percpu_ref_switch_to_atomic().  If @ref is
+ * dying or dead, the actual switching takes place on the following
+ * percpu_ref_reinit().
+ *
+ * This function normally doesn't block and can be called from any context
+ * but it may block if @ref is in the process of switching to atomic mode
+ * by percpu_ref_switch_atomic().
+ */
+void percpu_ref_switch_to_percpu(struct percpu_ref *ref)
+{
+	ref->force_atomic = false;
+
+	/* a dying or dead ref can't be switched to percpu mode w/o reinit */
+	if (!(ref->percpu_count_ptr & __PERCPU_REF_DEAD))
+		__percpu_ref_switch_to_percpu(ref);
 }
 
 /**
@@ -164,39 +284,48 @@ static void percpu_ref_kill_rcu(struct rcu_head *rcu)
  *
  * Equivalent to percpu_ref_kill() but also schedules kill confirmation if
  * @confirm_kill is not NULL.  @confirm_kill, which may not block, will be
- * called after @ref is seen as dead from all CPUs - all further
- * invocations of percpu_ref_tryget() will fail.  See percpu_ref_tryget()
- * for more details.
+ * called after @ref is seen as dead from all CPUs at which point all
+ * further invocations of percpu_ref_tryget_live() will fail.  See
+ * percpu_ref_tryget_live() for details.
+ *
+ * This function normally doesn't block and can be called from any context
+ * but it may block if @confirm_kill is specified and @ref is in the
+ * process of switching to atomic mode by percpu_ref_switch_atomic().
  *
- * Due to the way percpu_ref is implemented, @confirm_kill will be called
- * after at least one full RCU grace period has passed but this is an
- * implementation detail and callers must not depend on it.
+ * Due to the way percpu_ref is implemented, @confirm_switch will be called
+ * after at least one full sched RCU grace period has passed but this is an
+ * implementation detail and must not be depended upon.
  */
 void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
 				 percpu_ref_func_t *confirm_kill)
 {
-	WARN_ONCE(ref->pcpu_count_ptr & PCPU_REF_DEAD,
-		  "percpu_ref_kill() called more than once!\n");
+	WARN_ONCE(ref->percpu_count_ptr & __PERCPU_REF_DEAD,
+		  "%s called more than once on %pf!", __func__, ref->release);
 
-	ref->pcpu_count_ptr |= PCPU_REF_DEAD;
-	ref->confirm_kill = confirm_kill;
-
-	call_rcu_sched(&ref->rcu, percpu_ref_kill_rcu);
+	ref->percpu_count_ptr |= __PERCPU_REF_DEAD;
+	__percpu_ref_switch_to_atomic(ref, confirm_kill);
+	percpu_ref_put(ref);
 }
 EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm);
 
-/*
- * XXX: Temporary kludge to work around SCSI blk-mq stall.  Used only by
- * block/blk-mq.c::blk_mq_freeze_queue().  Will be removed during v3.18
- * devel cycle.  Do not use anywhere else.
+/**
+ * percpu_ref_reinit - re-initialize a percpu refcount
+ * @ref: perpcu_ref to re-initialize
+ *
+ * Re-initialize @ref so that it's in the same state as when it finished
+ * percpu_ref_init() ignoring %PERCPU_REF_INIT_DEAD.  @ref must have been
+ * initialized successfully and reached 0 but not exited.
+ *
+ * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
+ * this function is in progress.
  */
-void __percpu_ref_kill_expedited(struct percpu_ref *ref)
+void percpu_ref_reinit(struct percpu_ref *ref)
 {
-	WARN_ONCE(ref->pcpu_count_ptr & PCPU_REF_DEAD,
-		  "percpu_ref_kill() called more than once on %pf!",
-		  ref->release);
+	WARN_ON_ONCE(!percpu_ref_is_zero(ref));
 
-	ref->pcpu_count_ptr |= PCPU_REF_DEAD;
-	synchronize_sched_expedited();
-	percpu_ref_kill_rcu(&ref->rcu);
+	ref->percpu_count_ptr &= ~__PERCPU_REF_DEAD;
+	percpu_ref_get(ref);
+	if (!ref->force_atomic)
+		__percpu_ref_switch_to_percpu(ref);
 }
+EXPORT_SYMBOL_GPL(percpu_ref_reinit);
diff --git a/lib/percpu_counter.c b/lib/percpu_counter.c
index 7dd33577b905..48144cdae819 100644
--- a/lib/percpu_counter.c
+++ b/lib/percpu_counter.c
@@ -112,13 +112,15 @@ s64 __percpu_counter_sum(struct percpu_counter *fbc)
 }
 EXPORT_SYMBOL(__percpu_counter_sum);
 
-int __percpu_counter_init(struct percpu_counter *fbc, s64 amount,
+int __percpu_counter_init(struct percpu_counter *fbc, s64 amount, gfp_t gfp,
 			  struct lock_class_key *key)
 {
+	unsigned long flags __maybe_unused;
+
 	raw_spin_lock_init(&fbc->lock);
 	lockdep_set_class(&fbc->lock, key);
 	fbc->count = amount;
-	fbc->counters = alloc_percpu(s32);
+	fbc->counters = alloc_percpu_gfp(s32, gfp);
 	if (!fbc->counters)
 		return -ENOMEM;
 
@@ -126,9 +128,9 @@ int __percpu_counter_init(struct percpu_counter *fbc, s64 amount,
 
 #ifdef CONFIG_HOTPLUG_CPU
 	INIT_LIST_HEAD(&fbc->list);
-	spin_lock(&percpu_counters_lock);
+	spin_lock_irqsave(&percpu_counters_lock, flags);
 	list_add(&fbc->list, &percpu_counters);
-	spin_unlock(&percpu_counters_lock);
+	spin_unlock_irqrestore(&percpu_counters_lock, flags);
 #endif
 	return 0;
 }
@@ -136,15 +138,17 @@ EXPORT_SYMBOL(__percpu_counter_init);
 
 void percpu_counter_destroy(struct percpu_counter *fbc)
 {
+	unsigned long flags __maybe_unused;
+
 	if (!fbc->counters)
 		return;
 
 	debug_percpu_counter_deactivate(fbc);
 
 #ifdef CONFIG_HOTPLUG_CPU
-	spin_lock(&percpu_counters_lock);
+	spin_lock_irqsave(&percpu_counters_lock, flags);
 	list_del(&fbc->list);
-	spin_unlock(&percpu_counters_lock);
+	spin_unlock_irqrestore(&percpu_counters_lock, flags);
 #endif
 	free_percpu(fbc->counters);
 	fbc->counters = NULL;
@@ -173,7 +177,7 @@ static int percpu_counter_hotcpu_callback(struct notifier_block *nb,
 		return NOTIFY_OK;
 
 	cpu = (unsigned long)hcpu;
-	spin_lock(&percpu_counters_lock);
+	spin_lock_irq(&percpu_counters_lock);
 	list_for_each_entry(fbc, &percpu_counters, list) {
 		s32 *pcount;
 		unsigned long flags;
@@ -184,7 +188,7 @@ static int percpu_counter_hotcpu_callback(struct notifier_block *nb,
 		*pcount = 0;
 		raw_spin_unlock_irqrestore(&fbc->lock, flags);
 	}
-	spin_unlock(&percpu_counters_lock);
+	spin_unlock_irq(&percpu_counters_lock);
 #endif
 	return NOTIFY_OK;
 }
diff --git a/lib/proportions.c b/lib/proportions.c
index 05df84801b56..6f724298f67a 100644
--- a/lib/proportions.c
+++ b/lib/proportions.c
@@ -73,7 +73,7 @@
 #include <linux/proportions.h>
 #include <linux/rcupdate.h>
 
