| Age | Commit message (Collapse) | Author | Files |
|---|
|
Append to the value with appendVarFlag() instead of setting it outright,
so that we can also append to it in other places. Accordingly, this
varflag is pipe-separated (since we want to be able to exclude any
string fragment, in this case including the leading space), thus put a
leading pipe character to play nicely with any existing value.
Signed-off-by: Paul Eggleton <paul.eggleton@linux.intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
Single long commands are difficult to read and maintain. Split it to make
it more human-friendly.
Signed-off-by: Leonardo Sandoval <leonardo.sandoval.gonzalez@linux.intel.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
As of this commit:
39f5a05152aa0c3503735e18dd3b4c066b284107
patchelf no longer inflates file sizes. Since the files are no longer
inflated by patchelf, we can skip using cp with the --sparse option.
More details as to how patchelf has changed are available in that
commit log.
Signed-off-by: Stephano Cetola <stephano.cetola@linux.intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
When we switched to recipe specific sysroots (rss), performance took a nose dive. Its
easy to blame rss but it turns out not to be entirely at fault.
Three configurations are compared here:
a) Pre-RSS (revision 45df694a9f472ac2f684aadac4d864c3dfdc48a7)
b) Post-RSS (revision 226a508da955439b881b2f0a544a3aee76e59919)
c) as b) with this change
Overall build times:
a) 22794.25user 2687.88system 30:32.84elapsed 1390%CPU (0avgtext+0avgdata 919056maxresident)k
b) 22677.25user 3238.79system 36:16.68elapsed 1190%CPU (0avgtext+0avgdata 918896maxresident)k
c) 23571.84user 3383.65system 31:36.83elapsed 1421%CPU (0avgtext+0avgdata 919068maxresident)k
For the overall build and sstate directories, du -s shows:
a)
3992588 build-pre-rss/sstate-cache
30804484 build-pre-rss/tmp
b)
4013272 build-with-rss/sstate-cache
36519084 build-with-rss/tmp
c)
4014744 build-with-rss2/sstate-cache
35336960 build-with-rss2/tmp
However more worryingly:
$ du -s build-pre-rss/tmp/sysroots/
2506092 build-pre-rss/tmp/sysroots/
$ du -s build-with-rss/tmp/sysroots-components/
3790712 build-with-rss/tmp/sysroots-components/
$ du -s build-with-rss2/tmp/sysroots-components/
2467544 build-with-rss2/tmp/sysroots-components/
These numbers *should* be equivalent but as you can see, b) is ~1.2GB larger. The reason turned out
to be patchelf. Taking a specific binary from a specific recipe, bc from bc-native, in a) its 82kb
(stripped) yet in b) its 2.17MB.
$ ./patchelf --set-interpreter /bin/rp bc
warning: working around a Linux kernel bug by creating a hole of 2084864 bytes in ‘bc’
https://github.com/NixOS/patchelf/blob/master/src/patchelf.cc#L710 shows that this "hole" is just
padded zeros using memset, its not a proper sparse hole.
This patch copies files with cp --sparse=always after modifying them with patchelf, then replacing
the original file. The better fix will be to fix this in patchself itself and seek() there
when writing the new file but that means new uninative tarballs and will take a bit of work
so I'm proposing this workaround in the meantime.
Also, this patch drops error handling since subprocess check_output() tracebacks will print this
information if the command fails so we can simplify the code.
|
|
If you use subprocess.check_output() the traceback will contain the output
when the command fails which is very useful for debugging. There is no
good reason not to use this everywhere.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
This patch is comparatively large and invasive. It does only do one thing, switching the
system to build using recipe specific sysroots and where changes could be isolated from it,
that has been done.
With the current single sysroot approach, its possible for software to find things which
aren't in their dependencies. This leads to a determinism problem and is a growing issue in
several of the market segments where OE makes sense. The way to solve this problem for OE is
to have seperate sysroots for each recipe and these will only contain the dependencies for
that recipe.
