lib: sparse: allocate FASTBOOT_MAX_BLK_WRITE instead of small number
PATCH v2 1/1 lib: sparse: allocate FASTBOOT_MAX_BLK_WRITE instead of small number
Commit 62649165cb02 ("lib: sparse: Make CHUNK_TYPE_RAW buffer aligned") fixed cache alignment for systems with a D-CACHE.
However it introduced some performance regressions [1] on system flashing huge images, such as Android.
On AM62x SK EVM, we also observe such performance penalty:
Sending sparse 'super' 1/2 (768793 KB) OKAY [ 23.954s] Writing 'super' OKAY [ 75.926s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.641s] Writing 'super' OKAY [ 62.849s]
Finished. Total time: 182.474s
The reason for this is that we use an arbitrary small buffer (info->blksz * 100) for transferring.
Fix it by using a bigger buffer (info->blksz * FASTBOOT_MAX_BLK_WRITE) as suggested in the original's patch review [2].
With this patch, performance impact is mitigated:
Sending sparse 'super' 1/2 (768793 KB) OKAY [ 23.912s] Writing 'super' OKAY [ 15.780s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.581s] Writing 'super' OKAY [ 17.192s]
Finished. Total time: 76.569s
[1] https://lore.kernel.org/r/[email protected]
[2] https://lore.kernel.org/r/all/[email protected]/
---
Changes in v2:
- Use FASTBOOT_MAX_BLK_WRITE instead of blkcnt (Qianfan)
- Link to v1: https://lore.kernel.org/r/[email protected]
--- drivers/fastboot/fb_mmc.c | 2 -- include/image-sparse.h | 2 ++ lib/image-sparse.c | 3 ++- 3 files changed, 4 insertions(+), 3 deletions(-)
drivers/fastboot/fb_mmc.c +0 -2
@@ -19,8 +19,6 @@#include <linux/compat.h> #include <android_image.h> -#define FASTBOOT_MAX_BLK_WRITE 16384 -#define BOOT_PARTITION_NAME "boot"struct fb_mmc_sparse {
include/image-sparse.h +2 -0
@@ -7,6 +7,8 @@#include <part.h> #include <sparse_format.h> +#define FASTBOOT_MAX_BLK_WRITE 16384 +#define ROUNDUP (x, y) (((x) + ((y) - 1)) & ~((y) - 1))struct sparse_storage {
lib/image-sparse.c +2 -1
@@ -55,7 +55,8 @@ static lbaint_t write_sparse_chunk_raw(struct sparse_storage *info,void *data, char *response) { - lbaint_t n = blkcnt, write_blks, blks = 0, aligned_buf_blks = 100; + lbaint_t n = blkcnt, write_blks, blks = 0; +lbaint_t aligned_buf_blks =FASTBOOT_MAX_BLK_WRITE ;uint32_t *aligned_buf =NULL ;if (CONFIG_IS_ENABLED (SYS_DCACHE_OFF )) {
---
base-commit: 923de765ee1a5b26310f02cb42dcbad9e2b011c5
change-id: 20230616-sparse-flash-fix-9c2852aa8d16
Best regards,
Signature · 1 line
Mattijs Korpershoek <[email protected]>
Quoted 59 lines · Mattijs Korpershoek
Commit
62649165cb02("lib: sparse: MakeCHUNK_TYPE_RAWbuffer aligned") fixed cache alignment for systems with a D-CACHE.However it introduced some performance regressions [1] on system flashing huge images, such as Android.
