The 'overrides' here are PACKAGE_ARCH based and hence not overrides
as such and the _append wasn't working in many cases. This adjusts the
code to use PACKAGE_ARCH as the accessor and ensures the variables
work as expected. This fixes various segfaults and ensures postinsts
run at build time rather than on the target system.

The bug was introduced in http://git.yoctoproject.org/cgit.cgi/poky/commit/?id=7efad8a1b56df6ee07c12ad360c0493d7b1d6d23.

Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Signed-off-by: Ross Burton <ross.burton@intel.com>

* in some cases (e.g. with external toolchain which doesn't respect our
  reasonably old version set in OLDEST_KERNEL) it's possible to have libc
  which requires newer kernel than what we have on builders, qemu supports
  -r param to use different uname than what's returned by host system.
* change qemu_run_binary to pass -r ${OLDEST_KERNEL} and add the same to
  QEMU_OPTIONS which are used by qemuwrapper-cross
* maybe we should eventually convert all qemu_run_binary usages always include
  qemuwrapper-cross dependency and always call qemu through that (it
  seems very strange that qemu_target_binary is called from qemuwrapper
  and for allarch recipes it can return qemu-allarch as qemu binary).

  qemu_run_binary is used by:
  meta/classes/gtk-immodules-cache.bbclass:            ${@qemu_run_binary(d, '$D', '${bindir}/gtk-query-immodules-$maj_ver.0')} \
  meta/classes/qemu.bbclass:def qemu_run_binary(data, rootfs_path, binary):
  meta/recipes-core/systemd/systemd_213.bb:         ${@qemu_run_binary(d, '$D', '${base_bindir}/udevadm')} hwdb --update \
  meta/recipes-graphics/pango/pango.inc:    ${@qemu_run_binary(d, '$D','${bindir}/${MLPREFIX}pango-querymodules')} \

  and qemuwrapper directly by:
  scripts/postinst-intercepts/update_font_cache:PSEUDO_UNLOAD=1 qemuwrapper -L $D -E LD_LIBRARY_PATH=$D/${libdir}:$D/${base_libdir}\
  scripts/postinst-intercepts/update_pixbuf_cache:PSEUDO_UNLOAD=1 qemuwrapper -L $D -E LD_LIBRARY_PATH=$D/${libdir}:$D/${base_libdir}\

Signed-off-by: Martin Jansa <Martin.Jansa@gmail.com>
Signed-off-by: Saul Wold <sgw@linux.intel.com>
Signed-off-by: Ross Burton <ross.burton@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>

The QEMU_OPTIONS variables belong in qemu.bbclass so move them there. The
only users of them inherit qemu.bbclass. There is no point in pushing
these into every recipe.

Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>

This reverts commit 9f5a6f89d9f4a6c7bed3b163e6eaa764d762f523.

The reason for reverting this is:
 * qemuwrapper has now a fallback method;
 * when using multilib, calling qemu_target_binary from recipes would
   always point to the qemu binary corresponding to the machine
   architecture. Hence, postinstalls needing to use qemu would call the
   wrong qemu user emulation binary;

Signed-off-by: Laurentiu Palcu <laurentiu.palcu@intel.com>
Signed-off-by: Saul Wold <sgw@linux.intel.com>

Postinstalls that use qemu are throwing a segmentation fault when
building for qemux86-64 on a 64bit host (it might also happen for
qemux86 if building on a 32bit host but I didn't test). It looks like
qemu looks for ld.so.cache which is not found because it is generated
after rootfs_(rpm|ipk|deb)_do_rootfs is called and then it tries to load
libraries from the default paths (which are the host's). In order to
avoid this, pass the LD_LIBRARY_PATH explicitly to the target's dynamic
loader.

Signed-off-by: Laurentiu Palcu <laurentiu.palcu@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>

The intercept scripts fail to run on 32 bit hosts. Apparently, the
current approach worked on 64 bit hosts due to the larger virtual address
space (probably). On 32 bit hosts, however, calling the target binary like:

qemu-arm ld-linux.so --library-path /lib:/usr/lib arm_binary

fails with:

arm_binary: error while loading shared libraries: arm_binary: failed to
map segment from shared object: Operation not permitted

When run like this, qemu-arm fails to map the arm_binary executable in
memory because it's hitting the lower limit of
/proc/sys/vm/mmap_min_addr. That's because it loads the
ld-linux.so binary successfully, taking into account mmap_min_addr, runs
it, and then ld-linux.so will map the arm_binary at a fixed address but this
will fail because it is below mmap_min_addr. The qemu's guest base probing,
apparently, doesn't work fine when a program runs inside other.

One way around this would be to set mmap_min_addr to 0 (on recent
distributions is set to 65536 to avoid "kernel NULL pointer dereference"
defects) but this approach is not safe.

The other way is to call the binary directly but providing qemu with a
prefix (-L option) in order to find the elf interpreter correctly. This
way, both the target binary and dynamic loader are mapped into memory
under qemu's control and, only after, the dynamic loader is started.

[YOCTO #4179]

Signed-off-by: Laurentiu Palcu <laurentiu.palcu@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>

qemu user binaries sometimes segfault when running them through pseudo.
So, set PSEUDO_UNLOAD to 1 before running any qemu binary.

[YOCTO #3788]

Signed-off-by: Laurentiu Palcu <laurentiu.palcu@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>

For example with a lib32 multilib, we need to still use the 64 bit
qemu binary in case we do encounter a 64 bit binary.

Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>

When qemu bbclass is inherited from a recipe that is not architecture
dependent, qemu_run_binary will return "qemu-allarch". However this
binary does not exist. Instead, return "qemuwrapper" which will, in
turn, execute the right binary for the target the image was built for.

[YOCTO #2599]

Signed-off-by: Laurentiu Palcu <laurentiu.palcu@intel.com>
Signed-off-by: Saul Wold <sgw@linux.intel.com>

[YOCTO #3602]

Signed-off-by: Laurentiu Palcu <laurentiu.palcu@intel.com>
Signed-off-by: Saul Wold <sgw@linux.intel.com>