On 12/22/22 01:21, Rick Chen wrote:
When fit image boots from ram, the payload will be prepared in the address of SPL_LOAD_FIT_ADDRESS. In spl fit generic flow, it will malloc another memory address and copy whole fit image to this malloc address. But it is un-necessary for booting from RAM.
This should mostly be solved by using `mkimage -E` or binman's `fit,external-offset`. Then only the FIT structure, without any data, will be copied.
This patch improves this flow by declare the board_spl_fit_buffer_addr() to replace the original one. The larger image size (eq: Kernel Image 10~20MB), it can save more booting time.
Also enhance memcpy function by checking source and destination address. If they are the same address, just return and don't copy data anymore.
Signed-off-by: Rick Chen rick@andestech.com
arch/riscv/lib/memcpy.S | 2 ++ arch/riscv/lib/spl.c | 16 ++++++++++++++++ 2 files changed, 18 insertions(+)
diff --git a/arch/riscv/lib/memcpy.S b/arch/riscv/lib/memcpy.S index 00672c19ad..9884077c93 100644 --- a/arch/riscv/lib/memcpy.S +++ b/arch/riscv/lib/memcpy.S @@ -9,6 +9,7 @@ /* void *memcpy(void *, const void *, size_t) */ ENTRY(__memcpy) WEAK(memcpy)
- beq a0, a1, .copy_end
I see one call to memcpy() in common/spl/spl_ram.c. I think it would make more sense to do the check there instead of adding a branch to every memcpy() call.
/* Save for return value */ mv t6, a0
@@ -121,6 +122,7 @@ WEAK(memcpy) 2:
mv a0, t6 +.copy_end: ret
.Lmisaligned_word_copy: diff --git a/arch/riscv/lib/spl.c b/arch/riscv/lib/spl.c index f4d3b67e5d..18f86ee207 100644 --- a/arch/riscv/lib/spl.c +++ b/arch/riscv/lib/spl.c @@ -61,3 +61,19 @@ void __noreturn jump_to_image_no_args(struct spl_image_info *spl_image) #endif image_entry(gd->arch.boot_hart, fdt_blob); }
+#ifdef CONFIG_SPL_RAM_SUPPORT +struct legacy_img_hdr *spl_get_load_buffer(ssize_t offset, size_t size) +{
- return (void *)(CONFIG_SPL_LOAD_FIT_ADDRESS + offset);
+}
+void *board_spl_fit_buffer_addr(ulong fit_size, int sectors, int bl_len) +{
- void *buf;
- buf = spl_get_load_buffer(0, sectors * bl_len);
- return buf;
+} +#endif
You cannot provide strong definitions of these functions at an architecture level, as this prevents any board from overriding them.
Regards, Samuel