On AM62x SK EVM, we also observe such performance penalty:
Sending sparse 'super' 1/2 (768793 KB) OKAY [ 23.954s] Writing 'super' OKAY [ 75.926s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.641s] Writing 'super' OKAY [ 62.849s]Finished. Total time: 182.474s
The reason for this is that we use an arbitrary small buffer (
info->blksz* 100) for transferring.Fix it by using a bigger buffer (
info->blksz*FASTBOOT_MAX_BLK_WRITE) as suggested in the original's patch review [2].With this patch, performance impact is mitigated:
Sending sparse 'super' 1/2 (768793 KB) OKAY [ 23.912s] Writing 'super' OKAY [ 15.780s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.581s] Writing 'super' OKAY [ 17.192s]Finished. Total time: 76.569s
[1] https://lore.kernel.org/r/[email protected]
[2] https://lore.kernel.org/r/all/[email protected]/Fixes 62649165cb02 ("lib: sparse: Make CHUNK_TYPE_RAW buffer aligned")Signed-off-by Mattijs Korpershoek <[email protected]>---
Changes in v2:
- UseFASTBOOT_MAX_BLK_WRITEinstead of blkcnt (Qianfan)
- Link to v1: https://lore.kernel.org/r/[email protected]--- drivers/fastboot/fb_mmc.c | 2 -- include/image-sparse.h | 2 ++ lib/image-sparse.c | 3 ++- 3 files changed, 4 insertions(+), 3 deletions(-)drivers/fastboot/fb_mmc.c +0
-2@@ -19,8 +19,6 @@#include <linux/compat.h> #include <android_image.h> -#define FASTBOOT_MAX_BLK_WRITE 16384 -#define BOOT_PARTITION_NAME "boot"struct fb_mmc_sparse {include/image-sparse.h +0
-0
include/image-sparse.h line 10
Quoted 4 lines
include/image-sparse.h +1
-0@@ -7,6 +7,8 @@#include <part.h> #include <sparse_format.h> +#define FASTBOOT_MAX_BLK_WRITE 16384
Hi:
Just a personal suggestion, define sometings like FASTBOOT_MAX_BLK_WRITE in image-sparse.c is very strange.
Or maybe define a marco such as SPARSE_MAX_BLK_WRITE and set a default value to 16384, and leave some comments for why we choice this value.
Thanks.
Quoted 24 lines
include/image-sparse.h +1
-0+#define ROUNDUP (x, y) (((x) + ((y) - 1)) & ~((y) - 1))struct sparse_storage {lib/image-sparse.c +2
-2@@ -55,7 +55,8 @@ static lbaint_t write_sparse_chunk_raw(struct sparse_storage *info,void *data, char *response) { - lbaint_t n = blkcnt, write_blks, blks = 0, aligned_buf_blks = 100; + lbaint_t n = blkcnt, write_blks, blks = 0; +lbaint_t aligned_buf_blks =FASTBOOT_MAX_BLK_WRITE ;uint32_t *aligned_buf =NULL ;if (CONFIG_IS_ENABLED (SYS_DCACHE_OFF )) { ---base-commit:
923de765ee1a5b26310f02cb42dcbad9e2b011c5
change-id: 20230616-sparse-flash-fix-9c2852aa8d16Best regards,
Hi Qianfan,
Thank you for your review.
Quoted 63 lines · qianfan
Quoted 62 lines · Mattijs Korpershoek
Commit
62649165cb02("lib: sparse: MakeCHUNK_TYPE_RAWbuffer aligned") fixed cache alignment for systems with a D-CACHE.However it introduced some performance regressions [1] on system flashing huge images, such as Android.
