Add 2 seconds sleep after each forcibly killed/tried-to-kill process
in the final process termination loop in sysupgrade stage2.
This is needed especially for qualcommax/ipq807x, where ath11k
wireless driver may have a long 10-20 seconds delay after termination
before actually getting killed. This often breaks sysupgrade.
The current KILL loop in kill_remaining does all 10 kill attempts
consecutively without any delay, as evidenced here in a failing sysupgrade.
It does not allow any time for the process to finalize its internal
termination.
Sat Sep 2 19:05:56 EEST 2023 upgrade: Sending TERM to remaining processes ...
Sat Sep 2 19:05:56 EEST 2023 upgrade: Sending signal TERM to hostapd (2122)
Sat Sep 2 19:05:56 EEST 2023 upgrade: Sending signal TERM to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending KILL to remaining processes ...
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2122)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Sending signal KILL to hostapd (2138)
Sat Sep 2 19:06:00 EEST 2023 upgrade: Failed to kill all processes.
sysupgrade aborted with return code: 256
The change in this commit adds a 2 seconds delay after each kill attempt
in order to allow some processes to more gracefully handle their
internal termination.
The result is like this:
Sun Sep 3 11:15:10 EEST 2023 upgrade: Sending TERM to remaining processes ...
Sun Sep 3 11:15:10 EEST 2023 upgrade: Sending signal TERM to hostapd (2309)
Sun Sep 3 11:15:10 EEST 2023 upgrade: Sending signal TERM to hostapd (2324)
Sun Sep 3 11:15:14 EEST 2023 upgrade: Sending KILL to remaining processes ...
Sun Sep 3 11:15:14 EEST 2023 upgrade: Sending signal KILL to hostapd (2309)
[ 699.827521] br-lan: port 7(hn5wpa2r) entered disabled state
[ 699.908673] device hn5wpa2r left promiscuous mode
[ 699.908721] br-lan: port 7(hn5wpa2r) entered disabled state
[ 701.038029] br-lan: port 6(hn5wpa3) entered disabled state
Sun Sep 3 11:15:16 EEST 2023 upgrade: Sending signal KILL to hostapd (2324)
[ 702.058256] br-lan: port 5(hn2wlan) entered disabled state
[ 709.250063] stage2 (8237): drop_caches: 3
Sun Sep 3 11:15:25 EEST 2023 upgrade: Switching to ramdisk...
The delay introduced here only kicks in if there is some process that
does not get terminated by the first TERM call. Then there is at least
one 2 sec wait after the first KILL loop round.
This commit is related to discussion in PRs #12235 and #12632
Signed-off-by: Hannu Nyman <hannu.nyman@iki.fi>
Reviewed-by: Robert Marko <robimarko@gmail.com>
(cherry picked from commit 47d56ae546)
Some Broadcom MIPS devices require JFFS2 cleanmarkers to be present on the
kernel partition or the bootloader will identify the partition as corrupt and
won't boot the kernel.
Signed-off-by: Álvaro Fernández Rojas <noltari@gmail.com>
(cherry picked from commit 434df8df54)
When using "ubiformat" with stdin it requires passing image size using
the -S argument. Provide it just like we do for "ubiupdatevol".
This fixes:
ubiformat: error!: must use '-S' with non-zero value when reading from stdin
This change fixes sysupgrade for bcm53xx and bcm4908 NAND devices
possibly some other targets too.
Cc: Rodrigo Balerdi <lanchon@gmail.com>
Cc: Daniel Golle <daniel@makrotopia.org>
Fixes: 9710712120 ("base-files: accept gzipped nand sysupgrade images")
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
Acked-by: Daniel Golle <daniel@makrotopia.org>
Tested-by: Koen Vandeputte <koen.vandeputte@citymesh.com>
Starting from Linux Kernel version 6.3 UBI devices will no longer be
considered virtual, but rather have an MTD device parent. Hence they
will no longer be listed under /sys/devices/virtual/ubi which is
used in multiple places in OpenWrt. Prepare for future kernels by
using /sys/class/ubi instead of /sys/devuces/virtual/ubi.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
$BOOTDEV_MAJOR may be empty for many of the uevents parsed in this
function. This condition thus tends to fail benignly (we just skip to
the next device), but it can really clutter the stage2 sysupgrade
stderr, since it looks like the "=" operand doesn't have an appropriate
left-hand argument.
