From: Srividya Desireddy <[email protected]>
Date: Sun, 2 Jul 2017 19:15:37 +0530
Subject: [PATCH v2] zswap: Zero-filled pages handling
Zswap is a cache which compresses the pages that are being swapped out
and stores them into a dynamically allocated RAM-based memory pool.
Experiments have shown that around 10-20% of pages stored in zswap
are zero-filled pages (i.e. contents of the page are all zeros), but
these pages are handled as normal pages by compressing and allocating
memory in the pool.
This patch adds a check in zswap_frontswap_store() to identify zero-filled
page before compression of the page. If the page is a zero-filled page, set
zswap_entry.zeroflag and skip the compression of the page and alloction
of memory in zpool. In zswap_frontswap_load(), check if the zeroflag is
set for the page in zswap_entry. If the flag is set, memset the page with
zero. This saves the decompression time during load.
On Ubuntu PC with 2GB RAM, while executing kernel build and other test
scripts ~15% of pages in zswap were zero pages. With multimedia workload
more than 20% of zswap pages were found to be zero pages.
On a ARM Quad Core 32-bit device with 1.5GB RAM an average 10% of zero
pages were found in zswap (an average of 5000 zero pages found out of
~50000 pages stored in zswap) on launching and relaunching 15 applications.
The launch time of the applications improved by ~3%.
Test Parameters Baseline With patch Improvement
-----------------------------------------------------------
Total RAM 1343MB 1343MB
Available RAM 451MB 445MB -6MB
Avg. Memfree 69MB 70MB 1MB
Avg. Swap Used 226MB 215MB -11MB
Avg. App entry time 644msec 623msec 3%
With patch, every page swapped to zswap is checked if it is a zero
page or not and for all the zero pages compression and memory allocation
operations are skipped. Overall there is an improvement of 30% in zswap
store time.
In case of non-zero pages there is no overhead during zswap page load. For
zero pages there is a improvement of more than 60% in the zswap load time
as the zero page decompression is avoided.
The below table shows the execution time profiling of the patch.
Zswap Store Operation Baseline With patch % Improvement
--------------------------------------------------------------
* Zero page check -- 22.5ms
(for non-zero pages)
* Zero page check -- 24ms
(for zero pages)
* Compression time 55ms --
(of zero pages)
* Allocation time 14ms --
(to store compressed
zero pages)
-------------------------------------------------------------
Total 69ms 46.5ms 32%
Zswap Load Operation Baseline With patch % Improvement
-------------------------------------------------------------
* Decompression time 30.4ms --
(of zero pages)
* Zero page check + -- 10.04ms
memset operation
(of zero pages)
-------------------------------------------------------------
Total 30.4ms 10.04ms 66%
*The execution times may vary with test device used.
Signed-off-by: Srividya Desireddy <[email protected]>
---
mm/zswap.c | 46 ++++++++++++++++++++++++++++++++++++++++++----
1 file changed, 42 insertions(+), 4 deletions(-)
diff --git a/mm/zswap.c b/mm/zswap.c
index eedc278..edc584b 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -49,6 +49,8 @@
static u64 zswap_pool_total_size;
/* The number of compressed pages currently stored in zswap */
static atomic_t zswap_stored_pages = ATOMIC_INIT(0);
+/* The number of zero filled pages swapped out to zswap */
+static atomic_t zswap_zero_pages = ATOMIC_INIT(0);
/*
* The statistics below are not protected from concurrent access for
@@ -145,7 +147,7 @@ struct zswap_pool {
* be held while changing the refcount. Since the lock must
* be held, there is no reason to also make refcount atomic.
* length - the length in bytes of the compressed page data. Needed during
- * decompression
+ * decompression. For a zero page length is 0.
* pool - the zswap_pool the entry's data is in
* handle - zpool allocation handle that stores the compressed page data
*/
@@ -320,8 +322,12 @@ static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
*/
static void zswap_free_entry(struct zswap_entry *entry)
{
- zpool_free(entry->pool->zpool, entry->handle);
- zswap_pool_put(entry->pool);
+ if (!entry->length)
+ atomic_dec(&zswap_zero_pages);
+ else {
+ zpool_free(entry->pool->zpool, entry->handle);
+ zswap_pool_put(entry->pool);
+ }
zswap_entry_cache_free(entry);
atomic_dec(&zswap_stored_pages);
zswap_update_total_size();
@@ -956,6 +962,19 @@ static int zswap_shrink(void)
return ret;
}
+static int zswap_is_page_zero_filled(void *ptr)
+{
+ unsigned int pos;
+ unsigned long *page;
+
+ page = (unsigned long *)ptr;
+ for (pos = 0; pos != PAGE_SIZE / sizeof(*page); pos++) {
+ if (page[pos])
+ return 0;
+ }
+ return 1;
+}
+
/*********************************
* frontswap hooks
**********************************/
@@ -996,6 +1015,15 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
goto reject;
}
+ src = kmap_atomic(page);
+ if (zswap_is_page_zero_filled(src)) {
+ kunmap_atomic(src);
+ entry->offset = offset;
+ entry->length = 0;
+ atomic_inc(&zswap_zero_pages);
+ goto insert_entry;
+ }
+
/* if entry is successfully added, it keeps the reference */
entry->pool = zswap_pool_current_get();
if (!entry->pool) {
@@ -1006,7 +1034,6 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
/* compress */
dst = get_cpu_var(zswap_dstmem);
tfm = *get_cpu_ptr(entry->pool->tfm);
- src = kmap_atomic(page);
ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
kunmap_atomic(src);
put_cpu_ptr(entry->pool->tfm);
@@ -1040,6 +1067,7 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
entry->handle = handle;
entry->length = dlen;
+insert_entry:
/* map */
spin_lock(&tree->lock);
do {
@@ -1092,6 +1120,13 @@ static int zswap_frontswap_load(unsigned type, pgoff_t offset,
}
spin_unlock(&tree->lock);
+ if (!entry->length) {
+ dst = kmap_atomic(page);
+ memset(dst, 0, PAGE_SIZE);
+ kunmap_atomic(dst);
+ goto freeentry;
+ }
+
/* decompress */
dlen = PAGE_SIZE;
src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
@@ -1104,6 +1139,7 @@ static int zswap_frontswap_load(unsigned type, pgoff_t offset,
zpool_unmap_handle(entry->pool->zpool, entry->handle);
BUG_ON(ret);
+freeentry:
spin_lock(&tree->lock);
zswap_entry_put(tree, entry);
spin_unlock(&tree->lock);
@@ -1212,6 +1248,8 @@ static int __init zswap_debugfs_init(void)
zswap_debugfs_root, &zswap_pool_total_size);
debugfs_create_atomic_t("stored_pages", S_IRUGO,
zswap_debugfs_root, &zswap_stored_pages);
+ debugfs_create_atomic_t("zero_pages", 0444,
+ zswap_debugfs_root, &zswap_zero_pages);
return 0;
}
--
1.7.9.5
On Sun, Jul 2, 2017 at 9:19 AM, Srividya Desireddy
<[email protected]> wrote:
> From: Srividya Desireddy <[email protected]>
> Date: Sun, 2 Jul 2017 19:15:37 +0530
> Subject: [PATCH v2] zswap: Zero-filled pages handling
>
> Zswap is a cache which compresses the pages that are being swapped out
> and stores them into a dynamically allocated RAM-based memory pool.
> Experiments have shown that around 10-20% of pages stored in zswap
> are zero-filled pages (i.e. contents of the page are all zeros), but
> these pages are handled as normal pages by compressing and allocating
> memory in the pool.
I am somewhat surprised that this many anon pages are zero filled.
If this is true, then maybe we should consider solving this at the
swap level in general, as we can de-dup zero pages in all swap
devices, not just zswap.
That being said, this is a fair small change and I don't see anything
objectionable. However, I do think the better solution would be to do
this at a higher level.
