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 ---
drivers/timer/Kconfig | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/timer/Kconfig b/drivers/timer/Kconfig index 5ab6749193c..ff434de6f7c 100644 --- a/drivers/timer/Kconfig +++ b/drivers/timer/Kconfig @@ -30,6 +30,9 @@ config TPL_TIMER 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