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
This change ensures compatibility with both types of sysupgrade-tar files.
1. For some boards like xiaomi,redmi-router-ax6s, sysupgrade-tar
is pack in directory `vendor,name/`
2. For some boards like xiaomi,mi-router-3g, sysupgrade-tar is pack
in directory `vendor_name/`
Signed-off-by: Chen Minqiang <ptpt52@gmail.com>
emmc_do_upgrade() relies on identify() from the nand.sh upgrade helper.
This only works because FEATURES=emmc targets also tend to include
FEATURES=nand.
Rename identify_magic() to identify_magic_long() to match the common.sh
style and make it clear it pairs with other *_long() variants (and not,
say *_word()).
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
This is used to access footer data in firmare files, and is simpler and
less error-prone than using 'dd' with calculated offsets.
Signed-off-by: Tony Ambardar <itugrok@yahoo.com>
There are some devices putting kernel and rootfs on separated
ubi volumes. To make OpenWrt compatible with their bootloader,
we need to put kernel and rootfs into separated ubi volumes.
Add support for CI_KERN_UBIPART and CI_ROOT_UBIPART for this
situation.
Signed-off-by: Chuanhong Guo <gch981213@gmail.com>
split ubi reformat/attach into nand_attach_ubi in preparation
for reusing this code in other functions.
Signed-off-by: Chuanhong Guo <gch981213@gmail.com>
Several Broadcom targets were using the nand_do_upgrade_success
shell function which has been removed by commit e25e6d8e54
("base-files: fix and clean up nand sysupgrade code"). Refactor the
new nand_do_upgrade to bring back nand_do_upgrade_success with the
behavior expected by those users.
Fixes: e25e6d8e54 ("base-files: fix and clean up nand sysupgrade code")
Reported-by: Chen Minqiang <ptpt52@gmail.com>
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
When firmware images only contained compressed kernels and squashfs roots,
uncompressed tar files were a good option. We are now using UBIFS images,
both raw and tarred, as well as ubinized (full UBI partition) images, all
of which benefit greatly from compression.
For example, a raw ubinized backup taken from a running Askey RT4230W REV6
(such full backups can be restored via the LUCI's sysupgrade UI) is over
400 MB, but compresses to less than 10 MB.
This commit adds support for gzipped versions of all file types already
accepted by the nand sysupgrade mechanism, be them raw or tarred.
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
It has been reported that ubinized nand sysupgrade fails under certain
circumstances, being unable to detach the existing ubi partition due to
volumes within the partition being mounted.
This is an attempt to solve such issues by unmounting and removing
ubiblock devices and unmounting ubi volumes within the target partition
prior to detaching and formatting it.
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
- Never return from 'nand_do_upgrade', not even in case of errors, as that
would cause execution of sysupgrade code not intended for NAND devices.
- Unify handling of sysupgrade success and failure.
- Detect and report more error conditions.
- Fix outdated/incorrect/unclear comments.
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
Downstream projects might re-generate device-specific configuration
based on OpenWrt's defaults on each upgrade, thus being unaffected by
forward- as well as backwards-breaking configuration.
Add a new sysupgrade parameter, which allows sysupgrades between minor
compat-versions. Upgrades will still fail upon mismatching major compat
versions.
Signed-off-by: David Bauer <mail@david-bauer.net>
Commit ecbcc0b595 bricks devices on which the raw kernel and UBI mtd
partitions overlap.
This is the case of the ZyXEL NR7101 for example. Its OEM bootloader has
no UBI support. OpenWrt splits the stock kernel mtd partition into a raw
kernel part used by the bootloader and a UBI part used to store rootfs
and rootfs_data. Running mtd erase on the complete partition during
sysupgrade erases the UBI part and results in a soft brick.
Arguably the best solution would be to fix the partition layouts so that
kernel and UBI partitions do not overlap, also including a stock_kernel
partition to help reverting to stock firmware. This would have the added
benefit of protecting UBI from kernel images that are excessively large.
Fixes: ecbcc0b595 ("base-files: safer sysupgrade.tar for kernel-out-of-UBI")
Reported-by: Bjørn Mork <bjorn@mork.no>
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
Attempt to minimize the time during which an interrupted nand sysupgrade
can lead to a non-functional device by flushing caches before starting
the upgrade procedure.