Its worth noting that this is not task specific sysroots and that OE's dependencies do vary
enormously by task. This did result in some implementation challenges. There is nothing stopping
the implementation of task specific sysroots at some later point based on this work but
that as deemed a bridge too far right now.
Implementation details:
* Rather than installing the sysroot artefacts into a combined sysroots, they are now placed in
TMPDIR/sysroot-components/PACKAGE_ARCH/PN.
* WORKDIR/recipe-sysroot and WORKDIR/recipe-sysroot-native are built by hardlinking in files
from the sysroot-component trees. These new directories are known as RECIPE_SYSROOT and
RECIPE_SYSROOT_NATIVE.
* This construction is primarily done by a new do_prepare_recipe_sysroot task which runs
before do_configure and consists of a call to the extend_recipe_sysroot function.
* Other tasks need things in the sysroot before/after this, e.g. do_patch needs quilt-native
and do_package_write_deb needs dpkg-native. The code therefore inspects the dependencies
for each task and adds extend_recipe_sysroot as a prefunc if it has populate_sysroot
dependencies.
* We have to do a search/replace 'fixme' operation on the files installed into the sysroot to
change hardcoded paths into the correct ones. We create a fixmepath file in the component
directory which lists the files which need this operation.
* Some files have "postinstall" commands which need to run against them, e.g. gdk-pixbuf each
time a new loader is added. These are handled by adding files in bindir with the name
prefixed by "postinst-" and are run in each sysroot as its created if they're present.
This did mean most sstate postinstalls have to be rewritten but there shouldn't be many of them.
* Since a recipe can have multiple tasks and these tasks can run against each other at the same
time we have to have a lock when we perform write operations against the sysroot. We also have
to maintain manifests of what we install against a task checksum of the dependency. If the
checksum changes, we remove its files and then add the new ones.
* The autotools logic for filtering the view of m4 files is no longer needed (and was the model
for the way extend_recipe_sysroot works).
* For autotools, we used to build a combined m4 macros directory which had both the native and
target m4 files. We can no longer do this so we use the target sysroot as the default and add
the native sysroot as an extra backup include path. If we don't do this, we'd have to build
target pkg-config before we could built anything using pkg-config for example (ditto gettext).
Such dependencies would be painful so we haven't required that.
* PKDDATA_DIR was moved out the sysroot and works as before using sstate to build a hybrid copy
for each machine. The paths therefore changed, the behaviour did not.
* The ccache class had to be reworked to function with rss.
* The TCBOOTSTRAP sysroot for compiler bootstrap is no longer needed but the -initial data
does have to be filtered out from the main recipe sysroots. Putting "-initial" in a normal
recipe name therefore remains a bad idea.
* The logic in insane needed tweaks to deal with the new path layout, as did the debug source
file extraction code in package.bbclass.
* The logic in sstate.bbclass had to be rewritten since it previously only performed search and
replace on extracted sstate and we now need this to happen even if the compiled path was
"correct". This in theory could cause a mild performance issue but since the sysroot data
was the main data that needed this and we'd have to do it there regardless with rss, I've opted
just to change the way the class for everything. The built output used to build the sstate output
is now retained and installed rather than deleted.
* The search and replace logic used in sstate objects also seemed weak/incorrect and didn't hold
up against testing. This has been rewritten too. There are some assumptions made about paths, we
save the 'proper' search and replace operations to fixmepath.cmd but then ignore this. What is
here works but is a little hardcoded and an area for future improvement.
* In order to work with eSDK we need a way to build something that looks like the old style sysroot.
"bitbake build-sysroots" will construct such a sysroot based on everything in the components
directory that matches the current MACHINE. It will allow transition of external tools and can
built target or native variants or both. It also supports a clean task. I'd suggest not relying on
this for anything other than transitional purposes though. To see XXX in that sysroot, you'd have
to have built that in a previous bitbake invocation.
* pseudo is run out of its components directory. This is fine as its statically linked.
* The hacks for wayland to see allarch dependencies in the multilib case are no longer needed
and can be dropped.
