From RM#231:

Looking through dmesg to see boot times I found the following things:

A bunch of eth drivers are compiled in. Only smsc95xx needs to be enabled and even that could be a loadable module.
    bcm2708_spi takes 3.5 seconds to load
    sda1 (ext4) takes 9 seconds to load
    wlan0 takes 8 seconds to load
    X (and the fbtft) takes 17 seconds to load

Getting from power on to mounting the SD card takes about 6 seconds on a Pi 2.

[    5.844178] EXT4-fs (mmcblk0p2): recovery complete
[    5.853241] EXT4-fs (mmcblk0p2): mounted filesystem with ordered data mode. Opts: (null)
[    5.853328] VFS: Mounted root (ext3 filesystem) readonly on device 179:2.
[    5.863535] devtmpfs: mounted
[    5.870637] Freeing unused kernel memory: 1024K
[    5.870975] Run /sbin/init as init process

Nearly all of this is getting USB and SD access setup. Then getting the RNG takes another 5 seconds.

[    6.156952] EXT4-fs (mmcblk0p2): re-mounted. Opts: (null)
[    9.468533] FAT-fs (sda1): Volume was not properly unmounted. Some data may be corrupt. Please run fsck.
[    9.486838] FAT-fs (sdb1): Volume was not properly unmounted. Some data may be corrupt. Please run fsck.
[    9.737072] EXT4-fs (mmcblk0p3): mounting ext3 file system using the ext4 subsystem
[    9.758496] EXT4-fs (mmcblk0p3): recovery complete
[    9.758531] EXT4-fs (mmcblk0p3): mounted filesystem with ordered data mode. Opts: (null)
[    9.979002] FAT-fs (mmcblk0p1): Volume was not properly unmounted. Some data may be corrupt. Please run fsck.
[   10.917545] random: crng init done

This can be fixed with preseeding. The rest of the time goes to loading modules. By the time the Bluetooth init script is run,
which is the last one that shows up in dmesg, we're at close to 30 seconds.

[   28.846777] Bluetooth: BNEP socket layer initialized

Things I can do¶

• Call log() in all my custom scripts so I can tell which ones are taking a long time. Or at least call log() from rcS to say which script is being called. That will get more info into dmesg.

• Put network startup in the background. Nothing needs it up front to start, I don't think.

• Add RNG seeding early in the boot sequence.

• I see this with at least some wifi drivers: unknown parameter '11n_disable' ignored. I should take those out from the drivers I've seen that message from (rt2800usb, carl9170).

• There is something happening here that's taking 8 seconds - find out what it is:

[   11.725594] usbcore: registered new interface driver rt2800usb
[   19.922011] snd_bcm2835: module is from the staging directory, the quality is unknown, you have been warned.

I think this is module loading from /etc/modules.conf. I might be able to slurp just the module names from the file and feed them as one request to modprobe.

I feel confident I can get this down to 15 seconds, maybe less, to get to the UI. And with the changes to bootsplash it will seem much quicker.

Module loading was a bottleneck. I fixed the init script to scarf all the module names into a single variable and pass them all at once to modprobe. I also switched from depmod to depmod -A which helped a bunch too. This shaves about 7 seconds off the boot time.

The next biggest bottleneck is mdev. It takes 9 seconds for that to complete. One possible solution is to write a C program that handles all the configuration of usbhandler and friends (from nldev). Still, mdev has to spawn that C program for each device it finds so it's hard to say how much time this would save. I can start by writing usbhandler in C and seeing if that helps.

After a little googling I found that mdev has its own init script that I've been overwriting. If I use it, which has the following:

echo /sbin/mdev >/proc/sys/kernel/hotplug
/sbin/mdev -s # coldplug modules
find /sys/ -name modalias -print0 | xargs -0 sort -u | tr '\n' '\0' | xargs -0 modprobe -abq

then it seems to boot about 6 seconds faster overall, but it also loads a lot more modules (due to that last line). In fact, it loaded the module for the Pi Zero W wifi even though I'd commented it out of the modules.conf and the node came up using that wifi instead of the usb dongle I had connected (which came up as wlan1). So it seems I should drop my mdev script, and possibly drop loading my modules as this takes care of it, and much faster.

Network bring up is about 5.5 seconds. I can put that into the background and hang dependent services off it, like smb.

smb startup, on its own, takes almost 9 seconds. Combined with the network that's almost 15 seconds that could be done in the background so the UI can be up faster.

I commented out all the modules in /etc/modules.conf and tried again. I found the following modules were not loaded automatically by the mdev changes:

snd-bcm2835
spi_bcm2835
uvcvideo

However, there are alot of new drivers loaded. One of them is hid_multitouch. I'm not certain yet but I may need to blacklist that in order for tslib to work. That said, I ran the following and it looks okay:

1. evtest
   No device specified, trying to scan all of /dev/input/event*
   Available devices:
   /dev/input/event0: ILITEK ILITEK-TP Mouse
   /dev/input/event1: HID 6901:2701
   /dev/input/event2: HID 6901:2701 Mouse
   /dev/input/event3: HID 6901:2701 Consumer Control
   /dev/input/event4: HID 6901:2701 System Control
   /dev/input/event5: ILITEK ILITEK-TP

Despite having hid_multitouch loaded, event5 is mapped correctly to the touchscreen and shows up correctly so maybe I don't need to blacklist that driver. More testing is required.

Compressing kernel.img¶

The defconfig for RPi uses gzip for compression and I wasn't even grabbing that image - I was grabbing the uncompressed image.
I've switched to LZ4 compression and am now copying the compressed image to the package tree.

Changes tested, committed and pushed.

Using initramfs¶

The following could be used with a stripped down initramfs to get a bootsplash faster. This would then switch to the animated splash in the init processing.

/etc/init.d/start.sh:
#!/bin/sh
mount -t proc proc /proc
mount -t sysfs sysfs /sys
mount -t devtmpfs devtmpfs /dev
# Mount RFS / do some critical stuff
mount /dev/mmcblk0p1 /media
fbsplash -s /media/splash.ppm -d /dev/fb0
mount -o move /proc /media/proc
mount -o move /sys /media/sys
mount -o move /dev /media/dev
# Switch to production system
exec switch_root /media /linuxrc

Improving RNG setup¶

Getting RNG setup early is partly done, but there is the usual problem of not just seeding, but increasing the count of available entropy. The former is already done. The latter is not, because that requires an ioctl call for RNDADDTOENTCNT.

I can add a C program to piboxlib/ptools that handles this. It's a simply utility for bumping the available entropy - nothing more - using the ioctl. I can then add that to S20urandom (which I'll have to dup in my buildroot skeleton from the buildroot target). And I can generate a seed file with 4096 bits of entropy and add it to the initial build.

But S20urandom will need to change to read the whole 4096 bytes - or however large the entropy pool actually is. Right now it just reads 512 bytes from /etc/random-seed (which is where the new initialized data should go during the build).

• /proc/sys/kernel/random/entropy_avail
• /proc/sys/kernel/random/pool_size
• man urandom

The man page for urandom explains how to do all of this.

urandom updates are ready for testing. I've added a custom S20urandom and generate the seed file in Buildroot's postbuild.sh script. Both were verified with shellcheck.

I still need to write the C program to do the ioctl and place it in libpibox.