Thanks,
Seth
>
> This patch adds a check in zswap_frontswap_store() to identify zero-filled
> page before compression of the page. If the page is a zero-filled page, set
> zswap_entry.zeroflag and skip the compression of the page and alloction
> of memory in zpool. In zswap_frontswap_load(), check if the zeroflag is
> set for the page in zswap_entry. If the flag is set, memset the page with
> zero. This saves the decompression time during load.
>
> On Ubuntu PC with 2GB RAM, while executing kernel build and other test
> scripts ~15% of pages in zswap were zero pages. With multimedia workload
> more than 20% of zswap pages were found to be zero pages.
>
> On a ARM Quad Core 32-bit device with 1.5GB RAM an average 10% of zero
> pages were found in zswap (an average of 5000 zero pages found out of
> ~50000 pages stored in zswap) on launching and relaunching 15 applications.
> The launch time of the applications improved by ~3%.
>
> Test Parameters Baseline With patch Improvement
> -----------------------------------------------------------
> Total RAM 1343MB 1343MB
> Available RAM 451MB 445MB -6MB
> Avg. Memfree 69MB 70MB 1MB
> Avg. Swap Used 226MB 215MB -11MB
> Avg. App entry time 644msec 623msec 3%
>
> With patch, every page swapped to zswap is checked if it is a zero
> page or not and for all the zero pages compression and memory allocation
> operations are skipped. Overall there is an improvement of 30% in zswap
> store time.
>
> In case of non-zero pages there is no overhead during zswap page load. For
> zero pages there is a improvement of more than 60% in the zswap load time
> as the zero page decompression is avoided.
> The below table shows the execution time profiling of the patch.
>
> Zswap Store Operation Baseline With patch % Improvement
> --------------------------------------------------------------
> * Zero page check -- 22.5ms
> (for non-zero pages)
> * Zero page check -- 24ms
> (for zero pages)
> * Compression time 55ms --
> (of zero pages)
> * Allocation time 14ms --
> (to store compressed
> zero pages)
> -------------------------------------------------------------
> Total 69ms 46.5ms 32%
>
> Zswap Load Operation Baseline With patch % Improvement
> -------------------------------------------------------------
> * Decompression time 30.4ms --
> (of zero pages)
> * Zero page check + -- 10.04ms
> memset operation
> (of zero pages)
> -------------------------------------------------------------
> Total 30.4ms 10.04ms 66%
>
> *The execution times may vary with test device used.
>
> Signed-off-by: Srividya Desireddy <[email protected]>
> ---
> mm/zswap.c | 46 ++++++++++++++++++++++++++++++++++++++++++----
> 1 file changed, 42 insertions(+), 4 deletions(-)
>
> diff --git a/mm/zswap.c b/mm/zswap.c
> index eedc278..edc584b 100644
> --- a/mm/zswap.c
> +++ b/mm/zswap.c
> @@ -49,6 +49,8 @@
> static u64 zswap_pool_total_size;
> /* The number of compressed pages currently stored in zswap */
> static atomic_t zswap_stored_pages = ATOMIC_INIT(0);
> +/* The number of zero filled pages swapped out to zswap */
> +static atomic_t zswap_zero_pages = ATOMIC_INIT(0);
>
> /*
> * The statistics below are not protected from concurrent access for
> @@ -145,7 +147,7 @@ struct zswap_pool {
> * be held while changing the refcount. Since the lock must
> * be held, there is no reason to also make refcount atomic.
> * length - the length in bytes of the compressed page data. Needed during
> - * decompression
> + * decompression. For a zero page length is 0.
