In initr_bootstage() we call bootstage_mark_name() which ends up calling
timer_get_us(). This call happens before initr_dm(), which inits driver
model.
On x86 we set gd->timer to NULL in the transition from board_init_f()
to board_init_r(). See board_init_f_r() for this assignment. So U-Boot
knows there is no timer available in the period immediately after
relocation.
On x86 the timer_get_us() call is implemented as calls to get_ticks() and
get_tbclk(). Both of these call dm_timer_init() to set up the timer, if
gd->timer is NULL and the early timer is not available.
However dm_timer_init() cannot succeed before initr_dm() is called.
So it seems that on x86 if we want to use CONFIG_BOOTSTAGE we must enable
CONFIG_TIMER_EARLY. Update the Kconfig to handle this.
Note: On most architectures we can rely on the pre-relocation memory still
being available, so that gd->timer pointers to a valid timer device and
everything works correctly. Admittedly this is not strictly correct since
the timer device is set up by pre-relocation U-Boot, but normally this is
fine. On x86 the 'CAR' (cache-as-RAM) memory used by pre-relocation U-Boot
disappears in board_init_f_r() and any attempt to access it will hang.
This is the reason why we must mark the timer as invalid when we get to
board_init_f_r().
Signed-off-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Bin Meng <bmeng.cn@gmail.com>
config TIMER_EARLY
bool "Allow timer to be used early in U-Boot"
depends on TIMER
+ # initr_bootstage() requires a timer and is called before initr_dm()
+ # so only the early timer is available
+ default y if X86 && BOOTSTAGE
help
In some cases the timer must be accessible before driver model is
active. Examples include when using CONFIG_TRACE to trace U-Boot's