Re: zram: per-cpu compression streams

[Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 2519 bytes --]

Hello Minchan,

On (03/25/16 08:41), Minchan Kim wrote:
[..]
> >  Test #10 iozone -t 10 -R -r 80K -s 0M -I +Z
> >    Initial write        3213973.56      2731512.62      4416466.25*
> >          Rewrite        3066956.44*     2693819.50       332671.94
> >             Read        7769523.25*     2681473.75       462840.44
> >          Re-read        5244861.75      5473037.00*      382183.03
> >     Reverse Read        7479397.25*     4869597.75       374714.06
> >      Stride read        5403282.50*     5385083.75       382473.44
> >      Random read        5131997.25      5176799.75*      380593.56
> >   Mixed workload        3998043.25      4219049.00*     1645850.45
> >     Random write        3452832.88      3290861.69      3588531.75*
> >           Pwrite        3757435.81      2711756.47      4561807.88*
> >            Pread        2743595.25*     2635835.00       412947.98
> >           Fwrite       16076549.00     16741977.25*    14797209.38
> >            Fread       23581812.62*    21664184.25      5064296.97
> >  =          real         0m44.490s       0m44.444s       0m44.609s
> >  =          user          0m0.054s        0m0.049s        0m0.055s
> >  =           sys          0m0.037s        0m0.046s        0m0.148s
> >  
> >  
> >  so when the number of active tasks become larger than the number
> >  of online CPUS, iozone reports a bit hard to understand data. I
> >  can assume that since now we keep the preemption disabled longer
> >  in write path, a concurrent operation (READ or WRITE) cannot preempt
> >  current anymore... slightly suspicious.
> >  
> >  the other hard to understand thing is why do READ-only tests have
> >  such a huge jitter. READ-only tests don't depend on streams, they
> >  don't even use them, we supply compressed data directly to
> >  decompression api.
> >  
> >  may be better retire iozone and never use it again.
> >  
> >  
> >  "118 insertions(+), 238 deletions(-)" the patches remove a big
> >  pile of code.
> 
> First of all, I appreciate you very much!

thanks!

> At a glance, on write workload, huge win but worth to investigate
> how such fluctuation/regression happens on read-related test
> (read and mixed workload).

yes, was going to investigate in more details but got interrupted,
will return back to it today/tomorrow.

> Could you send your patchset? I will test it.

oh, sorry, sure! attached (because it's not a real patch submission
yet, but they look more or less ready I guess).

patches are against next-20160324.

	-ss

[-- Attachment #2: 0001-zsmalloc-require-GFP-in-zs_malloc.patch --]
[-- Type: text/x-diff, Size: 3948 bytes --]

>From 6bf369f2180dad1c8013a4847ec09d3b9056e910 Mon Sep 17 00:00:00 2001
From: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Subject: [PATCH 1/2] zsmalloc: require GFP in zs_malloc()

Pass GFP flags to zs_malloc() instead of using fixed ones (set
during pool creation), so we can be more flexible. Apart from
that, this also align zs_malloc() interface with zspool/zbud.

Signed-off-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
---
 drivers/block/zram/zram_drv.c |  2 +-
 include/linux/zsmalloc.h      |  2 +-
 mm/zsmalloc.c                 | 15 ++++++---------
 3 files changed, 8 insertions(+), 11 deletions(-)

diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
index 370c2f7..9030992 100644
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c
@@ -717,7 +717,7 @@ static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index,
 			src = uncmem;
 	}
 
-	handle = zs_malloc(meta->mem_pool, clen);
+	handle = zs_malloc(meta->mem_pool, clen, GFP_NOIO | __GFP_HIGHMEM);
 	if (!handle) {
 		pr_err("Error allocating memory for compressed page: %u, size=%zu\n",
 			index, clen);
diff --git a/include/linux/zsmalloc.h b/include/linux/zsmalloc.h
index 34eb160..6d89f8b 100644
--- a/include/linux/zsmalloc.h
+++ b/include/linux/zsmalloc.h
@@ -44,7 +44,7 @@ struct zs_pool;
 struct zs_pool *zs_create_pool(const char *name, gfp_t flags);
 void zs_destroy_pool(struct zs_pool *pool);
 
