On Mon, Mar 16, 2015 at 3:53 PM, Maxime Bizon <mbizon@xxxxxxxxxx> wrote: > > On Thu, 2015-03-12 at 17:00 +0100, Nicolas Schichan wrote: > >> When called from prom init code, bcm63xx_gpio_init() will fail as it >> will call gpiochip_add() which relies on a working kmalloc() to alloc >> the gpio_desc array and kmalloc is not useable yet at prom init time. >> >> Move bcm63xx_gpio_init() to bcm63xx_register_devices() (an >> arch_initcall) where kmalloc works. > > no that patch is completely bogus: > > 1) bcm63xx_gpio_init() does more than registering the gpio_chip: look at > bcm63xx_gpio_out_low_reg_init(). > > We want at least the low lever helpers bcm_gpio_readl()/writel() to work > early. bcm_gpio_readl/bcm_gpio_writel() are completely unaffected by anything done by bcm63xx_gpio_out_low_reg_init(), all they do is access the gpio register space. And all of the functions using the gpio_out_low_reg are static and only accessible through the registered gpio chip, so there can't be any rouge accesses before the chip is registered. > 2) look at board_register_devices() in board_bcm963xx.c, it uses the > gpio API, but is called during arch_initcall() (there was an attempt to > move it later, but it has been reverted) > > so you cannot move that gpiochip registration later as-is, more > refactoring and *testing* is required. The only access done in board_register_devices() is the ephy-reset gpio, and that is still done *after* moving bcm63xx_gpio_init() to register_devices(). So I don't see how this breaks anything. But for the sake of the argument, let's give it a spin: root@OpenWrt:/# uname -a Linux OpenWrt 4.0.0-rc1+ #383 SMP Mon Mar 16 16:46:49 CET 2015 mips GNU/Linux root@OpenWrt:/# cat /sys/kernel/debug/gpio GPIOs 0-39, bcm63xx-gpio: gpio-4 (power ) out hi gpio-5 (stop ) out lo gpio-15 (adsl-fail ) out hi gpio-22 (ppp ) out hi gpio-23 (ppp-fail ) out hi gpio-36 (ephy-reset ) out hi Everything seems to work fine. Jonas
