aboutsummaryrefslogtreecommitdiff
path: root/device/cirbuf.c
diff options
context:
space:
mode:
authorLuca Dariz <luca@orpolo.org>2022-06-28 12:10:48 +0200
committerSamuel Thibault <samuel.thibault@ens-lyon.org>2022-08-27 21:13:11 +0200
commita3db37b8837c091f6018257f2de5bb740350f3f5 (patch)
tree686b03952c144196b9411bd8b03d9ea3692a5782 /device/cirbuf.c
parent4cd8d01832be3100b89f9f12a7f828b787e2f901 (diff)
downloadgnumach-a3db37b8837c091f6018257f2de5bb740350f3f5.tar.gz
gnumach-a3db37b8837c091f6018257f2de5bb740350f3f5.tar.bz2
gnumach-a3db37b8837c091f6018257f2de5bb740350f3f5.zip
x86_64: fix exception stack alignment
* i386/i386/pcb.c: - increase alignment of pcb cache to 16 - ensure the stack is properly aligned when switching ktss * i386/i386/thread.h: - add padding tomake iss field end aligned to 16 bytes * i386/i386/trap.c: - ensure the state we get after the trap points to the correct place in the pcb structure When handling exceptions from IA-32e compatibility mode in user space, on a 64-bit kernel, the exception stack where error info is pushed needs to be aligned to 16 bytes (see Intel System Programming guide, $6.14.2) The exception stack frame is set in the middle of pcb->iss, but it's not always 16-byte aligned; to make sure it is, we increase the alignment of the pcb cache and add a padding field in the pcb structure. This issue resulted in a general protection failure due to CS being corrupted after a page fault. The corruption was happening when the exception stack frame was not properly aligned and a page fault happened; the error info was then pushed after re-aligning the stack, so the value of eflags was actually written in CS place and other fields were shifted too. It also makes sense to ensure this by adding two assertions, although these were primarly useful during debug. Signed-off-by: Luca Dariz <luca@orpolo.org> Message-Id: <20220628101054.446126-10-luca@orpolo.org>
Diffstat (limited to 'device/cirbuf.c')
0 files changed, 0 insertions, 0 deletions