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
Fix issues while retaining configuration during nand sysupgrade:
- abort configuration saving if data partition is not found
- generate diagnostics if saving fails (eg, because of lack of space)
- do not output "sysupgrade successful" in case of errors
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
Remove redundant check from nand ubinized sysupgrade code. This check
has already been done in the only caller of the affected function:
nand_do_upgrade.
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
Prepares code for ubirename-based safe sysupgrade implementation.
Fixes several issues:
- the special CI_KERNPART value "none" is ignored if an MTD partition
named "none" exists
- misleading variable names (such as has_kernel to mean "tar has kernel
and it should not be written to an MTD partition but a UBI volume")
- inconsistent treatment of zero-length tar member files
- inconsistent meaning of "0" and "" variable values
- redundant operations (unneeded untaring, repeated untaring, unneeded
partition lookups)
- inconsistent variable quoting
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
Ensure that the kernel CRC is invalidated while rootfs is being updated.
This allows the bootloader to detect an interrupted sysupgrade and fall
back to an alternate booting method, such as TFTP, instead of just going
ahead with normal boot and effectively bricking the device.
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
Ensure that the kernel CRC is invalidated while rootfs is being updated.
This allows the bootloader to detect an interrupted sysupgrade and fall
back to an alternate booting method, instead of just going ahead with
normal boot and effectively bricking the device.
Possible fallbacks include a recovery initramfs partition or UBI volume
and TFTP. See here for an example U-Boot configuration with fallbacks:
https://shorturl.at/befsA (https://github.com/Lanchon/openwrt-tr4400-v2/
blob/e7d707d6bd7839fbd0b8d0bd180fce451df77e47/install-recovery.sh#L52-L63)
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
Emit diagnostics if nand sysupgrade is aborted because UBI partition
cannot be attached. Also avoid redudndant checks.
Signed-off-by: Rodrigo Balerdi <lanchon@gmail.com>
Make sure sysupgrade on NAND also works in case of UBI volumes having
index >9. While at it, also make sure UBI device is detected and abort
in case it isn't. Use Shell built-in shorthand ':' instead of 'true'.
Fixes#9708
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
See firmware-utils.git commits [1], which implemented the cros-vbutil
verified-boot payload-packing tool, and extended ptgen for the CrOS
kernel partition type. With these, it's now possible to package kernel +
rootfs to make disk images that can boot a Chrome OS-based system (e.g.,
Chromebooks, or even a few AP models).
Regarding PARTUUID= changes: Chromium bootloaders work well with a
partition number offset (i.e., relative to the kernel partition), so
we'll be using a slightly different root UUID line.
NB: I've made this support specific to ip40xx for now, because I only
plan to support an IPQ4019-based AP that uses a Chromium-based
bootloader, but this image format can be used for essentially any
Chromebook, as well as the Google OnHub, a prior Chromium-based AP using
an IPQ8064 chipset.
[1]
ptgen: add Chromium OS kernel partition support
https://git.openwrt.org/?p=project/firmware-utils.git;a=commit;h=6c95945b5de973026dc6f52eb088d0943efa96bb
cros-vbutil: add Chrome OS vboot kernel-signing utility
https://git.openwrt.org/?p=project/firmware-utils.git;a=commit;h=8e7274e02fdc6f2cb61b415d6e5b2e1c7e977aa1
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
Rootfs overlays get created at a ROOTDEV_OVERLAY_ALIGN (64KiB)
alignment after the rootfs, but emmc_do_upgrade() is assuming
it comes at the very next 512-byte sector.
Suggested-by: Christian Lamparter <chunkeey@gmail.com>
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
(move spaces around, mention fstools' libtoolfs)
Signed-off-by: Christian Lamparter <chunkeey@gmail.com>
Not all targets create /var/lock or touch /var/lock/fw_printenv.lock in
their platform.sh. This is problematic as fw_printenv then fails in
case /var/lock/fw_printenv.lock has not been created by previous calls
to fw_printenv/fw_setenv before sysupgrade is run.
Targets using fw_printenv/fw_setenv during sysupgrade:
* ath79/*
* ipq40xx/*
* ipq806x/*
* kirkwood/*
* layerscape/*
* mediatek/mt7622
* mvebu/*
* ramips/*
* realtek/*
Targets currently using additional steps in /lib/upgrade/platform.sh
to make sure /var/lock/fw_printenv.lock (or at least /var/lock)
actually exists:
* ath79/* (openmesh devices)
* ipq40xx/* (linksys devices)
* ipq806x/* (linksys devices)
* kirkwood/* (linksys devices)
* layerscape/*
* mvebu/cortexa9 (linksys devices)
Given that accessing the U-Boot environment during sysupgrade is not
uncommon and the situation across targets is currently quite diverse,
just make sure both tools as well fw_env.config are always copied to
the ramdisk used for sysupgrade. Also make sure /var/lock always
exists.