On AM62x SK EVM, we also observe such performance penalty:
Sending sparse 'super' 1/2 (768793 KB) OKAY [ 23.954s] Writing 'super' OKAY [ 75.926s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.641s] Writing 'super' OKAY [ 62.849s]Finished. Total time: 182.474s
The reason for this is that we use an arbitrary small buffer (
info->blksz* 100) for transferring.Fix it by using a bigger buffer (
info->blksz*FASTBOOT_MAX_BLK_WRITE) as suggested in the original's patch review [2].With this patch, performance impact is mitigated:
Sending sparse 'super' 1/2 (768793 KB) OKAY [ 23.912s] Writing 'super' OKAY [ 15.780s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.581s] Writing 'super' OKAY [ 17.192s]Finished. Total time: 76.569s
[1] https://lore.kernel.org/r/[email protected]
[2] https://lore.kernel.org/r/all/[email protected]/Fixes 62649165cb02 ("lib: sparse: Make CHUNK_TYPE_RAW buffer aligned")Signed-off-by Mattijs Korpershoek <[email protected]>---
Changes in v2:
- UseFASTBOOT_MAX_BLK_WRITEinstead of blkcnt (Qianfan)
- Link to v1: https://lore.kernel.org/r/[email protected]--- drivers/fastboot/fb_mmc.c | 2 -- include/image-sparse.h | 2 ++ lib/image-sparse.c | 3 ++- 3 files changed, 4 insertions(+), 3 deletions(-)drivers/fastboot/fb_mmc.c +0
-2@@ -19,8 +19,6 @@#include <linux/compat.h> #include <android_image.h> -#define FASTBOOT_MAX_BLK_WRITE 16384 -#define BOOT_PARTITION_NAME "boot"struct fb_mmc_sparse {include/image-sparse.h +0
-0@@ -7,6 +7,8 @@#include <part.h> #include <sparse_format.h>
Quoted 8 lines
include/image-sparse.h line 10
Quoted 1 line
include/image-sparse.h +1
-0+#define FASTBOOT_MAX_BLK_WRITE 16384Hi:
Just a personal suggestion, define sometings like
FASTBOOT_MAX_BLK_WRITEinimage-sparse.cis very strange.Or maybe define a marco such as
SPARSE_MAX_BLK_WRITEand set a default value to 16384, and leave some comments for why we choice this value.
I don't agree with having a duplicating between SPARSE_MAX_BLK_WRITE and FASTBOOT_MAX_BLK_WRITE. code comments can rot as well.
And FASTBOOT_MAX_BLK_WRITE is used for both sparse and unsparsed image, which is why I chose to not rename it.
Quoted 26 lines
Thanks.
Quoted 24 lines
include/image-sparse.h +1
-0+#define ROUNDUP (x, y) (((x) + ((y) - 1)) & ~((y) - 1))struct sparse_storage {lib/image-sparse.c +2
-2@@ -55,7 +55,8 @@ static lbaint_t write_sparse_chunk_raw(struct sparse_storage *info,void *data, char *response) { - lbaint_t n = blkcnt, write_blks, blks = 0, aligned_buf_blks = 100; + lbaint_t n = blkcnt, write_blks, blks = 0; +lbaint_t aligned_buf_blks =FASTBOOT_MAX_BLK_WRITE ;uint32_t *aligned_buf =NULL ;if (CONFIG_IS_ENABLED (SYS_DCACHE_OFF )) { ---base-commit:
923de765ee1a5b26310f02cb42dcbad9e2b011c5
change-id: 20230616-sparse-flash-fix-9c2852aa8d16Best regards,
Quoted 24 lines · Mattijs Korpershoek
Commit
62649165cb02("lib: sparse: MakeCHUNK_TYPE_RAWbuffer aligned") fixed cache alignment for systems with a D-CACHE.However it introduced some performance regressions [1] on system flashing huge images, such as Android.
On AM62x SK EVM, we also observe such performance penalty:
Sending sparse 'super' 1/2 (768793 KB) OKAY [ 23.954s] Writing 'super' OKAY [ 75.926s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.641s] Writing 'super' OKAY [ 62.849s]Finished. Total time: 182.474s
The reason for this is that we use an arbitrary small buffer (
info->blksz* 100) for transferring.Fix it by using a bigger buffer (
info->blksz*FASTBOOT_MAX_BLK_WRITE) as suggested in the original's patch review [2].With this patch, performance impact is mitigated:
Sending sparse 'super' 1/2 (768793 KB) OKAY [ 23.912s] Writing 'super' OKAY [ 15.780s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.581s] Writing 'super' OKAY [ 17.192s]
Quoted 7 lines
Finished. Total time: 76.569s
[1] https://lore.kernel.org/r/[email protected]
[2] https://lore.kernel.org/r/all/[email protected]/Fixes 62649165cb02 ("lib: sparse: Make CHUNK_TYPE_RAW buffer aligned")Signed-off-by Mattijs Korpershoek <[email protected]>
Applied to u-boot/master, thanks!
Signature · 1 line
Tom