Re: libffi 3.3 release candidate 1
Hi,
On 10/25/19 10:04 PM, Andreas Schwab wrote:
On Okt 25 2019, John Paul Adrian Glaubitz wrote:
On 10/25/19 8:32 PM, Andreas Schwab wrote:
On Okt 25 2019, John Paul Adrian Glaubitz wrote:
It's qemu-user that we are using here, not qemu-system. qemu-user is more finicky
when it comes to running testsuites.
There is nothing finicky about this testsuite. It's pure user-space.
Yes, but qemu-user has known bugs
If it cannot even get the user-space ABI right it is useless.
Qemu-user was used for building Nokia Internet tablet + phone products
(Debian derivative) Maemo & Meego OSes (over a decade ago), until
Nokia shut down its Linux efforts (and eventually sold its phone
business to MicroSoft). I think user-space Qemu has in general been
useful for this kind of work quite long time.
Problems can be categorized to few classes:
1. CPU emulation bugs
- get more rare as emulation matures (new bugs may sometimes
be found e.g. with new compiler versions)
- obviously must be fixed for emu to be useful/trustworthy
2. User-space ABI translation
- rare, but expected to happen occasionally, as kernel & libc change
(kernel ABI extensions, libc starting to use something that wasn't
yet translated fully etc)
- should be easy to catch and in most cases straightforward to fix
3. Threading and other issues related to emulating just user-space
- not all issues are fixable
- AFAIK mainly an issue with some test-suites
- those packages just need to be built with system emulator.
User-space qemu is used for performance reasons, doing building
with magnitude slower system emulator is not too bad, as long as
majority of packages build fine with user-space qemu, and system
emulator can emulate large enough amount of RAM etc
An obvious conclusion from these is that any bugs with user-space
qemu need to be vetted for whether they belong to above categories,
before forwarding them to package maintainers.
- Eero
Reply to: