Re: [PATCH v20 3/4] qcow2: add zstd cluster compression

[Max Reitz]

On 28.04.20 09:23, Denis Plotnikov wrote:
> 
> 
> On 28.04.2020 09:16, Max Reitz wrote:
>> On 27.04.20 21:26, Denis Plotnikov wrote:
>>>
>>> On 27.04.2020 15:35, Max Reitz wrote:
 On 21.04.20 10:11, Denis Plotnikov wrote:
> zstd significantly reduces cluster compression time.
> It provides better compression performance maintaining
> the same level of the compression ratio in comparison with
> zlib, which, at the moment, is the only compression
> method available.
>
> The performance test results:
> Test compresses and decompresses qemu qcow2 image with just
> installed rhel-7.6 guest.
> Image cluster size: 64K. Image on disk size: 2.2G
>
> The test was conducted with brd disk to reduce the influence
> of disk subsystem to the test results.
> The results is given in seconds.
>
> compress cmd:
>     time ./qemu-img convert -O qcow2 -c -o
> compression_type=[zlib|zstd]
>     src.img [zlib|zstd]_compressed.img
> decompress cmd
>     time ./qemu-img convert -O qcow2
>     [zlib|zstd]_compressed.img uncompressed.img
>
>  compression   decompression
>    zlib   zstd   zlib zstd
> 
> real 65.5   16.3 (-75 %)    1.9  1.6 (-16 %)
> user 65.0   15.8    5.3  2.5
> sys   3.3    0.2    2.0  2.0
>
> Both ZLIB and ZSTD gave the same compression ratio: 1.57
> compressed image size in both cases: 1.4G
>
> Signed-off-by: Denis Plotnikov 
> Reviewed-by: Vladimir Sementsov-Ogievskiy 
> Reviewed-by: Alberto Garcia 
> QAPI part:
> Acked-by: Markus Armbruster 
> ---
>    docs/interop/qcow2.txt |   1 +
>    configure  |   2 +-
>    qapi/block-core.json   |   3 +-
>    block/qcow2-threads.c  | 157
> +
>    block/qcow2.c  |   7 ++
>    5 files changed, 168 insertions(+), 2 deletions(-)
 [...]

> diff --git a/block/qcow2-threads.c b/block/qcow2-threads.c
> index 7dbaf53489..0525718704 100644
> --- a/block/qcow2-threads.c
> +++ b/block/qcow2-threads.c
> @@ -28,6 +28,11 @@
>    #define ZLIB_CONST
>    #include 
>    +#ifdef CONFIG_ZSTD
> +#include 
> +#include 
> +#endif
> +
>    #include "qcow2.h"
>    #include "block/thread-pool.h"
>    #include "crypto.h"
> @@ -166,6 +171,148 @@ static ssize_t qcow2_zlib_decompress(void
> *dest, size_t dest_size,
>    return ret;
>    }
>    +#ifdef CONFIG_ZSTD
> +
> +/*
> + * qcow2_zstd_compress()
> + *
> + * Compress @src_size bytes of data using zstd compression method
> + *
> + * @dest - destination buffer, @dest_size bytes
> + * @src - source buffer, @src_size bytes
> + *
> + * Returns: compressed size on success
> + *  -ENOMEM destination buffer is not enough to store
> compressed data
> + *  -EIO    on any other error
> + */
> +static ssize_t qcow2_zstd_compress(void *dest, size_t dest_size,
> +   const void *src, size_t src_size)
> +{
> +    ssize_t ret;
> +    ZSTD_outBuffer output = { dest, dest_size, 0 };
> +    ZSTD_inBuffer input = { src, src_size, 0 };
 Minor style note: I think it’d be nicer to use designated initializers
 here.

