Hi Qianfan,
Thank you for your review.
On lun., juin 19, 2023 at 14:19, qianfan qianfanguijin@163.com wrote:
在 2023/6/16 21:26, Mattijs Korpershoek 写道:
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 * blkcnt) as suggested in the original's patch review [2].
With this patch, performance impact is mitigated: Sending sparse 'super' 1/2 (768793 KB) OKAY [ 24.006s] Writing 'super' OKAY [ 15.920s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.651s] Writing 'super' OKAY [ 14.665s] Finished. Total time: 74.346s
[1] https://lore.kernel.org/r/20221118121323.4009193-1-gary.bisson@boundarydevic... [2] https://lore.kernel.org/r/all/43e4c17c-4483-ec8e-f843-9b4c5569bd18@seco.com/
Fixes: 62649165cb02 ("lib: sparse: Make CHUNK_TYPE_RAW buffer aligned") Signed-off-by: Mattijs Korpershoek mkorpershoek@baylibre.com
lib/image-sparse.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/lib/image-sparse.c b/lib/image-sparse.c index 5ec0f94ab3eb..25aed0604192 100644 --- a/lib/image-sparse.c +++ b/lib/image-sparse.c @@ -55,7 +55,7 @@ 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, aligned_buf_blks = blkcnt;
Hi:
It's a good point that this code report the performance was affected by write large small mmc blks, not memory copy.
I believe memory copy also affects performance, but in my case, it has less impact than small mmc blks.
With 62649165cb02 reverted: Sending sparse 'super' 1/2 (768793 KB) OKAY [ 23.947s] Writing 'super' OKAY [ 12.983s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.600s] Writing 'super' OKAY [ 12.796s] Finished. Total time: 69.430s
With aligned_buf_blks = blkcnt: Sending sparse 'super' 1/2 (768793 KB) OKAY [ 24.072s] Writing 'super' OKAY [ 16.177s] Sending sparse 'super' 2/2 (629819 KB) OKAY [ 19.681s] Writing 'super' OKAY [ 14.845s] Finished. Total time: 74.919s
And I can not make sure whether memalign can always alloc such huge memory when we change the aligned_buf_blks to blkcnt.
Could you clarify the concern here? I've dumped blkcnt for my board (AM62x SK EVK) and the biggest blkcnt I found was: 131072
With info->blksz = 512, this gives me: 512 * 131072 = 67108864
Which is a memalign (memory alloc) of 64MB. Is 64MB really that big? (I don't realize it's that much)
Could you please set aligned_buf_blks to FASTBOOT_MAX_BLK_WRITE(16384) and test again?
With aligned_buf_blks = FASTBOOT_MAX_BLK_WRITE(16384): 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
So using FASTBOOT_MAX_BLK_WRITE is slightly worse than using blkcnt. But allocations (for blksz = 512) are smaller: 8MB instead of 64MB in my example.
I can spin up a v2 with FASTBOOT_MAX_BLK_WRITE but i'm waiting a little more feedback before doing so.
uint32_t *aligned_buf = NULL;
if (CONFIG_IS_ENABLED(SYS_DCACHE_OFF)) {
base-commit: 2f4664f5c3edc55b18d8906f256a4c8e303243c0 change-id: 20230616-sparse-flash-fix-9c2852aa8d16
Best regards,