-int prop_descriptor_init(struct prop_descriptor *pd, int shift)
+int prop_descriptor_init(struct prop_descriptor *pd, int shift, gfp_t gfp)
 {
 	int err;
 
@@ -83,11 +83,11 @@ int prop_descriptor_init(struct prop_descriptor *pd, int shift)
 	pd->index = 0;
 	pd->pg[0].shift = shift;
 	mutex_init(&pd->mutex);
-	err = percpu_counter_init(&pd->pg[0].events, 0);
+	err = percpu_counter_init(&pd->pg[0].events, 0, gfp);
 	if (err)
 		goto out;
 
-	err = percpu_counter_init(&pd->pg[1].events, 0);
+	err = percpu_counter_init(&pd->pg[1].events, 0, gfp);
 	if (err)
 		percpu_counter_destroy(&pd->pg[0].events);
 
@@ -188,12 +188,12 @@ prop_adjust_shift(int *pl_shift, unsigned long *pl_period, int new_shift)
 
 #define PROP_BATCH (8*(1+ilog2(nr_cpu_ids)))
 
-int prop_local_init_percpu(struct prop_local_percpu *pl)
+int prop_local_init_percpu(struct prop_local_percpu *pl, gfp_t gfp)
 {
 	raw_spin_lock_init(&pl->lock);
 	pl->shift = 0;
 	pl->period = 0;
-	return percpu_counter_init(&pl->events, 0);
+	return percpu_counter_init(&pl->events, 0, gfp);
 }
 
 void prop_local_destroy_percpu(struct prop_local_percpu *pl)
diff --git a/mm/backing-dev.c b/mm/backing-dev.c
index b27714f1b40f..12a992b62576 100644
--- a/mm/backing-dev.c
+++ b/mm/backing-dev.c
@@ -455,7 +455,7 @@ int bdi_init(struct backing_dev_info *bdi)
 	bdi_wb_init(&bdi->wb, bdi);
 
 	for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
-		err = percpu_counter_init(&bdi->bdi_stat[i], 0);
+		err = percpu_counter_init(&bdi->bdi_stat[i], 0, GFP_KERNEL);
 		if (err)
 			goto err;
 	}
@@ -470,7 +470,7 @@ int bdi_init(struct backing_dev_info *bdi)
 	bdi->write_bandwidth = INIT_BW;
 	bdi->avg_write_bandwidth = INIT_BW;
 
-	err = fprop_local_init_percpu(&bdi->completions);
+	err = fprop_local_init_percpu(&bdi->completions, GFP_KERNEL);
 
 	if (err) {
 err:
diff --git a/mm/mmap.c b/mm/mmap.c
index 16d19b48e2ad..93d28c7e5420 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -3202,7 +3202,7 @@ void __init mmap_init(void)
 {
 	int ret;
 
-	ret = percpu_counter_init(&vm_committed_as, 0);
+	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
 	VM_BUG_ON(ret);
 }
 
diff --git a/mm/nommu.c b/mm/nommu.c
index a881d9673c6b..bd1808e194a7 100644
--- a/mm/nommu.c
+++ b/mm/nommu.c
@@ -539,7 +539,7 @@ void __init mmap_init(void)
 {
 	int ret;
 
-	ret = percpu_counter_init(&vm_committed_as, 0);
+	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
 	VM_BUG_ON(ret);
 	vm_region_jar = KMEM_CACHE(vm_region, SLAB_PANIC);
 }
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index 35ca7102d421..ff24c9d83112 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -1777,7 +1777,7 @@ void __init page_writeback_init(void)
 	writeback_set_ratelimit();
 	register_cpu_notifier(&ratelimit_nb);
 
-	fprop_global_init(&writeout_completions);
+	fprop_global_init(&writeout_completions, GFP_KERNEL);
 }
 
 /**
diff --git a/mm/percpu-km.c b/mm/percpu-km.c
index 89633fefc6a2..10e3d0b8a86d 100644
--- a/mm/percpu-km.c
+++ b/mm/percpu-km.c
@@ -33,17 +33,14 @@
 
 #include <linux/log2.h>
 
-static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
+static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
+			       int page_start, int page_end)
 {
-	unsigned int cpu;
-
-	for_each_possible_cpu(cpu)
-		memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
-
 	return 0;
 }
 
-static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
+static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
+				  int page_start, int page_end)
 {
 	/* nada */
 }
@@ -70,6 +67,11 @@ static struct pcpu_chunk *pcpu_create_chunk(void)
 
 	chunk->data = pages;
 	chunk->base_addr = page_address(pages) - pcpu_group_offsets[0];
+
+	spin_lock_irq(&pcpu_lock);
+	pcpu_chunk_populated(chunk, 0, nr_pages);
+	spin_unlock_irq(&pcpu_lock);
+
 	return chunk;
 }
 
diff --git a/mm/percpu-vm.c b/mm/percpu-vm.c
index 51108165f829..538998a137d2 100644
--- a/mm/percpu-vm.c
+++ b/mm/percpu-vm.c
@@ -20,46 +20,25 @@ static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
 }
 
 /**
- * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
+ * pcpu_get_pages - get temp pages array
  * @chunk: chunk of interest
- * @bitmapp: output parameter for bitmap
- * @may_alloc: may allocate the array
  *
- * Returns pointer to array of pointers to struct page and bitmap,
- * both of which can be indexed with pcpu_page_idx().  The returned
- * array is cleared to zero and *@bitmapp is copied from
- * @chunk->populated.  Note that there is only one array and bitmap
- * and access exclusion is the caller's responsibility.
- *
- * CONTEXT:
- * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
- * Otherwise, don't care.
+ * Returns pointer to array of pointers to struct page which can be indexed
+ * with pcpu_page_idx().  Note that there is only one array and accesses
+ * should be serialized by pcpu_alloc_mutex.
  *
  * RETURNS:
- * Pointer to temp pages array on success, NULL on failure.
+ * Pointer to temp pages array on success.
  */
-static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
-					       unsigned long **bitmapp,
-					       bool may_alloc)
+static struct page **pcpu_get_pages(struct pcpu_chunk *chunk_alloc)
 {
 	static struct page **pages;
-	static unsigned long *bitmap;
 	size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
-	size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
-			     sizeof(unsigned long);
-
-	if (!pages || !bitmap) {
-		if (may_alloc && !pages)
-			pages = pcpu_mem_zalloc(pages_size);
-		if (may_alloc && !bitmap)
-			bitmap = pcpu_mem_zalloc(bitmap_size);
-		if (!pages || !bitmap)
-			return NULL;
-	}
 
-	bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
+	lockdep_assert_held(&pcpu_alloc_mutex);
 
-	*bitmapp = bitmap;
+	if (!pages)
+		pages = pcpu_mem_zalloc(pages_size);
 	return pages;
 }
 
@@ -67,7 +46,6 @@ static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
  * pcpu_free_pages - free pages which were allocated for @chunk
  * @chunk: chunk pages were allocated for
  * @pages: array of pages to be freed, indexed by pcpu_page_idx()
- * @populated: populated bitmap
  * @page_start: page index of the first page to be freed
  * @page_end: page index of the last page to be freed + 1
  *
@@ -75,8 +53,7 @@ static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
  * The pages were allocated for @chunk.
  */
 static void pcpu_free_pages(struct pcpu_chunk *chunk,
-			    struct page **pages, unsigned long *populated,
-			    int page_start, int page_end)
+			    struct page **pages, int page_start, int page_end)
 {
 	unsigned int cpu;
 	int i;
@@ -95,7 +72,6 @@ static void pcpu_free_pages(struct pcpu_chunk *chunk,
  * pcpu_alloc_pages - allocates pages for @chunk
  * @chunk: target chunk
  * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
- * @populated: populated bitmap
  * @page_start: page index of the first page to be allocated
  * @page_end: page index of the last page to be allocated + 1
  *
@@ -104,8 +80,7 @@ static void pcpu_free_pages(struct pcpu_chunk *chunk,
  * content of @pages and will pass it verbatim to pcpu_map_pages().
  */
 static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
-			    struct page **pages, unsigned long *populated,
-			    int page_start, int page_end)
+			    struct page **pages, int page_start, int page_end)
 {
 	const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
 	unsigned int cpu, tcpu;
@@ -164,7 +139,6 @@ static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
  * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
  * @chunk: chunk of interest
  * @pages: pages array which can be used to pass information to free
- * @populated: populated bitmap
  * @page_start: page index of the first page to unmap
  * @page_end: page index of the last page to unmap + 1
  *
@@ -175,8 +149,7 @@ static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
  * proper pre/post flush functions.
  */
 static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
-			     struct page **pages, unsigned long *populated,
-			     int page_start, int page_end)
+			     struct page **pages, int page_start, int page_end)
 {
 	unsigned int cpu;
 	int i;
@@ -192,8 +165,6 @@ static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
 		__pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
 				   page_end - page_start);
 	}
-
-	bitmap_clear(populated, page_start, page_end - page_start);
 }
 
 /**
@@ -228,7 +199,6 @@ static int __pcpu_map_pages(unsigned long addr, struct page **pages,
  * pcpu_map_pages - map pages into a pcpu_chunk
  * @chunk: chunk of interest
  * @pages: pages array containing pages to be mapped
- * @populated: populated bitmap
  * @page_start: page index of the first page to map
  * @page_end: page index of the last page to map + 1
  *
@@ -236,13 +206,11 @@ static int __pcpu_map_pages(unsigned long addr, struct page **pages,
  * caller is responsible for calling pcpu_post_map_flush() after all
  * mappings are complete.
  *
- * This function is responsible for setting corresponding bits in
- * @chunk->populated bitmap and whatever is necessary for reverse
- * lookup (addr -> chunk).
+ * This function is responsible for setting up whatever is necessary for
+ * reverse lookup (addr -> chunk).
  */
 static int pcpu_map_pages(struct pcpu_chunk *chunk,
-			  struct page **pages, unsigned long *populated,
-			  int page_start, int page_end)
+			  struct page **pages, int page_start, int page_end)
 {
 	unsigned int cpu, tcpu;
 	int i, err;
@@ -253,18 +221,12 @@ static int pcpu_map_pages(struct pcpu_chunk *chunk,
 				       page_end - page_start);
 		if (err < 0)
 			goto err;
-	}
 
-	/* mapping successful, link chunk and mark populated */
-	for (i = page_start; i < page_end; i++) {
-		for_each_possible_cpu(cpu)
+		for (i = page_start; i < page_end; i++)
 			pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
 					    chunk);
-		__set_bit(i, populated);
 	}
-
 	return 0;
-
 err:
 	for_each_possible_cpu(tcpu) {
 		if (tcpu == cpu)
@@ -299,123 +261,69 @@ static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
 /**
  * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
  * @chunk: chunk of interest
- * @off: offset to the area to populate
- * @size: size of the area to populate in bytes
+ * @page_start: the start page
+ * @page_end: the end page
  *
  * For each cpu, populate and map pages [@page_start,@page_end) into
- * @chunk.  The area is cleared on return.
+ * @chunk.
  *
  * CONTEXT:
  * pcpu_alloc_mutex, does GFP_KERNEL allocation.
  */
-static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
+static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
+			       int page_start, int page_end)
 {
-	int page_start = PFN_DOWN(off);
-	int page_end = PFN_UP(off + size);
-	int free_end = page_start, unmap_end = page_start;
 	struct page **pages;
-	unsigned long *populated;
-	unsigned int cpu;
-	int rs, re, rc;
-
-	/* quick path, check whether all pages are already there */
-	rs = page_start;
-	pcpu_next_pop(chunk, &rs, &re, page_end);
-	if (rs == page_start && re == page_end)
-		goto clear;
 
-	/* need to allocate and map pages, this chunk can't be immutable */
-	WARN_ON(chunk->immutable);
-
-	pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
+	pages = pcpu_get_pages(chunk);
 	if (!pages)
 		return -ENOMEM;
 
-	/* alloc and map */
-	pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
-		rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
-		if (rc)
-			goto err_free;
-		free_end = re;
-	}
+	if (pcpu_alloc_pages(chunk, pages, page_start, page_end))
+		return -ENOMEM;
 
-	pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
-		rc = pcpu_map_pages(chunk, pages, populated, rs, re);
-		if (rc)
-			goto err_unmap;
-		unmap_end = re;
+	if (pcpu_map_pages(chunk, pages, page_start, page_end)) {
+		pcpu_free_pages(chunk, pages, page_start, page_end);
+		return -ENOMEM;
 	}
 	pcpu_post_map_flush(chunk, page_start, page_end);
 
-	/* commit new bitmap */
-	bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
-clear:
-	for_each_possible_cpu(cpu)
-		memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
 	return 0;
-
-err_unmap:
-	pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
-	pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
-		pcpu_unmap_pages(chunk, pages, populated, rs, re);
-	pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
-err_free:
-	pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
-		pcpu_free_pages(chunk, pages, populated, rs, re);
-	return rc;
 }
 
