Dear Michael,
Ping!
In message 20101026210959.518031365D8@gemini.denx.de I wrote:
Dear Michael Zaidman,
In message d520c6ef298416a03789ebfa4e05e257b5331693.1284965175.git.michael.zaidman@gmail.com you wrote:
- Revives POST for blackfin arch;
- Removes redundant code: arch/blackfin/lib/post.c arch/powerpc/cpu/ppc4xx/commproc.c arch/powerpc/cpu/mpc512x/common.c
- fixes up the post_word_{load|store} usage.
Unfortunately it turns out that the code now contains a few nasty bugs...
...
#define CONFIG_SYS_GBL_DATA_OFFSET (CONFIG_SYS_INIT_RAM_END - CONFIG_SYS_GBL_DATA_SIZE) -#define CONFIG_SYS_POST_WORD_ADDR (CONFIG_SYS_GBL_DATA_OFFSET - 0x4) -#define CONFIG_SYS_INIT_SP_OFFSET CONFIG_SYS_POST_WORD_ADDR +#define CONFIG_SYS_INIT_SP_OFFSET (CONFIG_SYS_GBL_DATA_OFFSET - 0x4)
This is a seriously broken design, as it sneaks in storage for a variable in a storage location where it is not expected.
The "official" layout is that we have CONFIG_SYS_INIT_RAM_BYTES available; the top CONFIG_SYS_GBL_DATA_SIZE bytes (now GENERATED_GBL_DATA_SIZE) are used for global data, and the part below is used for the stack. No other room is reserved there.
Shifting down the stack by 4 bytes as it's done here causes that the stack is not correctly aligned any more, which may cause really nasty subsequent errors.
But it's even worse.
diff --git a/include/configs/mpc5121-common.h b/include/configs/mpc5121-common.h index 96fab20..afae1ab 100644 --- a/include/configs/mpc5121-common.h +++ b/include/configs/mpc5121-common.h
...
-#define CONFIG_SYS_POST_WORD_ADDR (CONFIG_SYS_GBL_DATA_OFFSET - 0x4) -#define CONFIG_SYS_INIT_SP_OFFSET CONFIG_SYS_POST_WORD_ADDR +#define CONFIG_SYS_INIT_SP_OFFSET (CONFIG_SYS_GBL_DATA_OFFSET - 0x4)
There the same is done, but what happens actually?
Have a look how the stack setup gets implemented in "arch/powerpc/cpu/mpc512x/start.S":
... 244 in_flash: 245 lis r1, (CONFIG_SYS_INIT_RAM_ADDR + CONFIG_SYS_GBL_DATA_OFFSET)@h 246 ori r1, r1, (CONFIG_SYS_INIT_RAM_ADDR + CONFIG_SYS_GBL_DATA_OFFSET)@l 247 248 li r0, 0 /* Make room for stack frame header and */ 249 stwu r0, -4(r1) /* clear final stack frame so that */ 250 stwu r0, -4(r1) /* stack backtraces terminate cleanly */ ...
As you can see, the code does not use CONFIG_SYS_INIT_SP_OFFSET at all; instead it performs a calculation which should be redundant, but in the current code it means that the location of the POST_WORD is right in the initial stack.
I did not check if the code for other processors has similar issues.
"Reserving" private storage like that is bad, as other involved parties probably have no knowledge of such a private reservation.
Why do we not simply reserve a word in the global data structure instead?
This bug needs pretty urgent fixing.
Best regards,
Wolfgang Denk
Viele Grüße,
Wolfgang Denk