> +    ZSTD_CCtx *cctx = ZSTD_createCCtx();
> +
> +    if (!cctx) {
> +    return -EIO;
> +    }
> +    /*
> + * Use the zstd streamed interface for symmetry with
> decompression,
> + * where streaming is essential since we don't record the exact
> + * compressed size.
> + *
> + * In the loop, we try to compress all the data into one zstd
> frame.
> + * ZSTD_compressStream2 potentially can finish a frame earlier
> + * than the full input data is consumed. That's why we are
> looping
> + * until all the input data is consumed.
> + */
> +    while (input.pos < input.size) {
> +    size_t zstd_ret;
> +    /*
> + * ZSTD spec: "You must continue calling
> ZSTD_compressStream2()
> + * with ZSTD_e_end until it returns 0, at which point you are
> + * free to start a new frame". We assume that "start a new
> frame"
> + * means call ZSTD_compressStream2 in the very beginning or
> when
> + * ZSTD_compressStream2 has returned with 0.
> + */
> +    do {
> +    zstd_ret = ZSTD_compressStream2(cctx, , ,
> ZSTD_e_end);
 The spec makes it sound to me like ZSTD_e_end will always complete in a
 single call if there’s enough space in the output buffer.  So 

Re: [PATCH v20 3/4] qcow2: add zstd cluster compression

[Denis Plotnikov]

On 28.04.2020 09:16, Max Reitz wrote:

On 27.04.20 21:26, Denis Plotnikov wrote:


On 27.04.2020 15:35, Max Reitz wrote:

On 21.04.20 10:11, Denis Plotnikov wrote:

zstd significantly reduces cluster compression time.
It provides better compression performance maintaining
the same level of the compression ratio in comparison with
zlib, which, at the moment, is the only compression
method available.

The performance test results:
Test compresses and decompresses qemu qcow2 image with just
installed rhel-7.6 guest.
Image cluster size: 64K. Image on disk size: 2.2G

The test was conducted with brd disk to reduce the influence
of disk subsystem to the test results.
The results is given in seconds.

compress cmd:
    time ./qemu-img convert -O qcow2 -c -o compression_type=[zlib|zstd]
    src.img [zlib|zstd]_compressed.img
decompress cmd
    time ./qemu-img convert -O qcow2
    [zlib|zstd]_compressed.img uncompressed.img

     compression   decompression
   zlib   zstd   zlib zstd

real 65.5   16.3 (-75 %)    1.9  1.6 (-16 %)
user 65.0   15.8    5.3  2.5
sys   3.3    0.2    2.0  2.0

Both ZLIB and ZSTD gave the same compression ratio: 1.57
compressed image size in both cases: 1.4G

Signed-off-by: Denis Plotnikov 
Reviewed-by: Vladimir Sementsov-Ogievskiy 
Reviewed-by: Alberto Garcia 
QAPI part:
Acked-by: Markus Armbruster 
---
   docs/interop/qcow2.txt |   1 +
   configure  |   2 +-
   qapi/block-core.json   |   3 +-
   block/qcow2-threads.c  | 157 +
   block/qcow2.c  |   7 ++
   5 files changed, 168 insertions(+), 2 deletions(-)

[...]


diff --git a/block/qcow2-threads.c b/block/qcow2-threads.c
index 7dbaf53489..0525718704 100644
--- a/block/qcow2-threads.c
+++ b/block/qcow2-threads.c
@@ -28,6 +28,11 @@
   #define ZLIB_CONST
   #include 
   +#ifdef CONFIG_ZSTD
+#include 
+#include 
+#endif
+
   #include "qcow2.h"
   #include "block/thread-pool.h"
   #include "crypto.h"
@@ -166,6 +171,148 @@ static ssize_t qcow2_zlib_decompress(void
*dest, size_t dest_size,
   return ret;
   }
   +#ifdef CONFIG_ZSTD
+
+/*
+ * qcow2_zstd_compress()
+ *
+ * Compress @src_size bytes of data using zstd compression method
+ *
+ * @dest - destination buffer, @dest_size bytes
+ * @src - source buffer, @src_size bytes
+ *
+ * Returns: compressed size on success
+ *  -ENOMEM destination buffer is not enough to store
compressed data
+ *  -EIO    on any other error
+ */
+static ssize_t qcow2_zstd_compress(void *dest, size_t dest_size,
+   const void *src, size_t src_size)
+{
+    ssize_t ret;
+    ZSTD_outBuffer output = { dest, dest_size, 0 };
+    ZSTD_inBuffer input = { src, src_size, 0 };

Minor style note: I think it’d be nicer to use designated initializers
here.