 /**
  * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
  * @chunk: chunk to depopulate
- * @off: offset to the area to depopulate
- * @size: size of the area to depopulate in bytes
+ * @page_start: the start page
+ * @page_end: the end page
  *
  * For each cpu, depopulate and unmap pages [@page_start,@page_end)
- * from @chunk.  If @flush is true, vcache is flushed before unmapping
- * and tlb after.
+ * from @chunk.
  *
  * CONTEXT:
  * pcpu_alloc_mutex.
  */
-static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
+static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
+				  int page_start, int page_end)
 {
-	int page_start = PFN_DOWN(off);
-	int page_end = PFN_UP(off + size);
 	struct page **pages;
-	unsigned long *populated;
-	int rs, re;
-
-	/* quick path, check whether it's empty already */
-	rs = page_start;
-	pcpu_next_unpop(chunk, &rs, &re, page_end);
-	if (rs == page_start && re == page_end)
-		return;
-
-	/* immutable chunks can't be depopulated */
-	WARN_ON(chunk->immutable);
 
 	/*
 	 * If control reaches here, there must have been at least one
 	 * successful population attempt so the temp pages array must
 	 * be available now.
 	 */
-	pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
+	pages = pcpu_get_pages(chunk);
 	BUG_ON(!pages);
 
 	/* unmap and free */
 	pcpu_pre_unmap_flush(chunk, page_start, page_end);
 
-	pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
-		pcpu_unmap_pages(chunk, pages, populated, rs, re);
+	pcpu_unmap_pages(chunk, pages, page_start, page_end);
 
 	/* no need to flush tlb, vmalloc will handle it lazily */
 
-	pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
-		pcpu_free_pages(chunk, pages, populated, rs, re);
-
-	/* commit new bitmap */
-	bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
+	pcpu_free_pages(chunk, pages, page_start, page_end);
 }
 
 static struct pcpu_chunk *pcpu_create_chunk(void)
diff --git a/mm/percpu.c b/mm/percpu.c
index da997f9800bd..014bab65e0ff 100644
--- a/mm/percpu.c
+++ b/mm/percpu.c
@@ -76,6 +76,10 @@
 
 #define PCPU_SLOT_BASE_SHIFT		5	/* 1-31 shares the same slot */
 #define PCPU_DFL_MAP_ALLOC		16	/* start a map with 16 ents */
+#define PCPU_ATOMIC_MAP_MARGIN_LOW	32
+#define PCPU_ATOMIC_MAP_MARGIN_HIGH	64
+#define PCPU_EMPTY_POP_PAGES_LOW	2
+#define PCPU_EMPTY_POP_PAGES_HIGH	4
 
 #ifdef CONFIG_SMP
 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
@@ -102,12 +106,16 @@ struct pcpu_chunk {
 	int			free_size;	/* free bytes in the chunk */
 	int			contig_hint;	/* max contiguous size hint */
 	void			*base_addr;	/* base address of this chunk */
+
 	int			map_used;	/* # of map entries used before the sentry */
 	int			map_alloc;	/* # of map entries allocated */
 	int			*map;		/* allocation map */
+	struct work_struct	map_extend_work;/* async ->map[] extension */
+
 	void			*data;		/* chunk data */
 	int			first_free;	/* no free below this */
 	bool			immutable;	/* no [de]population allowed */
+	int			nr_populated;	/* # of populated pages */
 	unsigned long		populated[];	/* populated bitmap */
 };
 
@@ -151,38 +159,33 @@ static struct pcpu_chunk *pcpu_first_chunk;
 static struct pcpu_chunk *pcpu_reserved_chunk;
 static int pcpu_reserved_chunk_limit;
 
+static DEFINE_SPINLOCK(pcpu_lock);	/* all internal data structures */
+static DEFINE_MUTEX(pcpu_alloc_mutex);	/* chunk create/destroy, [de]pop */
+
+static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
+
 /*
- * Synchronization rules.
- *
- * There are two locks - pcpu_alloc_mutex and pcpu_lock.  The former
- * protects allocation/reclaim paths, chunks, populated bitmap and
- * vmalloc mapping.  The latter is a spinlock and protects the index
- * data structures - chunk slots, chunks and area maps in chunks.
- *
- * During allocation, pcpu_alloc_mutex is kept locked all the time and
- * pcpu_lock is grabbed and released as necessary.  All actual memory
- * allocations are done using GFP_KERNEL with pcpu_lock released.  In
- * general, percpu memory can't be allocated with irq off but
- * irqsave/restore are still used in alloc path so that it can be used
- * from early init path - sched_init() specifically.
- *
- * Free path accesses and alters only the index data structures, so it
- * can be safely called from atomic context.  When memory needs to be
- * returned to the system, free path schedules reclaim_work which
- * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
- * reclaimed, release both locks and frees the chunks.  Note that it's
- * necessary to grab both locks to remove a chunk from circulation as
- * allocation path might be referencing the chunk with only
- * pcpu_alloc_mutex locked.
+ * The number of empty populated pages, protected by pcpu_lock.  The
+ * reserved chunk doesn't contribute to the count.
  */
-static DEFINE_MUTEX(pcpu_alloc_mutex);	/* protects whole alloc and reclaim */
-static DEFINE_SPINLOCK(pcpu_lock);	/* protects index data structures */
+static int pcpu_nr_empty_pop_pages;
 
-static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
+/*
+ * Balance work is used to populate or destroy chunks asynchronously.  We
+ * try to keep the number of populated free pages between
+ * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
+ * empty chunk.
+ */
+static void pcpu_balance_workfn(struct work_struct *work);
+static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
+static bool pcpu_async_enabled __read_mostly;
+static bool pcpu_atomic_alloc_failed;
 
-/* reclaim work to release fully free chunks, scheduled from free path */
-static void pcpu_reclaim(struct work_struct *work);
-static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
+static void pcpu_schedule_balance_work(void)
+{
+	if (pcpu_async_enabled)
+		schedule_work(&pcpu_balance_work);
+}
 
 static bool pcpu_addr_in_first_chunk(void *addr)
 {
@@ -315,6 +318,38 @@ static void pcpu_mem_free(void *ptr, size_t size)
 }
 
