Bug 12648 - opkg install reads all files on rootfs?
Summary: opkg install reads all files on rootfs?
Status: RESOLVED FIXED
Alias: None
Product: opkg
Classification: Yocto Project Subprojects
Component: opkg (show other bugs)
Version: 0.3.5
Hardware: All Multiple
: Undecided normal
Target Milestone: 0.4.0
Assignee: Alejandro del Castillo
QA Contact:
URL:
Whiteboard:
Depends on:
Blocks:
 
Reported: 2018-04-04 05:14 UTC by Robert Berger
Modified: 2018-11-14 21:57 UTC (History)
2 users (show)

See Also:
OS type for building Yocto: ---
Type of Regression: ---
Verified:
Documentation change: Don't know


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Robert Berger 2018-04-04 05:14:42 UTC
I'm on yocto 2.4.2 (but this problem bothers me for quite some time)

Here's what I do:

I have a rather big image. Something like a core-image-sato-sdk with some extra tings (including sources and debug info as well) and would like to show the usage of opkg.

The rootfs is exported via nfs:

root@multi-v7-ml:~# cat /proc/cmdline 
console=ttymxc1,115200n8 ip=192.168.42.11:192.168.42.1:192.168.42.254:255.255.255.0:imx6q-phytec-mira-rdk-nand:eth0:off nfsroot=192.168.42.1:/opt/poky/mira-rootfs,v3,tcp noinitrd nohlt panic=1 earlyprintk=ttymxc1,115200n8

And the pkg feed is on the same machine exported via a web server.

root@multi-v7-ml:~# opkg --version
opkg version 0.3.5 (libsolv 0.6.28)

root@multi-v7-ml:~# opkg update
Downloading http://192.168.42.1/imx6q-phytec-mira-rdk-nand/ext-sdk-repo/2.4.2/ext-sdk-ipk/all/Packages.gz.
Updated source 'all'.
Downloading http://192.168.42.1/imx6q-phytec-mira-rdk-nand/ext-sdk-repo/2.4.2/ext-sdk-ipk/armv7a-neon/Packages.gz.
Updated source 'armv7a-neon'.
Downloading http://192.168.42.1/imx6q-phytec-mira-rdk-nand/ext-sdk-repo/2.4.2/ext-sdk-ipk/multi_v7_ml/Packages.gz.
Updated source 'multi-v7-ml'.

root@multi-v7-ml:~# opkg info strace-doc
Package: strace-doc
Version: 4.18-r0.0
Status: unknown ok not-installed
Section: doc
Architecture: armv7a-neon
Maintainer: Poky <poky@yoctoproject.org>
MD5Sum: 820391789d9710186c3bee8549415d21
Size: 11184
Filename: strace-doc_4.18-r0.0_armv7a-neon.ipk
Source: strace_4.18.bb
Description: System call tracing tool - Documentation files
 System call tracing tool - Documentation files

root@multi-v7-ml:~# opkg list-installed | grep strace-doc
root@multi-v7-ml:~# opkg list | grep strace-doc
strace-doc - 4.18-r0.0 - System call tracing tool - Documentation files

---- so far so good ---

root@multi-v7-ml:~# time opkg install strace-doc
Installing strace-doc (4.18) on root
Downloading http://192.168.42.1/imx6q-phytec-mira-rdk-nand/ext-sdk-repo/2.4.2/ext-sdk-ipk/armv7a-neon/strace-doc_4.18-r0.0_armv7a-neon.ipk.
Configuring strace-doc.

real    3m41.485s
user    0m15.660s
sys     1m41.225s


which is quite long, I would say.

so I tried strace opkg install strace-doc and see:

...