* wic needed more extensive changes to work with rss and the fixes are in a separate commit series
* Various oe-selftest tweaks were needed since tests did assume the location to binaries and the
combined sysroot in several cases.
* Most missing dependencies this work found have been sent out as separate patches as they were found
but a few tweaks are still included here.
* A late addition is that extend_recipe_sysroot became multilib aware and able to populate multilib
sysroots. I had hoped not to have to add that complexity but the meta-environment recipe forced my
hand. That implementation can probably be neater but this is on the list of things to cleanup later
at this point.
In summary, the impact people will likely see after this change:
* Recipes may fail with missing dependencies, particularly native tools like gettext-native,
glib-2.0-native and libxml2.0-native. Some hosts have these installed and will mask these errors
* Any recipe/class using SSTATEPOSTINSTFUNCS will need that code rewriting into a postinst
* There was a separate patch series dealing with roots postinst native dependency issues. Any postinst
which expects native tools at rootfs time will need to mark that dependency with PACKAGE_WRITE_DEPS.
There could well be other issues. This has been tested repeatedly against our autobuilders and oe-selftest
and issues found have been fixed. We believe at least OE-Core is in good shape but that doesn't mean
we've found all the issues.
Also, the logging is a bit chatty at the moment. It does help if something goes wrong and goes to the
task logfiles, not the console so I've intentionally left this like that for now. We can turn it down
easily enough in due course.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
There were a few straggling expansion parameter removals left for
getVar/getVarFlag where the odd whitespace meant they were missed
on previous passes. There were also some plain broken ussages such
as:
d.getVar('ALTERNATIVE_TARGET', old_name, True)
path = d.getVar('PATH', d, True)
d.getVar('IMAGE_ROOTFS', 'True')
which I've corrected (they happend to work by luck).
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
This means that a user can change TMPDIR in a multiconfig situation
and still only have one path to the uninative setup. Without this change
its not possile to make such a setup work.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
getVar() now defaults to expanding by default, thus remove the True
option from getVar() calls with a regex search and replace.
Search made with the following regex: getVar ?\(( ?[^,()]*), True\)
Signed-off-by: Joshua Lock <joshua.g.lock@intel.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
Some c++ libraries fail to build if uninative is built
with gcc 5.x and host gcc version is either 4.8 or 4.9.
The issue should be solved by making separate uninative sstate
directory structure sstate-cache/universal-<gcc version> for host gcc
versions 4.8 and 4.9. This causes rebuilds of uninative if host gcc
is either 4.8 or 4.9 and it doesn't match gcc version used to build
uninative.
[YOCTO #10441]
Signed-off-by: Ed Bartosh <ed.bartosh@linux.intel.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
Create a common include file for compiler flags which allow native binaries to
be interoperable on a wide range of hosts. In particular the C++ ABI is
problematic so choose the CXX11 version to allow interoperation between gcc4 and
gcc5 based hosts. Moving this to a common include instead of uninative.bbclass
allows uninative to be configured later and used in the eSDK (where its
mandatory) even if the base configuration doesn't enable uninative by default
(e.g. nodistro in OE-Core).
[ YOCTO #10645 ]
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
There are various concerns about md5 so use sha256 instead.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
patchelf will understandably error out if there isn't a .interp section to
relocate, so don't try to relocate static binaries.
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
If no -std= option is passed to icu's configure, it defaults to CXX11.
This isn't what we want for uninative, so pass an explicit option
which selects an older ABI on newer versions of g++.
This avoids the __cxa_bad_array_new_length@CXXABI_1.3.8 symbol
being used.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
uninative hand codes the list of files which need relocation, add the libc
to that list to ensure GCONF_PATH is updated.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
We need to be able to update uninative if the version changes. To do this,
stash a checksum of the installed uninative tarball into a file. If this
changes, we update uninative.