+    ZSTD_CCtx *cctx = ZSTD_createCCtx();
+
+    if (!cctx) {
+    return -EIO;
+    }
+    /*
+ * Use the zstd streamed interface for symmetry with decompression,
+ * where streaming is essential since we don't record the exact
+ * compressed size.
+ *
+ * In the loop, we try to compress all the data into one zstd
frame.
+ * ZSTD_compressStream2 potentially can finish a frame earlier
+ * than the full input data is consumed. That's why we are looping
+ * until all the input data is consumed.
+ */
+    while (input.pos < input.size) {
+    size_t zstd_ret;
+    /*
+ * ZSTD spec: "You must continue calling ZSTD_compressStream2()
+ * with ZSTD_e_end until it returns 0, at which point you are
+ * free to start a new frame". We assume that "start a new
frame"
+ * means call ZSTD_compressStream2 in the very beginning or
when
+ * ZSTD_compressStream2 has returned with 0.
+ */
+    do {
+    zstd_ret = ZSTD_compressStream2(cctx, , ,
ZSTD_e_end);

The spec makes it sound to me like ZSTD_e_end will always complete in a
single call if there’s enough space in the output buffer.  So the only
team we have to loop would be when there isn’t enough space anyway:

It says this about ZSTD_e_end:

flush operation is the same, and follows same rules as calling
ZSTD_compressStream2() with ZSTD_e_flush.

Those rules being:

Note that, if `output->size` is too small, a single invocation with
ZSTD_e_flush might not be enough (return code > 0).

So it seems like it will only return a value > 0 if the output buffer is
definitely too small.

The spec also notes that the return value is greater than 0 if:

0 if some data still present within internal buffer (the value is

minimal estimation of remaining size),

So it’s a minimum estimate.  That’s another point that heavily implies
to me that if the return value were less than what’s left in the buffer,
the function wouldn’t 

Re: [PATCH v20 3/4] qcow2: add zstd cluster compression

[Max Reitz]

On 27.04.20 21:26, Denis Plotnikov wrote:
> 
> 
> On 27.04.2020 15:35, Max Reitz wrote:
>> On 21.04.20 10:11, Denis Plotnikov wrote:
>>> zstd significantly reduces cluster compression time.
>>> It provides better compression performance maintaining
>>> the same level of the compression ratio in comparison with
>>> zlib, which, at the moment, is the only compression
>>> method available.
>>>
>>> The performance test results:
>>> Test compresses and decompresses qemu qcow2 image with just
>>> installed rhel-7.6 guest.
>>> Image cluster size: 64K. Image on disk size: 2.2G
>>>
>>> The test was conducted with brd disk to reduce the influence
>>> of disk subsystem to the test results.
>>> The results is given in seconds.
>>>
>>> compress cmd:
>>>    time ./qemu-img convert -O qcow2 -c -o compression_type=[zlib|zstd]
>>>    src.img [zlib|zstd]_compressed.img
>>> decompress cmd
>>>    time ./qemu-img convert -O qcow2
>>>    [zlib|zstd]_compressed.img uncompressed.img
>>>
>>>     compression   decompression
>>>   zlib   zstd   zlib zstd
>>> 
>>> real 65.5   16.3 (-75 %)    1.9  1.6 (-16 %)
>>> user 65.0   15.8    5.3  2.5
>>> sys   3.3    0.2    2.0  2.0
>>>
>>> Both ZLIB and ZSTD gave the same compression ratio: 1.57
>>> compressed image size in both cases: 1.4G
>>>
>>> Signed-off-by: Denis Plotnikov 
>>> Reviewed-by: Vladimir Sementsov-Ogievskiy 
>>> Reviewed-by: Alberto Garcia 
>>> QAPI part:
>>> Acked-by: Markus Armbruster 
>>> ---
>>>   docs/interop/qcow2.txt |   1 +
>>>   configure  |   2 +-
>>>   qapi/block-core.json   |   3 +-
>>>   block/qcow2-threads.c  | 157 +
>>>   block/qcow2.c  |   7 ++
>>>   5 files changed, 168 insertions(+), 2 deletions(-)
>> [...]
>>
>>> diff --git a/block/qcow2-threads.c b/block/qcow2-threads.c
>>> index 7dbaf53489..0525718704 100644
>>> --- a/block/qcow2-threads.c
>>> +++ b/block/qcow2-threads.c
>>> @@ -28,6 +28,11 @@
>>>   #define ZLIB_CONST
>>>   #include 
>>>   +#ifdef CONFIG_ZSTD
>>> +#include 
>>> +#include 
>>> +#endif
>>> +
>>>   #include "qcow2.h"
>>>   #include "block/thread-pool.h"
>>>   #include "crypto.h"
>>> @@ -166,6 +171,148 @@ static ssize_t qcow2_zlib_decompress(void
>>> *dest, size_t dest_size,
>>>   return ret;
>>>   }
>>>   +#ifdef CONFIG_ZSTD
>>> +
>>> +/*
>>> + * qcow2_zstd_compress()
>>> + *
>>> + * Compress @src_size bytes of data using zstd compression method
>>> + *
>>> + * @dest - destination buffer, @dest_size bytes
>>> + * @src - source buffer, @src_size bytes
>>> + *
>>> + * Returns: compressed size on success
>>> + *  -ENOMEM destination buffer is not enough to store
>>> compressed data
>>> + *  -EIO    on any other error
>>> + */
>>> +static ssize_t qcow2_zstd_compress(void *dest, size_t dest_size,
>>> +   const void *src, size_t src_size)
>>> +{
>>> +    ssize_t ret;
>>> +    ZSTD_outBuffer output = { dest, dest_size, 0 };
>>> +    ZSTD_inBuffer input = { src, src_size, 0 };
>> Minor style note: I think it’d be nicer to use designated initializers
>> here.
>>
>>> +    ZSTD_CCtx *cctx = ZSTD_createCCtx();
>>> +
>>> +    if (!cctx) {
>>> +    return -EIO;
>>> +    }
>>> +    /*
>>> + * Use the zstd streamed interface for symmetry with decompression,
>>> + * where streaming is essential since we don't record the exact
>>> + * compressed size.
>>> + *
>>> + * In the loop, we try to compress all the data into one zstd
>>> frame.
>>> + * ZSTD_compressStream2 potentially can finish a frame earlier
>>> + * than the full input data is consumed. That's why we are looping
>>> + * until all the input data is consumed.
>>> + */
>>> +    while (input.pos < input.size) {
>>> +    size_t zstd_ret;
>>> +    /*
>>> + * ZSTD spec: "You must continue calling ZSTD_compressStream2()
>>> + * with ZSTD_e_end until it returns 0, at which point you are
>>> + * free to start a new frame". We assume that "start a new
>>> frame"
>>> + * means call ZSTD_compressStream2 in the very beginning or
>>> when
>>> + * ZSTD_compressStream2 has returned with 0.
>>> + */
>>> +    do {
>>> +    zstd_ret = ZSTD_compressStream2(cctx, , ,
>>> ZSTD_e_end);
>> The spec makes it sound to me like ZSTD_e_end will always complete in a
>> single call if there’s enough space in the output buffer.  So the only
>> team we have to loop would be when there isn’t enough space anyway:
>>
>> It says this about ZSTD_e_end:
>>> flush operation is the same, and follows same rules as calling
>>> ZSTD_compressStream2() with ZSTD_e_flush.
>> Those rules being:
>>> Note that, if `output->size` is too small, a single invocation with
>>> ZSTD_e_flush might not be enough (return code > 0).
>> So it 

Re: [PATCH v20 3/4] qcow2: add zstd cluster compression

[Denis Plotnikov]

On 27.04.2020 15:35, Max Reitz wrote:

On 21.04.20 10:11, Denis Plotnikov wrote:

zstd significantly reduces cluster compression time.
It provides better compression performance maintaining
the same level of the compression ratio in comparison with
zlib, which, at the moment, is the only compression
method available.