This now allows to remove copying of fw_printenv/fw_setenv as well as
fw_env.config, creation of /var/lock or even /var/lock/fw_printenv.lock
from lib/upgrade/platform.sh or files included there.
As the same applies also to 'fwtool' which is used by generic eMMC
sysupgrade, also always copy that to ramdisk.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
For sysupgrade on NAND/UBI devices there is the U-Boot environment
variable rootfs_data_max which can be used to limit the size of the
rootfs_data volume created on sysupgrade.
This stopped working reliable with recent kernels, probably due to a
race condition when reading the number of free erase blocks from sysfs
just after removing a volume.
Change the script to just try creating rootfs_data with the desired
size and retry with maximum size in case that fails. Hence calculating
the available size in the script can be dropped which works around the
problem.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
We were missing (not using) the last sector of each partition,
compared with the output of gparted.
Signed-off-by: Javier Marcet <javier@marcet.info>
[moved the dot]
Signed-off-by: Christian Lamparter <chunkeey@gmail.com>
Adds generic support for sysupgrading on eMMC-based devices.
Provide function emmc_do_upgrade and emmc_copy_config to be used in
/lib/upgrade/platform.sh instead of redundantly implementing the same
logic over and over again.
Similar to generic sysupgrade on NAND, use environment variables
CI_KERNPART, CI_ROOTPART and newly introduce CI_DATAPART to indicate
GPT partition names to be used. On devices with more than one MMC
block device, CI_ROOTDEV can be used to specify the MMC device for
partition name lookups.
Also allow to select block devices directly using EMMC_KERN_DEV,
EMMC_ROOT_DEV and EMMC_DATA_DEV, as using GPT partition names is not
always an option (e.g. when forced to use MBR).
To easily handle writing kernel and rootfs make use of sysupgrade.tar
format convention which is also already used for generic NAND support.
Signed-off-by: Enrico Mioso <mrkiko.rs@gmail.com>
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
CC: Li Zhang <li.zhang@gl-inet.com>
CC: TruongSinh Tran-Nguyen <i@truongsinh.pro>
Currently nand_upgrade_tar() will pass the kernel length
to nand_upgrade_prepare_ubi() in all cases except for when
the kernel is to be installed in a separate partition as a
binary with the MTD tool.
While this is fine for almost all cases newer MikroTik NAND
devices like hAP ac3 require the kernel to be installed as a
UBIFS packed UBI volume in its own partition.
So, since we have a custom recipe to use ubiformat to flash
the kernel in its partition it makes no sense for sysupgrade
to also install the kernel as a UBI volume in the "ubi"
partition as it only wastes space and will never be used.
So, simply check whether CI_KERNPART is set to "none" and
if so unset the "has_kernel" variable which will in turn
prevent the kernel length from being passed on and then
the kernel UBI volume wont be created for no usefull purpose.
The ath79 MikroTik NAND target has been setting CI_KERNPART
to "none" for a while now altough that was not preventing
the kernel to be installed as UBI volume as well.
Signed-off-by: Robert Marko <robimarko@gmail.com>
Simply reading /proc/*/stat as a space-separated string will not work
as the process name may itself contain spaces. Hence we must match on
the '(' and ')' characters around the process name and can then handle
the remaining string as space-separated values.
This fixes shell error messages which have been popping up the console
due to spaces in process names being interpreted as field separators.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
find_mmc_part provides a better alternative and all users of
get_partition_by_name have been removed.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
While an image layout based on MBR and 'bootfs' partition may be easy
to understand for users who are very used to the IBM PC and always have
the option to access the SD card outside of the device (and hence don't
really depend on other recovery methods or dual-boot), in my opinion
it's a dead end for many desirable features on embedded systems,
especially when managed remotely (and hence without an easy option to
access the SD card using another device in case things go wrong, for
example).
