kernel/power/swap.c | 22 +++++++++++++++++----- 1 file changed, 17 insertions(+), 5 deletions(-)
A new kernel parameter 'cmp_threads=' is introduced to
allow tuning the number of compression/decompression threads at boot.
Signed-off-by: Xueqin Luo <luoxueqin@kylinos.cn>
---
kernel/power/swap.c | 22 +++++++++++++++++-----
1 file changed, 17 insertions(+), 5 deletions(-)
diff --git a/kernel/power/swap.c b/kernel/power/swap.c
index ad13c461b657..43280e08a4ad 100644
--- a/kernel/power/swap.c
+++ b/kernel/power/swap.c
@@ -520,7 +520,8 @@ static int swap_writer_finish(struct swap_map_handle *handle,
#define CMP_SIZE (CMP_PAGES * PAGE_SIZE)
/* Maximum number of threads for compression/decompression. */
-#define CMP_THREADS 3
+#define CMP_MAX_THREADS 12
+static int cmp_threads = 3
/* Minimum/maximum number of pages for read buffering. */
#define CMP_MIN_RD_PAGES 1024
@@ -585,8 +586,8 @@ struct crc_data {
wait_queue_head_t go; /* start crc update */
wait_queue_head_t done; /* crc update done */
u32 *crc32; /* points to handle's crc32 */
- size_t *unc_len[CMP_THREADS]; /* uncompressed lengths */
- unsigned char *unc[CMP_THREADS]; /* uncompressed data */
+ size_t *unc_len[CMP_MAX_THREADS]; /* uncompressed lengths */
+ unsigned char *unc[CMP_MAX_THREADS]; /* uncompressed data */
};
/*
@@ -703,7 +704,7 @@ static int save_compressed_image(struct swap_map_handle *handle,
* footprint.
*/
nr_threads = num_online_cpus() - 1;
- nr_threads = clamp_val(nr_threads, 1, CMP_THREADS);
+ nr_threads = clamp_val(nr_threads, 1, cmp_threads);
page = (void *)__get_free_page(GFP_NOIO | __GFP_HIGH);
if (!page) {
@@ -1223,7 +1224,7 @@ static int load_compressed_image(struct swap_map_handle *handle,
* footprint.
*/
nr_threads = num_online_cpus() - 1;
- nr_threads = clamp_val(nr_threads, 1, CMP_THREADS);
+ nr_threads = clamp_val(nr_threads, 1, cmp_threads);
page = vmalloc(array_size(CMP_MAX_RD_PAGES, sizeof(*page)));
if (!page) {
@@ -1667,3 +1668,14 @@ static int __init swsusp_header_init(void)
}
core_initcall(swsusp_header_init);
+
+static int __init cmp_threads_setup(char *str)
+{
+ int rc = kstrtouint(str, 0, &cmp_threads);
+ if (rc)
+ return rc;
+ return 1;
+
+}
+
+__setup("cmp_threads=", cmp_threads_setup);
--
2.43.0On Tue, Aug 26, 2025 at 11:19 AM Xueqin Luo <luoxueqin@kylinos.cn> wrote:
>
> A new kernel parameter 'cmp_threads=' is introduced to
> allow tuning the number of compression/decompression threads at boot.
And why is it useful/needed?
> Signed-off-by: Xueqin Luo <luoxueqin@kylinos.cn>
> ---
> kernel/power/swap.c | 22 +++++++++++++++++-----
> 1 file changed, 17 insertions(+), 5 deletions(-)
>
> diff --git a/kernel/power/swap.c b/kernel/power/swap.c
> index ad13c461b657..43280e08a4ad 100644
> --- a/kernel/power/swap.c
> +++ b/kernel/power/swap.c
> @@ -520,7 +520,8 @@ static int swap_writer_finish(struct swap_map_handle *handle,
> #define CMP_SIZE (CMP_PAGES * PAGE_SIZE)
>
> /* Maximum number of threads for compression/decompression. */
> -#define CMP_THREADS 3
> +#define CMP_MAX_THREADS 12
> +static int cmp_threads = 3
>
> /* Minimum/maximum number of pages for read buffering. */
> #define CMP_MIN_RD_PAGES 1024
> @@ -585,8 +586,8 @@ struct crc_data {
> wait_queue_head_t go; /* start crc update */
> wait_queue_head_t done; /* crc update done */
> u32 *crc32; /* points to handle's crc32 */
> - size_t *unc_len[CMP_THREADS]; /* uncompressed lengths */
> - unsigned char *unc[CMP_THREADS]; /* uncompressed data */
> + size_t *unc_len[CMP_MAX_THREADS]; /* uncompressed lengths */
> + unsigned char *unc[CMP_MAX_THREADS]; /* uncompressed data */
> };
>
> /*
> @@ -703,7 +704,7 @@ static int save_compressed_image(struct swap_map_handle *handle,
> * footprint.
> */
> nr_threads = num_online_cpus() - 1;
> - nr_threads = clamp_val(nr_threads, 1, CMP_THREADS);
> + nr_threads = clamp_val(nr_threads, 1, cmp_threads);
>
> page = (void *)__get_free_page(GFP_NOIO | __GFP_HIGH);
> if (!page) {
> @@ -1223,7 +1224,7 @@ static int load_compressed_image(struct swap_map_handle *handle,
> * footprint.