 /**
+ * pcpu_count_occupied_pages - count the number of pages an area occupies
+ * @chunk: chunk of interest
+ * @i: index of the area in question
+ *
+ * Count the number of pages chunk's @i'th area occupies.  When the area's
+ * start and/or end address isn't aligned to page boundary, the straddled
+ * page is included in the count iff the rest of the page is free.
+ */
+static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i)
+{
+	int off = chunk->map[i] & ~1;
+	int end = chunk->map[i + 1] & ~1;
+
+	if (!PAGE_ALIGNED(off) && i > 0) {
+		int prev = chunk->map[i - 1];
+
+		if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE))
+			off = round_down(off, PAGE_SIZE);
+	}
+
+	if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) {
+		int next = chunk->map[i + 1];
+		int nend = chunk->map[i + 2] & ~1;
+
+		if (!(next & 1) && nend >= round_up(end, PAGE_SIZE))
+			end = round_up(end, PAGE_SIZE);
+	}
+
+	return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0);
+}
+
+/**
  * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
  * @chunk: chunk of interest
  * @oslot: the previous slot it was on
@@ -342,9 +377,14 @@ static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
 /**
  * pcpu_need_to_extend - determine whether chunk area map needs to be extended
  * @chunk: chunk of interest
+ * @is_atomic: the allocation context
  *
- * Determine whether area map of @chunk needs to be extended to
- * accommodate a new allocation.
+ * Determine whether area map of @chunk needs to be extended.  If
+ * @is_atomic, only the amount necessary for a new allocation is
+ * considered; however, async extension is scheduled if the left amount is
+ * low.  If !@is_atomic, it aims for more empty space.  Combined, this
+ * ensures that the map is likely to have enough available space to
+ * accomodate atomic allocations which can't extend maps directly.
  *
  * CONTEXT:
  * pcpu_lock.
@@ -353,15 +393,26 @@ static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
  * New target map allocation length if extension is necessary, 0
  * otherwise.
  */
-static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
+static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic)
 {
-	int new_alloc;
+	int margin, new_alloc;
+
+	if (is_atomic) {
+		margin = 3;
+
+		if (chunk->map_alloc <
+		    chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW &&
+		    pcpu_async_enabled)
+			schedule_work(&chunk->map_extend_work);
+	} else {
+		margin = PCPU_ATOMIC_MAP_MARGIN_HIGH;
+	}
 
-	if (chunk->map_alloc >= chunk->map_used + 3)
+	if (chunk->map_alloc >= chunk->map_used + margin)
 		return 0;
 
 	new_alloc = PCPU_DFL_MAP_ALLOC;
-	while (new_alloc < chunk->map_used + 3)
+	while (new_alloc < chunk->map_used + margin)
 		new_alloc *= 2;
 
 	return new_alloc;
@@ -418,11 +469,76 @@ out_unlock:
 	return 0;
 }
 
+static void pcpu_map_extend_workfn(struct work_struct *work)
+{
+	struct pcpu_chunk *chunk = container_of(work, struct pcpu_chunk,
+						map_extend_work);
+	int new_alloc;
+
+	spin_lock_irq(&pcpu_lock);
+	new_alloc = pcpu_need_to_extend(chunk, false);
+	spin_unlock_irq(&pcpu_lock);
+
+	if (new_alloc)
+		pcpu_extend_area_map(chunk, new_alloc);
+}
+
+/**
+ * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
+ * @chunk: chunk the candidate area belongs to
+ * @off: the offset to the start of the candidate area
+ * @this_size: the size of the candidate area
+ * @size: the size of the target allocation
+ * @align: the alignment of the target allocation
+ * @pop_only: only allocate from already populated region
+ *
+ * We're trying to allocate @size bytes aligned at @align.  @chunk's area
+ * at @off sized @this_size is a candidate.  This function determines
+ * whether the target allocation fits in the candidate area and returns the
+ * number of bytes to pad after @off.  If the target area doesn't fit, -1
+ * is returned.
+ *
+ * If @pop_only is %true, this function only considers the already
+ * populated part of the candidate area.
+ */
+static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
+			    int size, int align, bool pop_only)
+{
+	int cand_off = off;
+
+	while (true) {
+		int head = ALIGN(cand_off, align) - off;
+		int page_start, page_end, rs, re;
+
+		if (this_size < head + size)
+			return -1;
+
+		if (!pop_only)
+			return head;
+
+		/*
+		 * If the first unpopulated page is beyond the end of the
+		 * allocation, the whole allocation is populated;
+		 * otherwise, retry from the end of the unpopulated area.
+		 */
+		page_start = PFN_DOWN(head + off);
+		page_end = PFN_UP(head + off + size);
+
+		rs = page_start;
+		pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
+		if (rs >= page_end)
+			return head;
+		cand_off = re * PAGE_SIZE;
+	}
+}
+
 /**
  * pcpu_alloc_area - allocate area from a pcpu_chunk
  * @chunk: chunk of interest
  * @size: wanted size in bytes
  * @align: wanted align
+ * @pop_only: allocate only from the populated area
+ * @occ_pages_p: out param for the number of pages the area occupies
  *
  * Try to allocate @size bytes area aligned at @align from @chunk.
  * Note that this function only allocates the offset.  It doesn't
@@ -437,7 +553,8 @@ out_unlock:
  * Allocated offset in @chunk on success, -1 if no matching area is
  * found.
  */
-static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
+static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
+			   bool pop_only, int *occ_pages_p)
 {
 	int oslot = pcpu_chunk_slot(chunk);
 	int max_contig = 0;
@@ -453,11 +570,11 @@ static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
 		if (off & 1)
 			continue;
 
-		/* extra for alignment requirement */
-		head = ALIGN(off, align) - off;
-
 		this_size = (p[1] & ~1) - off;
-		if (this_size < head + size) {
+
+		head = pcpu_fit_in_area(chunk, off, this_size, size, align,
+					pop_only);
+		if (head < 0) {
 			if (!seen_free) {
 				chunk->first_free = i;
 				seen_free = true;
@@ -526,6 +643,7 @@ static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
 		chunk->free_size -= size;
 		*p |= 1;
 
+		*occ_pages_p = pcpu_count_occupied_pages(chunk, i);
 		pcpu_chunk_relocate(chunk, oslot);
 		return off;
 	}
@@ -541,6 +659,7 @@ static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
  * pcpu_free_area - free area to a pcpu_chunk
  * @chunk: chunk of interest
  * @freeme: offset of area to free
+ * @occ_pages_p: out param for the number of pages the area occupies
  *
  * Free area starting from @freeme to @chunk.  Note that this function
  * only modifies the allocation map.  It doesn't depopulate or unmap
@@ -549,7 +668,8 @@ static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
  * CONTEXT:
  * pcpu_lock.
  */
-static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
+static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme,
+			   int *occ_pages_p)
 {
 	int oslot = pcpu_chunk_slot(chunk);
 	int off = 0;
@@ -580,6 +700,8 @@ static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
 	*p = off &= ~1;
 	chunk->free_size += (p[1] & ~1) - off;
 
+	*occ_pages_p = pcpu_count_occupied_pages(chunk, i);
+
 	/* merge with next? */
 	if (!(p[1] & 1))
 		to_free++;
@@ -620,6 +742,7 @@ static struct pcpu_chunk *pcpu_alloc_chunk(void)
 	chunk->map_used = 1;
 
 	INIT_LIST_HEAD(&chunk->list);
+	INIT_WORK(&chunk->map_extend_work, pcpu_map_extend_workfn);
 	chunk->free_size = pcpu_unit_size;
 	chunk->contig_hint = pcpu_unit_size;
 
@@ -634,6 +757,50 @@ static void pcpu_free_chunk(struct pcpu_chunk *chunk)
 	pcpu_mem_free(chunk, pcpu_chunk_struct_size);
 }
 