Let me explain:
* using a MSDOS/VFAT filesystem to store kernel(s) is problematic, as a
single corruption of the bootfs can render the system into a state
that it no longer boots at all. This makes dual-boot useless, or at
least very tedious to setup with then 2 independent boot partitions
to avoid the single point of failure on a "hot" block (the FAT index
of the boot partition, written every time a file is changed in
bootfs). And well: most targets even store the bootloader environment
in a file in that very same FAT filesystem, hence it cannot be used
to script a reliable dual-boot method (as loading the environment
itself will already fail if the filesystem is corrupted).
* loading the kernel uImage from bootfs and using rootfs inside an
additional partition means the bootloader can only validate the
kernel -- if rootfs is broken or corrupted, this can lead to a reboot
loop, which is often a quite costly thing to happen in terms of
hardware lifetime.
* imitating MBR-boot behavior with a FAT-formatted bootfs partition
(like IBM PC in the 80s and 90s) is just one of many choices on
embedded targets. There are much better options with modern U-Boot
(which is what we use and build from source for all targets booting
off SD cards), see examples in mediatek/mt7622 and mediatek/mt7623.
Hence rename the 'sdcard' feature to 'legacy-sdcard', and prefix
functions with 'legacy_sdcard_' instead of 'sdcard_'.
Tested-by: Stijn Tintel <stijn@linux-ipv6.be>
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Add a generic sdcard upgrade method instead of duplicating code in yet
another target, and add a feature flag to only install this upgrade
method in targets that set this flag. Copied from mvebu.
Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Try umount on device mapper and loop devices still mounted, so the
subsequent call to disactivate all physical volumes and delete all
loop devices is more likely to succeed.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
If the busybox applet losetup was selected, `command -v` selects that
during sysupgrade. As this applet is in another path and doesn't cover
the '-D' option which is used to make sure user-defined loop devices
are no longer active during sysupgrade.
Detect losetup at the path of the full utility to avoid error messages
in case of the busybox applet being selected.
Reported-by: fda77 <fda77@users.noreply.github.com>
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
If one of the programmes is not running, then we see the following
output in the logs.
`killall: telnetd: no process killed`
To ensure that the log is clean, redirect the output to /dev/null
Signed-off-by: Florian Eckert <fe@dev.tdt.de>
The remaining vn calls have been ported to v.
Therefore, these functions are no longer needed and will be removed.
Signed-off-by: Florian Eckert <fe@dev.tdt.de>
The logging output should not only be displayed in the calling shell
session but also in the syslog. A sysupgrade and a configuration
import, export can thus be traced in the syslog.
Signed-off-by: Florian Eckert <fe@dev.tdt.de>
Calling `switch_to_ramfs()` will not copy the gzip executable
(/bin/gzip) to ramfs, but `/bin/zcat` will call `/bin/gzip` when
package gzip is installed, instead of the busybox-supplied zcat.
This will cause `zcat` to fail to find `gzip`, then cause the
sysupgrade to fail. Adding the `busybox` prefix here will solve
the problem.
Signed-off-by: Chuck Fan <fanck0605@qq.com>
bzip2 adds about 8kb of size. For tiny builds it's often disabled.
It's not directly used by stock OpenWrt programs.
Kernel images compressed with bzip2 are also not fully supported.
Signed-off-by: Sergey Ponomarev <stokito@gmail.com>
[fix \ indention]
Signed-off-by: Paul Spooren <mail@aparcar.org>
These processes are managed by procd and set to start again when killed
via the procd instance parameter "respawn" being set during init.
Example:
procd_set_param respawn 3600 1 0
When they are killed manually during sysupgrade,
they are started again in 5 seconds or less, depending on
how the "respawn" parameter is set.
Use procd through ubus to disable the instances that respawn them,
however, allow dnsmasq, netifd, and logd to restart for remote logging.
Properly closing all these processes increases free memory by about 3 MB,
which should help low memory devices upgrade without crashing.
For very low memory devices (set to 32 MB for now)
also kill dnsmasq, netifd, and logd for an additional 3 MB of free memory.
Also, bump sleep values to allow at least 10 seconds
for network interfaces and daemons
to come up after they are killed and restarted
before caches are dropped.
Signed-off-by: Michael Pratt <mcpratt@pm.me>
Remove vn call in favour of v call. This commit serves as preparation
for removing the v function call.
Signed-off-by: Florian Eckert <fe@dev.tdt.de>
[alter slightly to prevent double space after colon]
Signed-off-by: Adrian Schmutzler <freifunk@adrianschmutzler.de>