-unsigned long zs_malloc(struct zs_pool *pool, size_t size);
+unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t flags);
 void zs_free(struct zs_pool *pool, unsigned long obj);
 
 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
diff --git a/mm/zsmalloc.c b/mm/zsmalloc.c
index e72efb1..19027a1 100644
--- a/mm/zsmalloc.c
+++ b/mm/zsmalloc.c
@@ -247,7 +247,6 @@ struct zs_pool {
 	struct size_class **size_class;
 	struct kmem_cache *handle_cachep;
 
-	gfp_t flags;	/* allocation flags used when growing pool */
 	atomic_long_t pages_allocated;
 
 	struct zs_pool_stats stats;
@@ -295,10 +294,10 @@ static void destroy_handle_cache(struct zs_pool *pool)
 	kmem_cache_destroy(pool->handle_cachep);
 }
 
-static unsigned long alloc_handle(struct zs_pool *pool)
+static unsigned long alloc_handle(struct zs_pool *pool, gfp_t gfp)
 {
 	return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
-		pool->flags & ~__GFP_HIGHMEM);
+		gfp & ~__GFP_HIGHMEM);
 }
 
 static void free_handle(struct zs_pool *pool, unsigned long handle)
@@ -335,7 +334,7 @@ static void zs_zpool_destroy(void *pool)
 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
 			unsigned long *handle)
 {
-	*handle = zs_malloc(pool, size);
+	*handle = zs_malloc(pool, size, gfp);
 	return *handle ? 0 : -1;
 }
 static void zs_zpool_free(void *pool, unsigned long handle)
@@ -1391,7 +1390,7 @@ static unsigned long obj_malloc(struct page *first_page,
  * otherwise 0.
  * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
  */
-unsigned long zs_malloc(struct zs_pool *pool, size_t size)
+unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
 {
 	unsigned long handle, obj;
 	struct size_class *class;
@@ -1400,7 +1399,7 @@ unsigned long zs_malloc(struct zs_pool *pool, size_t size)
 	if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
 		return 0;
 
-	handle = alloc_handle(pool);
+	handle = alloc_handle(pool, gfp);
 	if (!handle)
 		return 0;
 
@@ -1413,7 +1412,7 @@ unsigned long zs_malloc(struct zs_pool *pool, size_t size)
 
 	if (!first_page) {
 		spin_unlock(&class->lock);
-		first_page = alloc_zspage(class, pool->flags);
+		first_page = alloc_zspage(class, gfp);
 		if (unlikely(!first_page)) {
 			free_handle(pool, handle);
 			return 0;
@@ -1954,8 +1953,6 @@ struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
 		prev_class = class;
 	}
 
-	pool->flags = flags;
-
 	if (zs_pool_stat_create(name, pool))
 		goto err;
 
-- 
2.8.0.rc0.1.gd285ab0


[-- Attachment #3: 0002-zram-switch-to-per-cpu-streams.patch --]
[-- Type: text/x-diff, Size: 13134 bytes --]

>From 02e997f42a77ebf9e2fdeb84c5e7eaf4dca21b37 Mon Sep 17 00:00:00 2001
From: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Subject: [PATCH 2/2] zram: switch to per-cpu streams

Use per-cpu compression streams.

Signed-off-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
---
 drivers/block/zram/zcomp.c    | 293 ++++++++++++------------------------------
 drivers/block/zram/zcomp.h    |  12 +-
 drivers/block/zram/zram_drv.c |  32 ++++-
 3 files changed, 110 insertions(+), 227 deletions(-)

diff --git a/drivers/block/zram/zcomp.c b/drivers/block/zram/zcomp.c
index 3ef42e5..d4159e4 100644
--- a/drivers/block/zram/zcomp.c
+++ b/drivers/block/zram/zcomp.c
@@ -13,6 +13,7 @@
 #include <linux/slab.h>
 #include <linux/wait.h>
 #include <linux/sched.h>
+#include <linux/cpu.h>
 
 #include "zcomp.h"
 #include "zcomp_lzo.h"
@@ -20,29 +21,6 @@
 #include "zcomp_lz4.h"
 #endif
 