The performance test results:
Test compresses and decompresses qemu qcow2 image with just
installed rhel-7.6 guest.
Image cluster size: 64K. Image on disk size: 2.2G

The test was conducted with brd disk to reduce the influence
of disk subsystem to the test results.
The results is given in seconds.

compress cmd:
   time ./qemu-img convert -O qcow2 -c -o compression_type=[zlib|zstd]
   src.img [zlib|zstd]_compressed.img
decompress cmd
   time ./qemu-img convert -O qcow2
   [zlib|zstd]_compressed.img uncompressed.img

compression   decompression
  zlib   zstd   zlib zstd

real 65.5   16.3 (-75 %)1.9  1.6 (-16 %)
user 65.0   15.85.3  2.5
sys   3.30.22.0  2.0

Both ZLIB and ZSTD gave the same compression ratio: 1.57
compressed image size in both cases: 1.4G

Signed-off-by: Denis Plotnikov 
Reviewed-by: Vladimir Sementsov-Ogievskiy 
Reviewed-by: Alberto Garcia 
QAPI part:
Acked-by: Markus Armbruster 
---
  docs/interop/qcow2.txt |   1 +
  configure  |   2 +-
  qapi/block-core.json   |   3 +-
  block/qcow2-threads.c  | 157 +
  block/qcow2.c  |   7 ++
  5 files changed, 168 insertions(+), 2 deletions(-)

[...]


diff --git a/block/qcow2-threads.c b/block/qcow2-threads.c
index 7dbaf53489..0525718704 100644
--- a/block/qcow2-threads.c
+++ b/block/qcow2-threads.c
@@ -28,6 +28,11 @@
  #define ZLIB_CONST
  #include 
  
+#ifdef CONFIG_ZSTD

+#include 
+#include 
+#endif
+
  #include "qcow2.h"
  #include "block/thread-pool.h"
  #include "crypto.h"
@@ -166,6 +171,148 @@ static ssize_t qcow2_zlib_decompress(void *dest, size_t 
dest_size,
  return ret;
  }
  
+#ifdef CONFIG_ZSTD

+
+/*
+ * qcow2_zstd_compress()
+ *
+ * Compress @src_size bytes of data using zstd compression method
+ *
+ * @dest - destination buffer, @dest_size bytes
+ * @src - source buffer, @src_size bytes
+ *
+ * Returns: compressed size on success
+ *  -ENOMEM destination buffer is not enough to store compressed data
+ *  -EIOon any other error
+ */
+static ssize_t qcow2_zstd_compress(void *dest, size_t dest_size,
+   const void *src, size_t src_size)
+{
+ssize_t ret;
+ZSTD_outBuffer output = { dest, dest_size, 0 };
+ZSTD_inBuffer input = { src, src_size, 0 };

Minor style note: I think it’d be nicer to use designated initializers here.


+ZSTD_CCtx *cctx = ZSTD_createCCtx();
+
+if (!cctx) {
+return -EIO;
+}
+/*
+ * Use the zstd streamed interface for symmetry with decompression,
+ * where streaming is essential since we don't record the exact
+ * compressed size.
+ *
+ * In the loop, we try to compress all the data into one zstd frame.
+ * ZSTD_compressStream2 potentially can finish a frame earlier
+ * than the full input data is consumed. That's why we are looping
+ * until all the input data is consumed.
+ */
+while (input.pos < input.size) {
+size_t zstd_ret;
+/*
+ * ZSTD spec: "You must continue calling ZSTD_compressStream2()
+ * with ZSTD_e_end until it returns 0, at which point you are
+ * free to start a new frame". We assume that "start a new frame"
+ * means call ZSTD_compressStream2 in the very beginning or when
+ * ZSTD_compressStream2 has returned with 0.
+ */
+do {
+zstd_ret = ZSTD_compressStream2(cctx, , , ZSTD_e_end);