+/**
+ * pcpu_chunk_populated - post-population bookkeeping
+ * @chunk: pcpu_chunk which got populated
+ * @page_start: the start page
+ * @page_end: the end page
+ *
+ * Pages in [@page_start,@page_end) have been populated to @chunk.  Update
+ * the bookkeeping information accordingly.  Must be called after each
+ * successful population.
+ */
+static void pcpu_chunk_populated(struct pcpu_chunk *chunk,
+				 int page_start, int page_end)
+{
+	int nr = page_end - page_start;
+
+	lockdep_assert_held(&pcpu_lock);
+
+	bitmap_set(chunk->populated, page_start, nr);
+	chunk->nr_populated += nr;
+	pcpu_nr_empty_pop_pages += nr;
+}
+
+/**
+ * pcpu_chunk_depopulated - post-depopulation bookkeeping
+ * @chunk: pcpu_chunk which got depopulated
+ * @page_start: the start page
+ * @page_end: the end page
+ *
+ * Pages in [@page_start,@page_end) have been depopulated from @chunk.
+ * Update the bookkeeping information accordingly.  Must be called after
+ * each successful depopulation.
+ */
+static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
+				   int page_start, int page_end)
+{
+	int nr = page_end - page_start;
+
+	lockdep_assert_held(&pcpu_lock);
+
+	bitmap_clear(chunk->populated, page_start, nr);
+	chunk->nr_populated -= nr;
+	pcpu_nr_empty_pop_pages -= nr;
+}
+
 /*
  * Chunk management implementation.
  *
@@ -695,21 +862,23 @@ static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
  * @size: size of area to allocate in bytes
  * @align: alignment of area (max PAGE_SIZE)
  * @reserved: allocate from the reserved chunk if available
+ * @gfp: allocation flags
  *
- * Allocate percpu area of @size bytes aligned at @align.
- *
- * CONTEXT:
- * Does GFP_KERNEL allocation.
+ * Allocate percpu area of @size bytes aligned at @align.  If @gfp doesn't
+ * contain %GFP_KERNEL, the allocation is atomic.
  *
  * RETURNS:
  * Percpu pointer to the allocated area on success, NULL on failure.
  */
-static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
+static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
+				 gfp_t gfp)
 {
 	static int warn_limit = 10;
 	struct pcpu_chunk *chunk;
 	const char *err;
-	int slot, off, new_alloc;
+	bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
+	int occ_pages = 0;
+	int slot, off, new_alloc, cpu, ret;
 	unsigned long flags;
 	void __percpu *ptr;
 
@@ -728,7 +897,6 @@ static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
 		return NULL;
 	}
 
-	mutex_lock(&pcpu_alloc_mutex);
 	spin_lock_irqsave(&pcpu_lock, flags);
 
 	/* serve reserved allocations from the reserved chunk if available */
@@ -740,16 +908,18 @@ static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
 			goto fail_unlock;
 		}
 
-		while ((new_alloc = pcpu_need_to_extend(chunk))) {
+		while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) {
 			spin_unlock_irqrestore(&pcpu_lock, flags);
-			if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
+			if (is_atomic ||
+			    pcpu_extend_area_map(chunk, new_alloc) < 0) {
 				err = "failed to extend area map of reserved chunk";
-				goto fail_unlock_mutex;
+				goto fail;
 			}
 			spin_lock_irqsave(&pcpu_lock, flags);
 		}
 
-		off = pcpu_alloc_area(chunk, size, align);
+		off = pcpu_alloc_area(chunk, size, align, is_atomic,
+				      &occ_pages);
 		if (off >= 0)
 			goto area_found;
 
@@ -764,13 +934,15 @@ restart:
 			if (size > chunk->contig_hint)
 				continue;
 
-			new_alloc = pcpu_need_to_extend(chunk);
+			new_alloc = pcpu_need_to_extend(chunk, is_atomic);
 			if (new_alloc) {
+				if (is_atomic)
+					continue;
 				spin_unlock_irqrestore(&pcpu_lock, flags);
 				if (pcpu_extend_area_map(chunk,
 							 new_alloc) < 0) {
 					err = "failed to extend area map";
-					goto fail_unlock_mutex;
+					goto fail;
 				}
 				spin_lock_irqsave(&pcpu_lock, flags);
 				/*
@@ -780,74 +952,134 @@ restart:
 				goto restart;
 			}
 
-			off = pcpu_alloc_area(chunk, size, align);
+			off = pcpu_alloc_area(chunk, size, align, is_atomic,
+					      &occ_pages);
 			if (off >= 0)
 				goto area_found;
 		}
 	}
 
-	/* hmmm... no space left, create a new chunk */
 	spin_unlock_irqrestore(&pcpu_lock, flags);
 
-	chunk = pcpu_create_chunk();
-	if (!chunk) {
-		err = "failed to allocate new chunk";
-		goto fail_unlock_mutex;
+	/*
+	 * No space left.  Create a new chunk.  We don't want multiple
+	 * tasks to create chunks simultaneously.  Serialize and create iff
+	 * there's still no empty chunk after grabbing the mutex.
+	 */
+	if (is_atomic)
+		goto fail;
+
+	mutex_lock(&pcpu_alloc_mutex);
+
+	if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
+		chunk = pcpu_create_chunk();
+		if (!chunk) {
+			mutex_unlock(&pcpu_alloc_mutex);
+			err = "failed to allocate new chunk";
+			goto fail;
+		}
+
+		spin_lock_irqsave(&pcpu_lock, flags);
+		pcpu_chunk_relocate(chunk, -1);
+	} else {
+		spin_lock_irqsave(&pcpu_lock, flags);
 	}
 
-	spin_lock_irqsave(&pcpu_lock, flags);
-	pcpu_chunk_relocate(chunk, -1);
+	mutex_unlock(&pcpu_alloc_mutex);
 	goto restart;
 
 area_found:
 	spin_unlock_irqrestore(&pcpu_lock, flags);
 
-	/* populate, map and clear the area */
-	if (pcpu_populate_chunk(chunk, off, size)) {
-		spin_lock_irqsave(&pcpu_lock, flags);
-		pcpu_free_area(chunk, off);
-		err = "failed to populate";
-		goto fail_unlock;
+	/* populate if not all pages are already there */
+	if (!is_atomic) {
+		int page_start, page_end, rs, re;
+
+		mutex_lock(&pcpu_alloc_mutex);
+
+		page_start = PFN_DOWN(off);
+		page_end = PFN_UP(off + size);
+
+		pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
+			WARN_ON(chunk->immutable);
+
+			ret = pcpu_populate_chunk(chunk, rs, re);
+
+			spin_lock_irqsave(&pcpu_lock, flags);
+			if (ret) {
+				mutex_unlock(&pcpu_alloc_mutex);
+				pcpu_free_area(chunk, off, &occ_pages);
+				err = "failed to populate";
+				goto fail_unlock;
+			}
+			pcpu_chunk_populated(chunk, rs, re);
+			spin_unlock_irqrestore(&pcpu_lock, flags);
+		}
+
+		mutex_unlock(&pcpu_alloc_mutex);
 	}
 