-/*
- * single zcomp_strm backend
- */
-struct zcomp_strm_single {
-	struct mutex strm_lock;
-	struct zcomp_strm *zstrm;
-};
-
-/*
- * multi zcomp_strm backend
- */
-struct zcomp_strm_multi {
-	/* protect strm list */
-	spinlock_t strm_lock;
-	/* max possible number of zstrm streams */
-	int max_strm;
-	/* number of available zstrm streams */
-	int avail_strm;
-	/* list of available strms */
-	struct list_head idle_strm;
-	wait_queue_head_t strm_wait;
-};
-
 static struct zcomp_backend *backends[] = {
 	&zcomp_lzo,
 #ifdef CONFIG_ZRAM_LZ4_COMPRESS
@@ -93,188 +71,6 @@ static struct zcomp_strm *zcomp_strm_alloc(struct zcomp *comp, gfp_t flags)
 	return zstrm;
 }
 
-/*
- * get idle zcomp_strm or wait until other process release
- * (zcomp_strm_release()) one for us
- */
-static struct zcomp_strm *zcomp_strm_multi_find(struct zcomp *comp)
-{
-	struct zcomp_strm_multi *zs = comp->stream;
-	struct zcomp_strm *zstrm;
-
-	while (1) {
-		spin_lock(&zs->strm_lock);
-		if (!list_empty(&zs->idle_strm)) {
-			zstrm = list_entry(zs->idle_strm.next,
-					struct zcomp_strm, list);
-			list_del(&zstrm->list);
-			spin_unlock(&zs->strm_lock);
-			return zstrm;
-		}
-		/* zstrm streams limit reached, wait for idle stream */
-		if (zs->avail_strm >= zs->max_strm) {
-			spin_unlock(&zs->strm_lock);
-			wait_event(zs->strm_wait, !list_empty(&zs->idle_strm));
-			continue;
-		}
-		/* allocate new zstrm stream */
-		zs->avail_strm++;
-		spin_unlock(&zs->strm_lock);
-		/*
-		 * This function can be called in swapout/fs write path
-		 * so we can't use GFP_FS|IO. And it assumes we already
-		 * have at least one stream in zram initialization so we
-		 * don't do best effort to allocate more stream in here.
-		 * A default stream will work well without further multiple
-		 * streams. That's why we use NORETRY | NOWARN.
-		 */
-		zstrm = zcomp_strm_alloc(comp, GFP_NOIO | __GFP_NORETRY |
-					__GFP_NOWARN);
-		if (!zstrm) {
-			spin_lock(&zs->strm_lock);
-			zs->avail_strm--;
-			spin_unlock(&zs->strm_lock);
-			wait_event(zs->strm_wait, !list_empty(&zs->idle_strm));
-			continue;
-		}
-		break;
-	}
-	return zstrm;
-}
-
-/* add stream back to idle list and wake up waiter or free the stream */
-static void zcomp_strm_multi_release(struct zcomp *comp, struct zcomp_strm *zstrm)
-{
-	struct zcomp_strm_multi *zs = comp->stream;
-
-	spin_lock(&zs->strm_lock);
-	if (zs->avail_strm <= zs->max_strm) {
-		list_add(&zstrm->list, &zs->idle_strm);
-		spin_unlock(&zs->strm_lock);
-		wake_up(&zs->strm_wait);
-		return;
-	}
-
-	zs->avail_strm--;
-	spin_unlock(&zs->strm_lock);
-	zcomp_strm_free(comp, zstrm);
-}
-
-/* change max_strm limit */
-static bool zcomp_strm_multi_set_max_streams(struct zcomp *comp, int num_strm)
-{
-	struct zcomp_strm_multi *zs = comp->stream;
-	struct zcomp_strm *zstrm;
-
-	spin_lock(&zs->strm_lock);
-	zs->max_strm = num_strm;
-	/*
-	 * if user has lowered the limit and there are idle streams,