openat(AT_FDCWD, "///var/lib/opkg/info/bluez5-dbg.list", O_RDONLY|O_LARGEFILE) = 4
fstat64(4, {st_mode=S_IFREG|0644, st_size=31790, ...}) = 0
read(4, "/usr/lib/debug/usr/bin/bluemoon."..., 4096) = 4096
lstat64("/usr/lib/debug/usr/bin/bluemoon.debug", {st_mode=S_IFREG|0755, st_size=81848, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/tools/hciattach.c", {st_mode=S_IFREG|0644, st_size=32144, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/src/shared/ecc.h", {st_mode=S_IFREG|0644, st_size=2443, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/src/shared/gatt-helpers.c", {st_mode=S_IFREG|0644, st_size=33891, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/tools/hciattach.h", {st_mode=S_IFREG|0644, st_size=2379, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/lib/bnep.h", {st_mode=S_IFREG|0644, st_size=3938, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/profiles/input/suspend-none.c", {st_mode=S_IFREG|0644, st_size=1124, ...}) = 0
lstat64("/usr/lib/debug/usr/lib/bluez5/ptest/unit/test-ecc.debug", {st_mode=S_IFREG|0755, st_size=86116, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/src/shared/gatt-helpers.h", {st_mode=S_IFREG|0644, st_size=4436, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/unit/test-avdtp.c", {st_mode=S_IFREG|0644, st_size=41652, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/tools/parser/hidp.c", {st_mode=S_IFREG|0644, st_size=3098, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/plugins/policy.c", {st_mode=S_IFREG|0644, st_size=19928, ...}) = 0
lstat64("/usr/lib/debug/usr/libexec/bluetooth/obexd.debug", {st_mode=S_IFREG|0755, st_size=1147820, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/tools/parser/ericsson.c", {st_mode=S_IFREG|0644, st_size=1388, ...}) = 0
lstat64("/usr/src/debug/bluez5/5.46-r0/bluez-5.46/client/display.c", {st_mode=S_IFREG|0644, st_size=3640, ...}) = 0

...

opkg reads /usr/src/debug and /usr/lib/debug, meaning the source and debug packages.

And it does this each time.

It looks like the bigger the rootfs gets, the longer it takes to install a small package.

Can you please elaborate what's happening? I guess opkg tries to build some database and that's why it checks what's on the rootfs. If that's the case can I make it somehow non-volatile? Or even better pre-build it and add it to the rootfs?

root@multi-v7-ml:~# time opkg install strace-doc
No packages installed or removed.

real    3m30.172s
user    0m16.160s
sys     1m40.425s

It even takes 3m30 not to install anything;)

root@multi-v7-ml:~# opkg info strace-doc
Package: strace-doc
Version: 4.18-r0.0
Status: install ok installed
Section: doc
Architecture: armv7a-neon
Maintainer: Poky <poky@yoctoproject.org>
MD5Sum: 820391789d9710186c3bee8549415d21
Size: 11184
Filename: strace-doc_4.18-r0.0_armv7a-neon.ipk
Source: strace_4.18.bb
Description: System call tracing tool - Documentation files
 System call tracing tool - Documentation files
Installed-Size: 27614
Installed-Time: 1522825261
Comment 1 Alejandro del Castillo 2018-11-13 17:28:19 UTC
Today, I took a look at this and your right, this is a regression from 0.3.4. On commit 492dd1, the file mode (returned by lstat) was added to the list of things to track on a file. However, the mode information is not persisted, so opkg calls lstat on every installed file every time it runs, which is causing the performance slowdown.

Michael: I think we talked about possibly persisting the mode in the *.list file at some point? I remember there were concerns about backwards compatibility. Since the performance hit is too big, we might need to add a check to use the new format if present, the old if not. The drawback would be that tools that read the *.list file might break if we change the format...but I think I am OK with that since the *.list files are a detail implementation of opkg, with no guarantees of it staying the same.
Comment 2 Robert Berger 2018-11-13 18:54:52 UTC
@Alejandro del Castillo:

I don't quite understand what the reading of every file is needed for. Does it build some database?

Also I was wondering if this could be done when we build the image on the host and not on identical targets once on every target.
Comment 3 Alejandro del Castillo 2018-11-13 19:32:31 UTC
The following patchset can give you context: https://groups.google.com/forum/#!topic/opkg-devel/afrFJrAva-I.

The file mode for every file is a piece of metadata that is now needed, but is not being persisted. We should persist that information to avoid the expensive lstat call on every installed file.
Comment 4 Alejandro del Castillo 2018-11-14 20:24:22 UTC
fixed on commit 99c89c64665670a5547294e826bc709e218b1bd3  (Michael Hansen)
Comment 5 Alejandro del Castillo 2018-11-14 20:24:51 UTC
fixed