-	mutex_unlock(&pcpu_alloc_mutex);
+	if (chunk != pcpu_reserved_chunk)
+		pcpu_nr_empty_pop_pages -= occ_pages;
+
+	if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
+		pcpu_schedule_balance_work();
+
+	/* clear the areas and return address relative to base address */
+	for_each_possible_cpu(cpu)
+		memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
 
-	/* return address relative to base address */
 	ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
 	kmemleak_alloc_percpu(ptr, size);
 	return ptr;
 
 fail_unlock:
 	spin_unlock_irqrestore(&pcpu_lock, flags);
-fail_unlock_mutex:
-	mutex_unlock(&pcpu_alloc_mutex);
-	if (warn_limit) {
-		pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
-			   "%s\n", size, align, err);
+fail:
+	if (!is_atomic && warn_limit) {
+		pr_warning("PERCPU: allocation failed, size=%zu align=%zu atomic=%d, %s\n",
+			   size, align, is_atomic, err);
 		dump_stack();
 		if (!--warn_limit)
 			pr_info("PERCPU: limit reached, disable warning\n");
 	}
+	if (is_atomic) {
+		/* see the flag handling in pcpu_blance_workfn() */
+		pcpu_atomic_alloc_failed = true;
+		pcpu_schedule_balance_work();
+	}
 	return NULL;
 }
 
 /**
- * __alloc_percpu - allocate dynamic percpu area
+ * __alloc_percpu_gfp - allocate dynamic percpu area
  * @size: size of area to allocate in bytes
  * @align: alignment of area (max PAGE_SIZE)
+ * @gfp: allocation flags
  *
- * Allocate zero-filled percpu area of @size bytes aligned at @align.
- * Might sleep.  Might trigger writeouts.
- *
- * CONTEXT:
- * Does GFP_KERNEL allocation.
+ * Allocate zero-filled percpu area of @size bytes aligned at @align.  If
+ * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
+ * be called from any context but is a lot more likely to fail.
  *
  * RETURNS:
  * Percpu pointer to the allocated area on success, NULL on failure.
  */
+void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
+{
+	return pcpu_alloc(size, align, false, gfp);
+}
+EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
+
+/**
+ * __alloc_percpu - allocate dynamic percpu area
+ * @size: size of area to allocate in bytes
+ * @align: alignment of area (max PAGE_SIZE)
+ *
+ * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
+ */
 void __percpu *__alloc_percpu(size_t size, size_t align)
 {
-	return pcpu_alloc(size, align, false);
+	return pcpu_alloc(size, align, false, GFP_KERNEL);
 }
 EXPORT_SYMBOL_GPL(__alloc_percpu);
 
@@ -869,44 +1101,121 @@ EXPORT_SYMBOL_GPL(__alloc_percpu);
  */
 void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
 {
-	return pcpu_alloc(size, align, true);
+	return pcpu_alloc(size, align, true, GFP_KERNEL);
 }
 
 /**
- * pcpu_reclaim - reclaim fully free chunks, workqueue function
+ * pcpu_balance_workfn - manage the amount of free chunks and populated pages
  * @work: unused
  *
  * Reclaim all fully free chunks except for the first one.
- *
- * CONTEXT:
- * workqueue context.
  */
-static void pcpu_reclaim(struct work_struct *work)
+static void pcpu_balance_workfn(struct work_struct *work)
 {
-	LIST_HEAD(todo);
-	struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
+	LIST_HEAD(to_free);
+	struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
 	struct pcpu_chunk *chunk, *next;
+	int slot, nr_to_pop, ret;
 
+	/*
+	 * There's no reason to keep around multiple unused chunks and VM
+	 * areas can be scarce.  Destroy all free chunks except for one.
+	 */
 	mutex_lock(&pcpu_alloc_mutex);
 	spin_lock_irq(&pcpu_lock);
 
-	list_for_each_entry_safe(chunk, next, head, list) {
+	list_for_each_entry_safe(chunk, next, free_head, list) {
 		WARN_ON(chunk->immutable);
 
 		/* spare the first one */
-		if (chunk == list_first_entry(head, struct pcpu_chunk, list))
+		if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
 			continue;
 
-		list_move(&chunk->list, &todo);
+		list_move(&chunk->list, &to_free);
 	}
 
 	spin_unlock_irq(&pcpu_lock);
 
-	list_for_each_entry_safe(chunk, next, &todo, list) {
-		pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
+	list_for_each_entry_safe(chunk, next, &to_free, list) {
+		int rs, re;
+
+		pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) {
+			pcpu_depopulate_chunk(chunk, rs, re);
+			spin_lock_irq(&pcpu_lock);
+			pcpu_chunk_depopulated(chunk, rs, re);
+			spin_unlock_irq(&pcpu_lock);
+		}
 		pcpu_destroy_chunk(chunk);
 	}
 
+	/*
+	 * Ensure there are certain number of free populated pages for
+	 * atomic allocs.  Fill up from the most packed so that atomic
+	 * allocs don't increase fragmentation.  If atomic allocation
+	 * failed previously, always populate the maximum amount.  This
+	 * should prevent atomic allocs larger than PAGE_SIZE from keeping
+	 * failing indefinitely; however, large atomic allocs are not
+	 * something we support properly and can be highly unreliable and
+	 * inefficient.
+	 */
+retry_pop:
+	if (pcpu_atomic_alloc_failed) {
+		nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
+		/* best effort anyway, don't worry about synchronization */
+		pcpu_atomic_alloc_failed = false;
+	} else {
+		nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
+				  pcpu_nr_empty_pop_pages,
+				  0, PCPU_EMPTY_POP_PAGES_HIGH);
+	}
+
+	for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
+		int nr_unpop = 0, rs, re;
+
+		if (!nr_to_pop)
+			break;
+
+		spin_lock_irq(&pcpu_lock);
+		list_for_each_entry(chunk, &pcpu_slot[slot], list) {
+			nr_unpop = pcpu_unit_pages - chunk->nr_populated;
+			if (nr_unpop)
+				break;
+		}
+		spin_unlock_irq(&pcpu_lock);
+
+		if (!nr_unpop)
+			continue;
+
+		/* @chunk can't go away while pcpu_alloc_mutex is held */
+		pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) {
+			int nr = min(re - rs, nr_to_pop);
+
+			ret = pcpu_populate_chunk(chunk, rs, rs + nr);
+			if (!ret) {
+				nr_to_pop -= nr;
+				spin_lock_irq(&pcpu_lock);
+				pcpu_chunk_populated(chunk, rs, rs + nr);
+				spin_unlock_irq(&pcpu_lock);
+			} else {
+				nr_to_pop = 0;
+			}
+
+			if (!nr_to_pop)
+				break;
+		}
+	}
+
+	if (nr_to_pop) {
+		/* ran out of chunks to populate, create a new one and retry */
+		chunk = pcpu_create_chunk();
+		if (chunk) {
+			spin_lock_irq(&pcpu_lock);
+			pcpu_chunk_relocate(chunk, -1);
+			spin_unlock_irq(&pcpu_lock);
+			goto retry_pop;
+		}
+	}
+
 	mutex_unlock(&pcpu_alloc_mutex);
 }
 