The spec makes it sound to me like ZSTD_e_end will always complete in a
single call if there’s enough space in the output buffer.  So the only
team we have to loop would be when there isn’t enough space anyway:

It says this about ZSTD_e_end:

flush operation is the same, and follows same rules as calling
ZSTD_compressStream2() with ZSTD_e_flush.

Those rules being:

Note that, if `output->size` is too small, a single invocation with
ZSTD_e_flush might not be enough (return code > 0).

So it seems like it will only return a value > 0 if the output buffer is
definitely too small.

The spec also notes that the return value is greater than 0 if:

0 if some data still present within internal buffer (the value is

minimal estimation of remaining size),

So it’s a minimum estimate.  That’s another point that heavily implies
to me that if the return value were less than what’s left in the buffer,
the function wouldn’t return but still try to write it out, until it
realizes that there isn’t enough space in the output 

Re: [PATCH v20 3/4] qcow2: add zstd cluster compression

[Max Reitz]

On 21.04.20 10:11, Denis Plotnikov wrote:
> zstd significantly reduces cluster compression time.
> It provides better compression performance maintaining
> the same level of the compression ratio in comparison with
> zlib, which, at the moment, is the only compression
> method available.
> 
> The performance test results:
> Test compresses and decompresses qemu qcow2 image with just
> installed rhel-7.6 guest.
> Image cluster size: 64K. Image on disk size: 2.2G
> 
> The test was conducted with brd disk to reduce the influence
> of disk subsystem to the test results.
> The results is given in seconds.
> 
> compress cmd:
>   time ./qemu-img convert -O qcow2 -c -o compression_type=[zlib|zstd]
>   src.img [zlib|zstd]_compressed.img
> decompress cmd
>   time ./qemu-img convert -O qcow2
>   [zlib|zstd]_compressed.img uncompressed.img
> 
>compression   decompression
>  zlib   zstd   zlib zstd
> 
> real 65.5   16.3 (-75 %)1.9  1.6 (-16 %)
> user 65.0   15.85.3  2.5
> sys   3.30.22.0  2.0
> 
> Both ZLIB and ZSTD gave the same compression ratio: 1.57
> compressed image size in both cases: 1.4G
> 
> Signed-off-by: Denis Plotnikov 
> Reviewed-by: Vladimir Sementsov-Ogievskiy 
> Reviewed-by: Alberto Garcia 
> QAPI part:
> Acked-by: Markus Armbruster 
> ---
>  docs/interop/qcow2.txt |   1 +
>  configure  |   2 +-
>  qapi/block-core.json   |   3 +-
>  block/qcow2-threads.c  | 157 +
>  block/qcow2.c  |   7 ++
>  5 files changed, 168 insertions(+), 2 deletions(-)

[...]

> diff --git a/block/qcow2-threads.c b/block/qcow2-threads.c
> index 7dbaf53489..0525718704 100644
> --- a/block/qcow2-threads.c
> +++ b/block/qcow2-threads.c
> @@ -28,6 +28,11 @@
>  #define ZLIB_CONST
>  #include 
>  
> +#ifdef CONFIG_ZSTD
> +#include 
> +#include 
> +#endif
> +
>  #include "qcow2.h"
>  #include "block/thread-pool.h"
>  #include "crypto.h"
> @@ -166,6 +171,148 @@ static ssize_t qcow2_zlib_decompress(void *dest, size_t 
> dest_size,
>  return ret;
>  }
>  
> +#ifdef CONFIG_ZSTD
> +
> +/*
> + * qcow2_zstd_compress()
> + *
> + * Compress @src_size bytes of data using zstd compression method
> + *
> + * @dest - destination buffer, @dest_size bytes
> + * @src - source buffer, @src_size bytes
> + *
> + * Returns: compressed size on success
> + *  -ENOMEM destination buffer is not enough to store compressed data
> + *  -EIOon any other error
> + */
> +static ssize_t qcow2_zstd_compress(void *dest, size_t dest_size,
> +   const void *src, size_t src_size)
> +{
> +ssize_t ret;
> +ZSTD_outBuffer output = { dest, dest_size, 0 };
> +ZSTD_inBuffer input = { src, src_size, 0 };

