cyclic: get rid of cyclic_init()

Currently, we must call cyclic_init() at some point before
cyclic_register() becomes possible. That turns out to be somewhat
awkward, especially with SPL, and has resulted in a watchdog callback
not being registered, thus causing the board to prematurely reset.

We already rely on gd->cyclic reliably being set to NULL by the asm
code that clears all of gd. Now that the cyclic list is a hlist, and
thus an empty list is represented by a NULL head pointer, and struct
cyclic_drv has no other members, we can just as well drop a level of
indirection and put the hlist_head directly in struct
global_data. This doesn't increase the size of struct global_data,
gets rid of an early malloc(), and generates slightly smaller code.

But primarily, this avoids having to call cyclic_init() early; the cyclic
infrastructure is simply ready to register callbacks as soon as we
enter C code.

We can still end up with schedule() being called from asm very early,
so we still need to check that gd itself has been properly initialized
[*], but once it has, gd->cyclic_list is perfectly fine to access, and
will just be an empty list.

As for cyclic_uninit(), it was never really the opposite of
cyclic_init() since it didn't free the struct cyclic_drv nor set
gd->cyclic to NULL. Rename it to cyclic_unregister_all() and use that
in test/, and also insert a call at the end of the board_init_f
sequence so that gd->cyclic_list is a fresh empty list before we enter
board_init_r().

A small piece of ugliness is that I had to add a cast in
cyclic_get_list() to silence a "discards 'volatile' qualifier"
warning, but that is completely equivalent to the existing handling of
the uclass_root_s list_head member.

[*] I'm not really sure where we guarantee that the register used for
gd contains 0 until it gets explicitly initialized, but that must be
the case, otherwise testing gd for being NULL would not make much sense.

Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk>
Reviewed-by: Stefan Roese <sr@denx.de>
Tested-by: Stefan Roese <sr@denx.de>
Tested-by: Tim Harvey <tharvey@gateworks.com> # imx8mm-venice-*

1 files changed, 10 insertions, 1 deletions

diff --git a/common/board_f.c b/common/board_f.c
index faa096479b3..7c26e235fc0 100644
--- a/common/board_f.c
+++ b/common/board_f.c
@@ -839,7 +839,6 @@ static const init_fnc_t init_sequence_f[] = {
 	initf_malloc,
 	log_init,
 	initf_bootstage,	/* uses its own timer, so does not need DM */
-	cyclic_init,
 	event_init,
 #ifdef CONFIG_BLOBLIST
 	bloblist_init,
@@ -957,6 +956,16 @@ static const init_fnc_t init_sequence_f[] = {
 	do_elf_reloc_fixups,
 #endif
 	clear_bss,
+	/*
+	 * Deregister all cyclic functions before relocation, so that
+	 * gd->cyclic_list does not contain any references to pre-relocation
+	 * devices. Drivers will register their cyclic functions anew when the
+	 * devices are probed again.
+	 *
+	 * This should happen as late as possible so that the window where a
+	 * watchdog device is not serviced is as small as possible.
+	 */
+	cyclic_unregister_all,
 #if !defined(CONFIG_ARM) && !defined(CONFIG_SANDBOX) && \
 		!CONFIG_IS_ENABLED(X86_64)
 	jump_to_copy,