@@ -924,7 +1233,7 @@ void free_percpu(void __percpu *ptr)
 	void *addr;
 	struct pcpu_chunk *chunk;
 	unsigned long flags;
-	int off;
+	int off, occ_pages;
 
 	if (!ptr)
 		return;
@@ -938,7 +1247,10 @@ void free_percpu(void __percpu *ptr)
 	chunk = pcpu_chunk_addr_search(addr);
 	off = addr - chunk->base_addr;
 
-	pcpu_free_area(chunk, off);
+	pcpu_free_area(chunk, off, &occ_pages);
+
+	if (chunk != pcpu_reserved_chunk)
+		pcpu_nr_empty_pop_pages += occ_pages;
 
 	/* if there are more than one fully free chunks, wake up grim reaper */
 	if (chunk->free_size == pcpu_unit_size) {
@@ -946,7 +1258,7 @@ void free_percpu(void __percpu *ptr)
 
 		list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
 			if (pos != chunk) {
-				schedule_work(&pcpu_reclaim_work);
+				pcpu_schedule_balance_work();
 				break;
 			}
 	}
@@ -1336,11 +1648,13 @@ int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
 	 */
 	schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
 	INIT_LIST_HEAD(&schunk->list);
+	INIT_WORK(&schunk->map_extend_work, pcpu_map_extend_workfn);
 	schunk->base_addr = base_addr;
 	schunk->map = smap;
 	schunk->map_alloc = ARRAY_SIZE(smap);
 	schunk->immutable = true;
 	bitmap_fill(schunk->populated, pcpu_unit_pages);
+	schunk->nr_populated = pcpu_unit_pages;
 
 	if (ai->reserved_size) {
 		schunk->free_size = ai->reserved_size;
@@ -1364,11 +1678,13 @@ int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
 	if (dyn_size) {
 		dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
 		INIT_LIST_HEAD(&dchunk->list);
+		INIT_WORK(&dchunk->map_extend_work, pcpu_map_extend_workfn);
 		dchunk->base_addr = base_addr;
 		dchunk->map = dmap;
 		dchunk->map_alloc = ARRAY_SIZE(dmap);
 		dchunk->immutable = true;
 		bitmap_fill(dchunk->populated, pcpu_unit_pages);
+		dchunk->nr_populated = pcpu_unit_pages;
 
 		dchunk->contig_hint = dchunk->free_size = dyn_size;
 		dchunk->map[0] = 1;
@@ -1379,6 +1695,8 @@ int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
 
 	/* link the first chunk in */
 	pcpu_first_chunk = dchunk ?: schunk;
+	pcpu_nr_empty_pop_pages +=
+		pcpu_count_occupied_pages(pcpu_first_chunk, 1);
 	pcpu_chunk_relocate(pcpu_first_chunk, -1);
 
 	/* we're done */
@@ -1932,8 +2250,6 @@ void __init setup_per_cpu_areas(void)
 
 	if (pcpu_setup_first_chunk(ai, fc) < 0)
 		panic("Failed to initialize percpu areas.");
-
-	pcpu_free_alloc_info(ai);
 }
 
 #endif	/* CONFIG_SMP */
@@ -1967,3 +2283,15 @@ void __init percpu_init_late(void)
 		spin_unlock_irqrestore(&pcpu_lock, flags);
 	}
 }
+
+/*
+ * Percpu allocator is initialized early during boot when neither slab or
+ * workqueue is available.  Plug async management until everything is up
+ * and running.
+ */
+static int __init percpu_enable_async(void)
+{
+	pcpu_async_enabled = true;
+	return 0;
+}
+subsys_initcall(percpu_enable_async);
diff --git a/mm/shmem.c b/mm/shmem.c
index 4fad61bb41e5..cd6fc7590e54 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -2995,7 +2995,7 @@ int shmem_fill_super(struct super_block *sb, void *data, int silent)
 #endif
 
 	spin_lock_init(&sbinfo->stat_lock);
-	if (percpu_counter_init(&sbinfo->used_blocks, 0))
+	if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
 		goto failed;
 	sbinfo->free_inodes = sbinfo->max_inodes;
 
diff --git a/net/dccp/proto.c b/net/dccp/proto.c
index 97b0fcc79547..5ab6627cf370 100644
--- a/net/dccp/proto.c
+++ b/net/dccp/proto.c
@@ -1115,7 +1115,7 @@ static int __init dccp_init(void)
 
 	BUILD_BUG_ON(sizeof(struct dccp_skb_cb) >
 		     FIELD_SIZEOF(struct sk_buff, cb));
-	rc = percpu_counter_init(&dccp_orphan_count, 0);
+	rc = percpu_counter_init(&dccp_orphan_count, 0, GFP_KERNEL);
 	if (rc)
 		goto out_fail;
 	rc = -ENOBUFS;
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 461003d258ba..86023b9be47f 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -3071,8 +3071,8 @@ void __init tcp_init(void)
 
 	BUILD_BUG_ON(sizeof(struct tcp_skb_cb) > sizeof(skb->cb));
 
-	percpu_counter_init(&tcp_sockets_allocated, 0);
-	percpu_counter_init(&tcp_orphan_count, 0);
+	percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
+	percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
 	tcp_hashinfo.bind_bucket_cachep =
 		kmem_cache_create("tcp_bind_bucket",
 				  sizeof(struct inet_bind_bucket), 0,
diff --git a/net/ipv4/tcp_memcontrol.c b/net/ipv4/tcp_memcontrol.c
index 3af522622fad..1d191357bf88 100644
--- a/net/ipv4/tcp_memcontrol.c
+++ b/net/ipv4/tcp_memcontrol.c
@@ -32,7 +32,7 @@ int tcp_init_cgroup(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
 		res_parent = &parent_cg->memory_allocated;
 
 	res_counter_init(&cg_proto->memory_allocated, res_parent);
-	percpu_counter_init(&cg_proto->sockets_allocated, 0);
+	percpu_counter_init(&cg_proto->sockets_allocated, 0, GFP_KERNEL);
 
 	return 0;
 }
diff --git a/net/sctp/protocol.c b/net/sctp/protocol.c
index 9d2c6c9facb6..8f34b27d5775 100644
--- a/net/sctp/protocol.c
+++ b/net/sctp/protocol.c
@@ -1341,7 +1341,7 @@ static __init int sctp_init(void)
 	if (!sctp_chunk_cachep)
 		goto err_chunk_cachep;
 
-	status = percpu_counter_init(&sctp_sockets_allocated, 0);
+	status = percpu_counter_init(&sctp_sockets_allocated, 0, GFP_KERNEL);
 	if (status)
 		goto err_percpu_counter_init;