Minor style note: I think it’d be nicer to use designated initializers here.

> +ZSTD_CCtx *cctx = ZSTD_createCCtx();
> +
> +if (!cctx) {
> +return -EIO;
> +}
> +/*
> + * Use the zstd streamed interface for symmetry with decompression,
> + * where streaming is essential since we don't record the exact
> + * compressed size.
> + *
> + * In the loop, we try to compress all the data into one zstd frame.
> + * ZSTD_compressStream2 potentially can finish a frame earlier
> + * than the full input data is consumed. That's why we are looping
> + * until all the input data is consumed.
> + */
> +while (input.pos < input.size) {
> +size_t zstd_ret;
> +/*
> + * ZSTD spec: "You must continue calling ZSTD_compressStream2()
> + * with ZSTD_e_end until it returns 0, at which point you are
> + * free to start a new frame". We assume that "start a new frame"
> + * means call ZSTD_compressStream2 in the very beginning or when
> + * ZSTD_compressStream2 has returned with 0.
> + */
> +do {
> +zstd_ret = ZSTD_compressStream2(cctx, , , 
> ZSTD_e_end);

The spec makes it sound to me like ZSTD_e_end will always complete in a
single call if there’s enough space in the output buffer.  So the only
team we have to loop would be when there isn’t enough space anyway:

It says this about ZSTD_e_end:
> flush operation is the same, and follows same rules as calling
> ZSTD_compressStream2() with ZSTD_e_flush.

Those rules being:
> Note that, if `output->size` is too small, a single invocation with
> ZSTD_e_flush might not be enough (return code > 0).

So it seems like it will only return a value > 0 if the output buffer is
definitely too small.

The spec also notes that the return value is greater than 0 if:
> >0 if some data still present within internal buffer (the value is
> minimal estimation of remaining size),

So it’s a minimum estimate.  That’s another point that heavily implies
to me that if the return value 

[PATCH v20 3/4] qcow2: add zstd cluster compression

[Denis Plotnikov]

zstd significantly reduces cluster compression time.
It provides better compression performance maintaining
the same level of the compression ratio in comparison with
zlib, which, at the moment, is the only compression
method available.

The performance test results:
Test compresses and decompresses qemu qcow2 image with just
installed rhel-7.6 guest.
Image cluster size: 64K. Image on disk size: 2.2G

The test was conducted with brd disk to reduce the influence
of disk subsystem to the test results.
The results is given in seconds.

compress cmd:
  time ./qemu-img convert -O qcow2 -c -o compression_type=[zlib|zstd]
  src.img [zlib|zstd]_compressed.img
decompress cmd
  time ./qemu-img convert -O qcow2
  [zlib|zstd]_compressed.img uncompressed.img

   compression   decompression
 zlib   zstd   zlib zstd

real 65.5   16.3 (-75 %)1.9  1.6 (-16 %)
user 65.0   15.85.3  2.5
sys   3.30.22.0  2.0

Both ZLIB and ZSTD gave the same compression ratio: 1.57
compressed image size in both cases: 1.4G

Signed-off-by: Denis Plotnikov 
Reviewed-by: Vladimir Sementsov-Ogievskiy 
Reviewed-by: Alberto Garcia 
QAPI part:
Acked-by: Markus Armbruster 
---
 docs/interop/qcow2.txt |   1 +
 configure  |   2 +-
 qapi/block-core.json   |   3 +-
 block/qcow2-threads.c  | 157 +
 block/qcow2.c  |   7 ++
 5 files changed, 168 insertions(+), 2 deletions(-)

diff --git a/docs/interop/qcow2.txt b/docs/interop/qcow2.txt
index 640e0eca40..18a77f737e 100644
--- a/docs/interop/qcow2.txt
+++ b/docs/interop/qcow2.txt
@@ -209,6 +209,7 @@ version 2.
 