-	 * immediately free as much streams (and memory) as we can.
-	 */
-	while (zs->avail_strm > num_strm && !list_empty(&zs->idle_strm)) {
-		zstrm = list_entry(zs->idle_strm.next,
-				struct zcomp_strm, list);
-		list_del(&zstrm->list);
-		zcomp_strm_free(comp, zstrm);
-		zs->avail_strm--;
-	}
-	spin_unlock(&zs->strm_lock);
-	return true;
-}
-
-static void zcomp_strm_multi_destroy(struct zcomp *comp)
-{
-	struct zcomp_strm_multi *zs = comp->stream;
-	struct zcomp_strm *zstrm;
-
-	while (!list_empty(&zs->idle_strm)) {
-		zstrm = list_entry(zs->idle_strm.next,
-				struct zcomp_strm, list);
-		list_del(&zstrm->list);
-		zcomp_strm_free(comp, zstrm);
-	}
-	kfree(zs);
-}
-
-static int zcomp_strm_multi_create(struct zcomp *comp, int max_strm)
-{
-	struct zcomp_strm *zstrm;
-	struct zcomp_strm_multi *zs;
-
-	comp->destroy = zcomp_strm_multi_destroy;
-	comp->strm_find = zcomp_strm_multi_find;
-	comp->strm_release = zcomp_strm_multi_release;
-	comp->set_max_streams = zcomp_strm_multi_set_max_streams;
-	zs = kmalloc(sizeof(struct zcomp_strm_multi), GFP_KERNEL);
-	if (!zs)
-		return -ENOMEM;
-
-	comp->stream = zs;
-	spin_lock_init(&zs->strm_lock);
-	INIT_LIST_HEAD(&zs->idle_strm);
-	init_waitqueue_head(&zs->strm_wait);
-	zs->max_strm = max_strm;
-	zs->avail_strm = 1;
-
-	zstrm = zcomp_strm_alloc(comp, GFP_KERNEL);
-	if (!zstrm) {
-		kfree(zs);
-		return -ENOMEM;
-	}
-	list_add(&zstrm->list, &zs->idle_strm);
-	return 0;
-}
-
-static struct zcomp_strm *zcomp_strm_single_find(struct zcomp *comp)
-{
-	struct zcomp_strm_single *zs = comp->stream;
-	mutex_lock(&zs->strm_lock);
-	return zs->zstrm;
-}
-
-static void zcomp_strm_single_release(struct zcomp *comp,
-		struct zcomp_strm *zstrm)
-{
-	struct zcomp_strm_single *zs = comp->stream;
-	mutex_unlock(&zs->strm_lock);
-}
-
-static bool zcomp_strm_single_set_max_streams(struct zcomp *comp, int num_strm)
-{
-	/* zcomp_strm_single support only max_comp_streams == 1 */
-	return false;
-}
-
-static void zcomp_strm_single_destroy(struct zcomp *comp)
-{
-	struct zcomp_strm_single *zs = comp->stream;
-	zcomp_strm_free(comp, zs->zstrm);
-	kfree(zs);
-}
-
-static int zcomp_strm_single_create(struct zcomp *comp)
-{
-	struct zcomp_strm_single *zs;
-
-	comp->destroy = zcomp_strm_single_destroy;
-	comp->strm_find = zcomp_strm_single_find;
-	comp->strm_release = zcomp_strm_single_release;
-	comp->set_max_streams = zcomp_strm_single_set_max_streams;
-	zs = kmalloc(sizeof(struct zcomp_strm_single), GFP_KERNEL);
-	if (!zs)
-		return -ENOMEM;
-
-	comp->stream = zs;
-	mutex_init(&zs->strm_lock);
-	zs->zstrm = zcomp_strm_alloc(comp, GFP_KERNEL);
-	if (!zs->zstrm) {
-		kfree(zs);
-		return -ENOMEM;
-	}
-	return 0;
-}
-
 /* show available compressors */
 ssize_t zcomp_available_show(const char *comp, char *buf)
 {
@@ -301,17 +97,17 @@ bool zcomp_available_algorithm(const char *comp)
 
 bool zcomp_set_max_streams(struct zcomp *comp, int num_strm)
 {
-	return comp->set_max_streams(comp, num_strm);
+	return true;
 }
 