For cleaner download messages, we place the tarballs into directories
based on the checksum.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
Move duplicate code into a common function
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
We may see binaries built with gcc5 run or linked into gcc4 environment
so use the older libstdc++ standard for now until we don't support gcc4
on the host system. https://wiki.debian.org/GCC5 has more details about this.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
The previous attempt at soft-failing when uninative was enabled didn't actually
work, because the workers didn't evaluate the function that actually enabled
uninative.
In a BuildStarted handler we can check if we need to download or extract the
uninative tarball.
In a ConfigParsed handler on the workers we can check if the uninative loader is
present, and if so enable it.
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
Instead of raising a generic Exception that can't be handled specifically, raise
a ValueError. Also update the callers so any unexpected exceptions are not
ignored.
Also, rename isBigEngian() to isBigEndian().
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
When patchelf-uninative fails, reporting only the exit code
as done by subprocess.check_call() is not enough to understand
the problem. We also need to capture and report the output
of the command.
Signed-off-by: Patrick Ohly <patrick.ohly@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
We need to improve the error handling here, things were breaking and
yet the user wasn't seeing the issues. We need to skip libraries as
we process the files.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
If the local fetcher is used then files are not actually fetched into DL_DIR, so
check if this happened and if required add a symlink to the real file.
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
When uninative was changed to use it's own sysroot the path to patchelf lost
${bindir_native}, so add it back.
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
Currently this code installs into the standard sysroot, however this causes
some conflicts when linking since the linker can look specifically for
versioned .so files (e.g. like libpthreads.so.0). This breaks builds
of util-linux-native for example.
The easiest solution is to install uninative into its own separate sysroot.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
Originally, the idea was that the init environment would handle
fetching or providing the binary shim that uninative needs.
This turns out to be ugly, especially when you consider proxy
environments and so on getting involved. Instead, lets therefore
support our fetcher which already handles all this.
The distro is expected to setup configuration like:
UNINATIVE_URL ?= "http://mydomain/mypath/"
UNINATIVE_CHECKSUM[i586] =
"md5sum1"
UNINATIVE_CHECKSUM[x86_64] = "md5sum2"
and then it should all work if the user inherits the uninative class.
This patch also improves the error handling in the class to give more
user readable error messages.
If the shim binary is already provided, the system will just use that
and ignore the url information.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
|
Previously UNINATIVE_LOADER was always ld-linux-x86-64.so.2. That is
incorrect when the host is 32-bit.
This change also changes to using ?= so the user can override
UNINATIVE_LOADER if so desired.
[YOCTO #8124]
Signed-off-by: Randy Witt <randy.e.witt@linux.intel.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
|
|
These patches are the start of a new idea, a way of allowing a single set of
cross/native sstate to work over mutliple distros, even old ones.
The assumption is that our own C library is basically up to date. We build
and share a small tarball (~2MB) of a prebuilt copy of this along with a
patchelf binary (which sadly is C++ based so libstdc++ is in there). This
tarball can be generated from our usual SDK generation process through
the supplied recipe, uninative-tarball.
At the start of the build, if its not been extracted into the sysroot, this
tarball is extracted there and configured for the specified path.
When we install binaries from a "uninative" sstate feed, we change the
dynamic loader to point at this dynamic loader and C librbary. This works
exactly the same way as our relocatable SDK does. The only real difference
is a switch to use patchelf, so even if the interpreter section is too small,
it can still adjust the binary.
Right now this implements a working proof of concept. If you build the tarball
and place it at the head of the tree (in COREBASE), you can run a build from
sstate and successfully build packages and construct images.
There is some improvement needed, its hardcoded for x86_64 right now, its trivial
to add 32 bit support too. The tarball isn't fetched right now, there is just a
harcoded path assumption and there is no error handling. I haven't figured
out the best delivery mechanism for that yet. BuildStarted is probably not
the right event to hook on either.
I've merged this to illustrate how with a small change, we might make the
native/cross sstate much more reusable and hence improve the accessibility
of lower overhead builds. With this change, its possible the Yocto Project may
be able to support a configured sstate mirror out the box. This also has
positive implications for our developer workflow/SDK improvements.
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
|