> * pool - the zswap_pool the entry's data is in
> * handle - zpool allocation handle that stores the compressed page data
> */
> @@ -320,8 +322,12 @@ static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
> */
> static void zswap_free_entry(struct zswap_entry *entry)
> {
> - zpool_free(entry->pool->zpool, entry->handle);
> - zswap_pool_put(entry->pool);
> + if (!entry->length)
> + atomic_dec(&zswap_zero_pages);
> + else {
> + zpool_free(entry->pool->zpool, entry->handle);
> + zswap_pool_put(entry->pool);
> + }
> zswap_entry_cache_free(entry);
> atomic_dec(&zswap_stored_pages);
> zswap_update_total_size();
> @@ -956,6 +962,19 @@ static int zswap_shrink(void)
> return ret;
> }
>
> +static int zswap_is_page_zero_filled(void *ptr)
> +{
> + unsigned int pos;
> + unsigned long *page;
> +
> + page = (unsigned long *)ptr;
> + for (pos = 0; pos != PAGE_SIZE / sizeof(*page); pos++) {
> + if (page[pos])
> + return 0;
> + }
> + return 1;
> +}
> +
> /*********************************
> * frontswap hooks
> **********************************/
> @@ -996,6 +1015,15 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
> goto reject;
> }
>
> + src = kmap_atomic(page);
> + if (zswap_is_page_zero_filled(src)) {
> + kunmap_atomic(src);
> + entry->offset = offset;
> + entry->length = 0;
> + atomic_inc(&zswap_zero_pages);
> + goto insert_entry;
> + }
> +
> /* if entry is successfully added, it keeps the reference */
> entry->pool = zswap_pool_current_get();
> if (!entry->pool) {
> @@ -1006,7 +1034,6 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
> /* compress */
> dst = get_cpu_var(zswap_dstmem);
> tfm = *get_cpu_ptr(entry->pool->tfm);
> - src = kmap_atomic(page);
> ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
> kunmap_atomic(src);
> put_cpu_ptr(entry->pool->tfm);
> @@ -1040,6 +1067,7 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
> entry->handle = handle;
> entry->length = dlen;
>
> +insert_entry:
> /* map */
> spin_lock(&tree->lock);
> do {
> @@ -1092,6 +1120,13 @@ static int zswap_frontswap_load(unsigned type, pgoff_t offset,
> }
> spin_unlock(&tree->lock);
>
> + if (!entry->length) {
> + dst = kmap_atomic(page);
> + memset(dst, 0, PAGE_SIZE);
> + kunmap_atomic(dst);
> + goto freeentry;
> + }
> +
> /* decompress */
> dlen = PAGE_SIZE;
> src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
> @@ -1104,6 +1139,7 @@ static int zswap_frontswap_load(unsigned type, pgoff_t offset,
> zpool_unmap_handle(entry->pool->zpool, entry->handle);
> BUG_ON(ret);
>
> +freeentry:
> spin_lock(&tree->lock);
> zswap_entry_put(tree, entry);
> spin_unlock(&tree->lock);
> @@ -1212,6 +1248,8 @@ static int __init zswap_debugfs_init(void)
> zswap_debugfs_root, &zswap_pool_total_size);
> debugfs_create_atomic_t("stored_pages", S_IRUGO,
> zswap_debugfs_root, &zswap_stored_pages);
> + debugfs_create_atomic_t("zero_pages", 0444,
> + zswap_debugfs_root, &zswap_zero_pages);
>
> return 0;
> }
> --
> 1.7.9.5
On (07/02/17 20:28), Seth Jennings wrote:
> On Sun, Jul 2, 2017 at 9:19 AM, Srividya Desireddy
> > Zswap is a cache which compresses the pages that are being swapped out
> > and stores them into a dynamically allocated RAM-based memory pool.
> > Experiments have shown that around 10-20% of pages stored in zswap
> > are zero-filled pages (i.e. contents of the page are all zeros), but
> > these pages are handled as normal pages by compressing and allocating
> > memory in the pool.
>
> I am somewhat surprised that this many anon pages are zero filled.
>
> If this is true, then maybe we should consider solving this at the
> swap level in general, as we can de-dup zero pages in all swap
> devices, not just zswap.