 struct zcomp_strm *zcomp_strm_find(struct zcomp *comp)
 {
-	return comp->strm_find(comp);
+	return *get_cpu_ptr(comp->stream);
 }
 
 void zcomp_strm_release(struct zcomp *comp, struct zcomp_strm *zstrm)
 {
-	comp->strm_release(comp, zstrm);
+	put_cpu_ptr(comp->stream);
 }
 
 int zcomp_compress(struct zcomp *comp, struct zcomp_strm *zstrm,
@@ -327,9 +123,83 @@ int zcomp_decompress(struct zcomp *comp, const unsigned char *src,
 	return comp->backend->decompress(src, src_len, dst);
 }
 
+static int __zcomp_cpu_notifier(struct zcomp *comp,
+		unsigned long action, unsigned long cpu)
+{
+	struct zcomp_strm *zstrm;
+
+	switch (action) {
+	case CPU_UP_PREPARE:
+		if (WARN_ON(*per_cpu_ptr(comp->stream, cpu)))
+			break;
+		zstrm = zcomp_strm_alloc(comp, GFP_KERNEL);
+		if (IS_ERR_OR_NULL(zstrm)) {
+			pr_err("Can't allocate a compression stream\n");
+			return NOTIFY_BAD;
+		}
+		*per_cpu_ptr(comp->stream, cpu) = zstrm;
+		break;
+	case CPU_DEAD:
+	case CPU_UP_CANCELED:
+		zstrm = *per_cpu_ptr(comp->stream, cpu);
+		if (!IS_ERR_OR_NULL(zstrm))
+			zcomp_strm_free(comp, zstrm);
+		*per_cpu_ptr(comp->stream, cpu) = NULL;
+		break;
+	default:
+		break;
+	}
+	return NOTIFY_OK;
+}
+
+static int zcomp_cpu_notifier(struct notifier_block *nb,
+		unsigned long action, void *pcpu)
+{
+	unsigned long cpu = (unsigned long)pcpu;
+	struct zcomp *comp = container_of(nb, typeof(*comp), notifier);
+
+	return __zcomp_cpu_notifier(comp, action, cpu);
+}
+
+static int zcomp_init(struct zcomp *comp)
+{
+	unsigned long cpu;
+	int ret;
+
+	comp->notifier.notifier_call = zcomp_cpu_notifier;
+
+	comp->stream = alloc_percpu(struct zcomp_strm *);
+	if (!comp->stream)
+		return -ENOMEM;
+
+	cpu_notifier_register_begin();
+	for_each_online_cpu(cpu) {
+		ret = __zcomp_cpu_notifier(comp, CPU_UP_PREPARE, cpu);
+		if (ret == NOTIFY_BAD)
+			goto cleanup;
+	}
+	__register_cpu_notifier(&comp->notifier);
+	cpu_notifier_register_done();
+	return 0;
+
+cleanup:
+	for_each_online_cpu(cpu)
+		__zcomp_cpu_notifier(comp, CPU_UP_CANCELED, cpu);
+	cpu_notifier_register_done();
+	return -ENOMEM;
+}
+
 void zcomp_destroy(struct zcomp *comp)
 {
-	comp->destroy(comp);
+	unsigned long cpu;
+
+	cpu_notifier_register_begin();
+	for_each_online_cpu(cpu)
+		__zcomp_cpu_notifier(comp, CPU_UP_CANCELED, cpu);
+	__unregister_cpu_notifier(&comp->notifier);
+	cpu_notifier_register_done();
+
+	free_percpu(comp->stream);
 	kfree(comp);
 }
 
@@ -356,10 +226,7 @@ struct zcomp *zcomp_create(const char *compress, int max_strm)
 		return ERR_PTR(-ENOMEM);
 
 	comp->backend = backend;
-	if (max_strm > 1)
-		error = zcomp_strm_multi_create(comp, max_strm);
-	else
-		error = zcomp_strm_single_create(comp);
+	error = zcomp_init(comp);
 	if (error) {
 		kfree(comp);
 		return ERR_PTR(error);
diff --git a/drivers/block/zram/zcomp.h b/drivers/block/zram/zcomp.h
index b7d2a4b..aba8c21 100644
--- a/drivers/block/zram/zcomp.h
+++ b/drivers/block/zram/zcomp.h
@@ -10,8 +10,6 @@
 #ifndef _ZCOMP_H_
 #define _ZCOMP_H_
 