 Available compression type values:
 0: zlib 
+1: zstd 
 
 
 === Header padding ===
diff --git a/configure b/configure
index 23b5e93752..4e3a1690ea 100755
--- a/configure
+++ b/configure
@@ -1861,7 +1861,7 @@ disabled with --disable-FEATURE, default is enabled if 
available:
   lzfse   support of lzfse compression library
   (for reading lzfse-compressed dmg images)
   zstdsupport for zstd compression library
-  (for migration compression)
+  (for migration compression and qcow2 cluster compression)
   seccomp seccomp support
   coroutine-pool  coroutine freelist (better performance)
   glusterfs   GlusterFS backend
diff --git a/qapi/block-core.json b/qapi/block-core.json
index 1522e2983f..6fbacddab2 100644
--- a/qapi/block-core.json
+++ b/qapi/block-core.json
@@ -4293,11 +4293,12 @@
 # Compression type used in qcow2 image file
 #
 # @zlib: zlib compression, see 
+# @zstd: zstd compression, see 
 #
 # Since: 5.1
 ##
 { 'enum': 'Qcow2CompressionType',
-  'data': [ 'zlib' ] }
+  'data': [ 'zlib', { 'name': 'zstd', 'if': 'defined(CONFIG_ZSTD)' } ] }
 
 ##
 # @BlockdevCreateOptionsQcow2:
diff --git a/block/qcow2-threads.c b/block/qcow2-threads.c
index 7dbaf53489..0525718704 100644
--- a/block/qcow2-threads.c
+++ b/block/qcow2-threads.c
@@ -28,6 +28,11 @@
 #define ZLIB_CONST
 #include 
 
+#ifdef CONFIG_ZSTD
+#include 
+#include 
+#endif
+
 #include "qcow2.h"
 #include "block/thread-pool.h"
 #include "crypto.h"
@@ -166,6 +171,148 @@ static ssize_t qcow2_zlib_decompress(void *dest, size_t 
dest_size,
 return ret;
 }
 
+#ifdef CONFIG_ZSTD
+
+/*
+ * qcow2_zstd_compress()
+ *
+ * Compress @src_size bytes of data using zstd compression method
+ *
+ * @dest - destination buffer, @dest_size bytes
+ * @src - source buffer, @src_size bytes
+ *
+ * Returns: compressed size on success
+ *  -ENOMEM destination buffer is not enough to store compressed data
+ *  -EIOon any other error
+ */
+static ssize_t qcow2_zstd_compress(void *dest, size_t dest_size,
+   const void *src, size_t src_size)
+{
+ssize_t ret;
+ZSTD_outBuffer output = { dest, dest_size, 0 };
+ZSTD_inBuffer input = { src, src_size, 0 };
+ZSTD_CCtx *cctx = ZSTD_createCCtx();
+
+if (!cctx) {
+return -EIO;
+}
+/*
+ * Use the zstd streamed interface for symmetry with decompression,
+ * where streaming is essential since we don't record the exact
+ * compressed size.
+ *
+ * In the loop, we try to compress all the data into one zstd frame.
+ * ZSTD_compressStream2 potentially can finish a frame earlier
+ * than the full input data is consumed. That's why we are looping
+ * until all the input data is consumed.
+ */
+while (input.pos < input.size) {
+size_t zstd_ret;
+/*
+ * ZSTD spec: "You must continue calling ZSTD_compressStream2()
+ * with ZSTD_e_end until it returns 0, at