>
> That being said, this is a fair small change and I don't see anything
> objectionable. However, I do think the better solution would be to do
> this at a higher level.
zero-filled pages are just 1 case. in general, it's better
to handle pages that are memset-ed with the same value (e.g.
memset(page, 0x01, page_size)). which includes, but not
limited to, 0x00. zram does it.
-ss
On Wed, Jul 6, 2017 at 10:49 AM, Sergey Senozhatsky wrote:
> On (07/02/17 20:28), Seth Jennings wrote:
>> On Sun, Jul 2, 2017 at 9:19 AM, Srividya Desireddy
>> > Zswap is a cache which compresses the pages that are being swapped out
>> > and stores them into a dynamically allocated RAM-based memory pool.
>> > Experiments have shown that around 10-20% of pages stored in zswap
>> > are zero-filled pages (i.e. contents of the page are all zeros), but
>> > these pages are handled as normal pages by compressing and allocating
>> > memory in the pool.
>>
>> I am somewhat surprised that this many anon pages are zero filled.
>>
>> If this is true, then maybe we should consider solving this at the
>> swap level in general, as we can de-dup zero pages in all swap
>> devices, not just zswap.
>>
>> That being said, this is a fair small change and I don't see anything
>> objectionable. However, I do think the better solution would be to do
> this at a higher level.
>
Thank you for your suggestion. It is a better solution to handle
zero-filled pages before swapping-out to zswap. Since, Zram is already
handles Zero pages internally, I considered to handle within Zswap.
In a long run, we can work on it to commonly handle zero-filled anon
pages.
> zero-filled pages are just 1 case. in general, it's better
> to handle pages that are memset-ed with the same value (e.g.
> memset(page, 0x01, page_size)). which includes, but not
> limited to, 0x00. zram does it.
>
> -ss
It is a good solution to extend zero-filled pages handling to same value
pages. I will work on to identify the percentage of same value pages
excluding zero-filled pages in Zswap and will get back.
- Srividya
On Thu, Jul 6, 2017 at 5:29 AM, Srividya Desireddy
<[email protected]> wrote:
> On Wed, Jul 6, 2017 at 10:49 AM, Sergey Senozhatsky wrote:
>> On (07/02/17 20:28), Seth Jennings wrote:
>>> On Sun, Jul 2, 2017 at 9:19 AM, Srividya Desireddy
>>> > Zswap is a cache which compresses the pages that are being swapped out
>>> > and stores them into a dynamically allocated RAM-based memory pool.
>>> > Experiments have shown that around 10-20% of pages stored in zswap
>>> > are zero-filled pages (i.e. contents of the page are all zeros), but
>>> > these pages are handled as normal pages by compressing and allocating
>>> > memory in the pool.
>>>
>>> I am somewhat surprised that this many anon pages are zero filled.
>>>
>>> If this is true, then maybe we should consider solving this at the
>>> swap level in general, as we can de-dup zero pages in all swap
>>> devices, not just zswap.
>>>
>>> That being said, this is a fair small change and I don't see anything
>>> objectionable. However, I do think the better solution would be to do
>> this at a higher level.
>>
>
> Thank you for your suggestion. It is a better solution to handle
> zero-filled pages before swapping-out to zswap. Since, Zram is already
> handles Zero pages internally, I considered to handle within Zswap.
> In a long run, we can work on it to commonly handle zero-filled anon
> pages.
>
>> zero-filled pages are just 1 case. in general, it's better
>> to handle pages that are memset-ed with the same value (e.g.
>> memset(page, 0x01, page_size)). which includes, but not
>> limited to, 0x00. zram does it.
>>
>> -ss
>
> It is a good solution to extend zero-filled pages handling to same value
> pages. I will work on to identify the percentage of same value pages
> excluding zero-filled pages in Zswap and will get back.
Yes, this sounds like a good modification to the patch. Also, unless
anyone else disagrees, it may be good to control this with a module
param - in case anyone has a use case that they know won't be helped
by this, and the extra overhead of checking each page is wasteful.
Probably should default to enabled.
>
> - Srividya