-#include <linux/mutex.h>
-
 struct zcomp_strm {
 	/* compression/decompression buffer */
 	void *buffer;
@@ -21,8 +19,6 @@ struct zcomp_strm {
 	 * working memory)
 	 */
 	void *private;
-	/* used in multi stream backend, protected by backend strm_lock */
-	struct list_head list;
 };
 
 /* static compression backend */
@@ -41,13 +37,9 @@ struct zcomp_backend {
 
 /* dynamic per-device compression frontend */
 struct zcomp {
-	void *stream;
+	struct zcomp_strm * __percpu *stream;
 	struct zcomp_backend *backend;
-
-	struct zcomp_strm *(*strm_find)(struct zcomp *comp);
-	void (*strm_release)(struct zcomp *comp, struct zcomp_strm *zstrm);
-	bool (*set_max_streams)(struct zcomp *comp, int num_strm);
-	void (*destroy)(struct zcomp *comp);
+	struct notifier_block notifier;
 };
 
 ssize_t zcomp_available_show(const char *comp, char *buf);
diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
index 9030992..c5e682c 100644
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c
@@ -650,7 +650,7 @@ static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index,
 {
 	int ret = 0;
 	size_t clen;
-	unsigned long handle;
+	unsigned long handle = 0;
 	struct page *page;
 	unsigned char *user_mem, *cmem, *src, *uncmem = NULL;
 	struct zram_meta *meta = zram->meta;
@@ -673,9 +673,8 @@ static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index,
 			goto out;
 	}
 
-	zstrm = zcomp_strm_find(zram->comp);
+compress_again:
 	user_mem = kmap_atomic(page);
-
 	if (is_partial_io(bvec)) {
 		memcpy(uncmem + offset, user_mem + bvec->bv_offset,
 		       bvec->bv_len);
@@ -699,6 +698,7 @@ static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index,
 		goto out;
 	}
 
+	zstrm = zcomp_strm_find(zram->comp);
 	ret = zcomp_compress(zram->comp, zstrm, uncmem, &clen);
 	if (!is_partial_io(bvec)) {
 		kunmap_atomic(user_mem);
@@ -710,6 +710,7 @@ static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index,
 		pr_err("Compression failed! err=%d\n", ret);
 		goto out;
 	}
+
 	src = zstrm->buffer;
 	if (unlikely(clen > max_zpage_size)) {
 		clen = PAGE_SIZE;
@@ -717,8 +718,31 @@ static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index,
 			src = uncmem;
 	}
 
-	handle = zs_malloc(meta->mem_pool, clen, GFP_NOIO | __GFP_HIGHMEM);
+	/*
+	 * handle allocation has 2 paths:
+	 * a) fast path is executed with preemption disabled (for
+	 *  per-cpu streams) and has __GFP_DIRECT_RECLAIM bit clear,
+	 *  since we can't sleep;
+	 * b) slow path enables preemption and attempts to alocate
+	 *  the page with __GFP_DIRECT_RECLAIM bit set. we have to
+	 *  put per-cpu compression stream and, thus, to re-do
+	 *  the compression once handle is allocated.
+	 *
+	 * if we have a 'non-null' handle here then we are coming
+	 * from the slow path and handle has already been allocated.
+	 */
+	if (!handle)
+		handle = zs_malloc(meta->mem_pool, clen,
+				__GFP_KSWAPD_RECLAIM |
+				__GFP_NOWARN |
+				__GFP_HIGHMEM);
 	if (!handle) {
+		zcomp_strm_release(zram->comp, zstrm);
+		handle = zs_malloc(meta->mem_pool, clen,
+				GFP_NOIO | __GFP_HIGHMEM);
+		if (handle)
+			goto compress_again;
+
 		pr_err("Error allocating memory for compressed page: %u, size=%zu\n",
 			index, clen);
 		ret = -ENOMEM;
-- 
2.8.0.rc0.1.gd285ab0