Some of our TARGET_* constant definitions for ioctls were
wrong because the ioctl number is based on the sizeof()
the type passed to the TARGET_IO* macros, and we were
passing a host type rather than a target type.
This was originally reported as a bug where the
FS_IOC_{GET,SET}FLAGS ioctls weren't working for 32-bit
arm guests on x86-64 hosts.
I did a quick audit of all the uses of the TARGET_IO* macros
in syscall_defs.h, and:
* FS_IOC_GETFLAGS/SETFLAGS are indeed wrong
* 3 ioctls to do with magtapes are also wrong
* TARGET_FS_IOC_FIEMAP, TARGET_FICLONERANGE and
TARGET_SOUND_MIXER_INFO take a host struct which
is defined such that it's the same size for all archs
* lots and lots of ioctls use 'int', which is OK as for
us 'abi_int' is always 32 bits (it might have different
alignment requirements, but they don't matter for this
purpose)
* TARGET_SIOCPGRP takes a pid_t, which is always 'int'
This patchset fixes the bugs in the first two bullet
points, and leaves everything else alone since it doesn't
actually manifest as wrong behaviour.
Patch 1 is tested by the LTP 'setxattrs3' test case.
There's no LTP test case for the magtape ioctls, so that
patch change is untested.
Peter Maydell (2):
linux-user: Fix target FS_IOC_GETFLAGS and FS_IOC_SETFLAGS numbers
linux-user: Fix TARGET_MTIOCTOP/MTIOCGET/MTIOCPOS values
linux-user/syscall_defs.h | 35 ++++++++++++++++++++++++++++++-----
1 file changed, 30 insertions(+), 5 deletions(-)
--
2.7.4