mtd: only use __xipram annotation when XIP_KERNEL is set

When XIP_KERNEL is enabled, some functions are defined in the .data
ELF section because we require them to be in RAM whenever we communicate
with the flash chip. However this causes problems when FTRACE is
enabled and gcc emits calls to __gnu_mcount_nc in the function
prolog:

drivers/built-in.o: In function `cfi_chip_setup':
:(.data+0x272fc): relocation truncated to fit: R_ARM_CALL against symbol `__gnu_mcount_nc' defined in .text section in arch/arm/kernel/built-in.o
drivers/built-in.o: In function `cfi_probe_chip':
:(.data+0x27de8): relocation truncated to fit: R_ARM_CALL against symbol `__gnu_mcount_nc' defined in .text section in arch/arm/kernel/built-in.o
/tmp/ccY172rP.s: Assembler messages:
/tmp/ccY172rP.s:70: Warning: ignoring changed section attributes for .data
/tmp/ccY172rP.s: Error: 1 warning, treating warnings as errors
make[5]: *** [drivers/mtd/chips/cfi_probe.o] Error 1
/tmp/ccK4rjeO.s: Assembler messages:
/tmp/ccK4rjeO.s:421: Warning: ignoring changed section attributes for .data
/tmp/ccK4rjeO.s: Error: 1 warning, treating warnings as errors
make[5]: *** [drivers/mtd/chips/cfi_util.o] Error 1
/tmp/ccUvhCYR.s: Assembler messages:
/tmp/ccUvhCYR.s:1895: Warning: ignoring changed section attributes for .data
/tmp/ccUvhCYR.s: Error: 1 warning, treating warnings as errors

Specifically, this does not work because the .data section is not
marked executable, which leads LD to not generate trampolines for
long calls.

This moves the __xipram functions into their own .xiptext section instead.
The section is still placed next to .data and located in RAM but is marked
executable, which avoids the build errors.

Also, we only need to place the XIP functions into a separate section
if both CONFIG_XIP_KERNEL and CONFIG_MTD_XIP are set: When only MTD_XIP
is used, the whole kernel is still in RAM and we do not need to worry
about pulling out the rug under it. When only XIP_KERNEL but not MTD_XIP
is set, the kernel is in some form of ROM, but we never write to it.

Note that MTD_XIP has been broken on ARM since around 2011 or 2012. I
have sent another patch[2] to fix compilation, which I plan to merge
through arm-soc unless there are objections. The obvious alternative
to that would be to completely rip out the MTD_XIP support from the
kernel, since obviously nobody has been using it in a long while.

Link: [1] https://patchwork.kernel.org/patch/8109771/
Link: [2] https://patchwork.kernel.org/patch/9855225/
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>

include/asm-generic/vmlinux.lds.h

@@ -203,6 +203,7 @@
  * pull in .data..stuff which has its own requirements. Same for bss.
  */
 #define DATA_DATA							\
+	*(.xiptext)							\
 	*(.data .data.[0-9a-zA-Z_]*)					\
 	*(.ref.data)							\
 	*(.data..shared_aligned) /* percpu related */			\

include/linux/mtd/xip.h

@@ -30,7 +30,9 @@
  * obviously not be running from flash.  The __xipram is therefore marking
  * those functions so they get relocated to ram.
  */
-#define __xipram noinline __attribute__ ((__section__ (".data")))
+#ifdef CONFIG_XIP_KERNEL
+#define __xipram noinline __attribute__ ((__section__ (".xiptext")))
+#endif
 
 /*
  * Each architecture has to provide the following macros.  They must access
@@ -90,10 +92,10 @@
 #define xip_cpu_idle()  do { } while (0)
 #endif
 
-#else
-
-#define __xipram
-
 #endif /* CONFIG_MTD_XIP */
 
+#ifndef __xipram
+#define __xipram
+#endif
+
 #endif /* __LINUX_MTD_XIP_H__ */