Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
To:     LKML <linux-kernel@vger.kernel.org>,
        Al Viro <viro@ZenIV.linux.org.uk>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>
From:   Rasmus Villemoes <linux@rasmusvillemoes.dk>
Subject: must populate_rootfs be synchronous?
Message-ID: <19574912-44b4-c1dc-44c3-67309968d465@rasmusvillemoes.dk>
Date:   Wed, 10 Jun 2020 13:43:56 +0200
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101
 Thunderbird/68.8.0
MIME-Version: 1.0
Content-Type: text/plain; charset=windows-1252
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

I have worked on several boards where there is a rather "aggressive"
external (gpio) watchdog, with timeouts between 250 and 800 ms. Combined
with a rather slow CPU and memory, these boards often fail to make it
through populate_rootfs, since unpacking the initramfs is rather
time-consuming, and we haven't gotten to get the watchdog driver handle
the watchdog. [No, GPIO_WATCHDOG_ARCH_INITCALL doesn't help, at least
not for all cases, probably because the device isn't "discovered" until
mpc83xx_declare_of_platform_devices runs, which is at device_initcall time].

So, those boards currently use some rather ugly patches to make them
boot. I'd like to get rid of those.

I assume there's a good reason populate_rootfs runs between fs_initcall
and device_initcalls - perhaps one of the latter wants some firmware or
do a request_module?

But, would it be possible to throw most of populate_rootfs into a work
item, add some globally visible DECLARE_COMPLETION(initramfs_unpacked)
which is complete_all'ed at the end, and then in the (I assume)
relatively few places that might need to look at the filesystem add a
wait_for_completion(initramfs_unpacked) - including of course right
before the console_on_rootfs() call in kernel_init_freeable() (so also
before we start asking whether there is /init or not).

Rasmus


PS: This is the slowness, .7 seconds to unpack 4MB - it got a bit better
when switching to lz4 compression, but still in the few hundreds of
milliseconds range, which is way too much given the external watchdog's
requirements:

 [    0.047970] Trying to unpack rootfs image as initramfs...
 [    0.768516] Freeing initrd memory: 3972K