> */
> nr_threads = num_online_cpus() - 1;
> - nr_threads = clamp_val(nr_threads, 1, CMP_THREADS);
> + nr_threads = clamp_val(nr_threads, 1, cmp_threads);
>
> page = vmalloc(array_size(CMP_MAX_RD_PAGES, sizeof(*page)));
> if (!page) {
> @@ -1667,3 +1668,14 @@ static int __init swsusp_header_init(void)
> }
>
> core_initcall(swsusp_header_init);
> +
> +static int __init cmp_threads_setup(char *str)
> +{
> + int rc = kstrtouint(str, 0, &cmp_threads);
> + if (rc)
> + return rc;
> + return 1;
> +
> +}
> +
> +__setup("cmp_threads=", cmp_threads_setup);
> --
> 2.43.0
>
在 2025/8/26 19:43, Rafael J. Wysocki 写道:
> On Tue, Aug 26, 2025 at 11:19 AM Xueqin Luo <luoxueqin@kylinos.cn> wrote:
>> A new kernel parameter 'cmp_threads=' is introduced to
>> allow tuning the number of compression/decompression threads at boot.
> And why is it useful/needed?
The number of compression/decompression threads directly impacts
hibernate and resume time.
In our tests(averaged over 10 runs):
cmp_threads hibernate(s) resume(s)
3 12.14 18.86
4 12.28 17.48
5 11.09 16.77
6 11.08 16.44
With 5–6 threads, resume latency improves by ~12% compared to 3 threads.
But on low-core systems,
more threads may cause contention. Making it configurable allows
integrators to balance performance
and CPU usage across different hardware without recompiling the kernel.
>> Signed-off-by: Xueqin Luo <luoxueqin@kylinos.cn>
>> ---
>> kernel/power/swap.c | 22 +++++++++++++++++-----
>> 1 file changed, 17 insertions(+), 5 deletions(-)
>>
>> diff --git a/kernel/power/swap.c b/kernel/power/swap.c
>> index ad13c461b657..43280e08a4ad 100644
>> --- a/kernel/power/swap.c
>> +++ b/kernel/power/swap.c
>> @@ -520,7 +520,8 @@ static int swap_writer_finish(struct swap_map_handle *handle,
>> #define CMP_SIZE (CMP_PAGES * PAGE_SIZE)
>>
>> /* Maximum number of threads for compression/decompression. */
>> -#define CMP_THREADS 3
>> +#define CMP_MAX_THREADS 12
>> +static int cmp_threads = 3
>>
>> /* Minimum/maximum number of pages for read buffering. */
>> #define CMP_MIN_RD_PAGES 1024
>> @@ -585,8 +586,8 @@ struct crc_data {
>> wait_queue_head_t go; /* start crc update */
>> wait_queue_head_t done; /* crc update done */
>> u32 *crc32; /* points to handle's crc32 */
>> - size_t *unc_len[CMP_THREADS]; /* uncompressed lengths */
>> - unsigned char *unc[CMP_THREADS]; /* uncompressed data */
>> + size_t *unc_len[CMP_MAX_THREADS]; /* uncompressed lengths */
>> + unsigned char *unc[CMP_MAX_THREADS]; /* uncompressed data */
>> };
>>
>> /*
>> @@ -703,7 +704,7 @@ static int save_compressed_image(struct swap_map_handle *handle,
>> * footprint.
>> */
>> nr_threads = num_online_cpus() - 1;
>> - nr_threads = clamp_val(nr_threads, 1, CMP_THREADS);
>> + nr_threads = clamp_val(nr_threads, 1, cmp_threads);
>>
>> page = (void *)__get_free_page(GFP_NOIO | __GFP_HIGH);
>> if (!page) {
>> @@ -1223,7 +1224,7 @@ static int load_compressed_image(struct swap_map_handle *handle,
>> * footprint.
>> */
>> nr_threads = num_online_cpus() - 1;
>> - nr_threads = clamp_val(nr_threads, 1, CMP_THREADS);
>> + nr_threads = clamp_val(nr_threads, 1, cmp_threads);
>>
>> page = vmalloc(array_size(CMP_MAX_RD_PAGES, sizeof(*page)));
>> if (!page) {
>> @@ -1667,3 +1668,14 @@ static int __init swsusp_header_init(void)
>> }
>>
>> core_initcall(swsusp_header_init);
>> +
>> +static int __init cmp_threads_setup(char *str)
>> +{
>> + int rc = kstrtouint(str, 0, &cmp_threads);
>> + if (rc)
>> + return rc;
>> + return 1;
>> +
>> +}
>> +
>> +__setup("cmp_threads=", cmp_threads_setup);
>> --
>> 2.43.0
>>
On Thu, Aug 28, 2025 at 5:30 AM luoxueqin <luoxueqin@kylinos.cn> wrote: > > > 在 2025/8/26 19:43, Rafael J. Wysocki 写道: > > On Tue, Aug 26, 2025 at 11:19 AM Xueqin Luo <luoxueqin@kylinos.cn> wrote: > >> A new kernel parameter 'cmp_threads=' is introduced to > >> allow tuning the number of compression/decompression threads at boot. > > And why is it useful/needed? > The number of compression/decompression threads directly impacts > hibernate and resume time. > In our tests(averaged over 10 runs): > cmp_threads hibernate(s) resume(s) > 3 12.14 18.86 > 4 12.28 17.48 > 5 11.09 16.77 > 6 11.08 16.44 > With 5–6 threads, resume latency improves by ~12% compared to 3 threads. > But on low-core systems, > more threads may cause contention. Making it configurable allows > integrators to balance performance > and CPU usage across different hardware without recompiling the kernel. So please add this information to the changelog of the patch and resend it. Thanks!
© 2016 - 2025 Red Hat, Inc.