powerpc/tm: Fix restoring FP/VMX facility incorrectly on interrupts
When in userspace and MSR FP=0 the hardware FP state is unrelated to
the current process. This is extended for transactions where if tbegin
is run with FP=0, the hardware checkpoint FP state will also be
unrelated to the current process. Due to this, we need to ensure this
hardware checkpoint is updated with the correct state before we enable
FP for this process.
Unfortunately we get this wrong when returning to a process from a
hardware interrupt. A process that starts a transaction with FP=0 can
take an interrupt. When the kernel returns back to that process, we
change to FP=1 but with hardware checkpoint FP state not updated. If
this transaction is then rolled back, the FP registers now contain the
wrong state.
The process looks like this:
Userspace: Kernel
Start userspace with MSR FP=0 TM=1
< -----
...
tbegin
bne Hardware interrupt ---- > <do_IRQ...> .... ret_from_except restore_math()
/* sees FP=0 */ restore_fp() tm_active_with_fp()
/* sees FP=1 (Incorrect) */ load_fp_state() FP = 0 -> 1
< ----- Return to userspace with MSR TM=1 FP=1 with junk in the FP TM checkpoint
TM rollback
reads FP junk
When returning from the hardware exception, tm_active_with_fp() is
incorrectly making restore_fp() call load_fp_state() which is setting
FP=1.
The fix is to remove tm_active_with_fp().
tm_active_with_fp() is attempting to handle the case where FP state
has been changed inside a transaction. In this case the checkpointed
and transactional FP state is different and hence we must restore the
FP state (ie. we can't do lazy FP restore inside a transaction that's
used FP). It's safe to remove tm_active_with_fp() as this case is
handled by restore_tm_state(). restore_tm_state() detects if FP has
been using inside a transaction and will set load_fp and call
restore_math() to ensure the FP state (checkpoint and transaction) is
restored.
This is a data integrity problem for the current process as the FP
registers are corrupted. It's also a security problem as the FP
registers from one process may be leaked to another.
Similarly for VMX.
A simple testcase to replicate this will be posted to
tools/testing/selftests/powerpc/tm/tm-poison.c
powerpc/tm: Fix FP/VMX unavailable exceptions inside a transaction
When we take an FP unavailable exception in a transaction we have to
account for the hardware FP TM checkpointed registers being
incorrect. In this case for this process we know the current and
checkpointed FP registers must be the same (since FP wasn't used
inside the transaction) hence in the thread_struct we copy the current
FP registers to the checkpointed ones.
This copy is done in tm_reclaim_thread(). We use thread->ckpt_regs.msr
to determine if FP was on when in userspace. thread->ckpt_regs.msr
represents the state of the MSR when exiting userspace. This is setup
by check_if_tm_restore_required().
Unfortunatley there is an optimisation in giveup_all() which returns
early if tsk->thread.regs->msr (via local variable `usermsr`) has
FP=VEC=VSX=SPE=0. This optimisation means that
check_if_tm_restore_required() is not called and hence
thread->ckpt_regs.msr is not updated and will contain an old value.
This can happen if due to load_fp=255 we start a userspace process
with MSR FP=1 and then we are context switched out. In this case
thread->ckpt_regs.msr will contain FP=1. If that same process is then
context switched in and load_fp overflows, MSR will have FP=0. If that
process now enters a transaction and does an FP instruction, the FP
unavailable will not update thread->ckpt_regs.msr (the bug) and MSR
FP=1 will be retained in thread->ckpt_regs.msr. tm_reclaim_thread()
will then not perform the required memcpy and the checkpointed FP regs
in the thread struct will contain the wrong values.
The code path for this happening is:
Userspace: Kernel Start userspace with MSR FP/VEC/VSX/SPE=0 TM=1 < -----
...
tbegin
bne
fp instruction FP unavailable ---- > fp_unavailable_tm() tm_reclaim_current() tm_reclaim_thread() giveup_all()
return early since FP/VMX/VSX=0
/* ckpt MSR not updated (Incorrect) */ tm_reclaim()
/* thread_struct ckpt FP regs contain junk (OK) */ /* Sees ckpt MSR FP=1 (Incorrect) */
no memcpy() performed
/* thread_struct ckpt FP regs not fixed (Incorrect) */ tm_recheckpoint()
/* Put junk in hardware checkpoint FP regs */ .... < ----- Return to userspace with MSR TM=1 FP=1 with junk in the FP TM checkpoint
TM rollback
reads FP junk
This is a data integrity problem for the current process as the FP
registers are corrupted. It's also a security problem as the FP
registers from one process may be leaked to another.
This patch moves up check_if_tm_restore_required() in giveup_all() to
ensure thread->ckpt_regs.msr is updated correctly.
A simple testcase to replicate this will be posted to
tools/testing/selftests/powerpc/tm/tm-poison.c
== Comment: #0 - Michael Ranweiler <email address hidden> - 2019-09-11 00:49:28 ==
There are two problems/CVEs for power that we'd appreciate adding:
https:/ /git.kernel. org/pub/ scm/linux/ kernel/ git/torvalds/ linux.git/ commit/ ?id=a8318c13e79 badb92bc6640704 a64cc022a6eb97
powerpc/tm: Fix restoring FP/VMX facility incorrectly on interrupts
When in userspace and MSR FP=0 the hardware FP state is unrelated to
the current process. This is extended for transactions where if tbegin
is run with FP=0, the hardware checkpoint FP state will also be
unrelated to the current process. Due to this, we need to ensure this
hardware checkpoint is updated with the correct state before we enable
FP for this process.
Unfortunately we get this wrong when returning to a process from a
hardware interrupt. A process that starts a transaction with FP=0 can
take an interrupt. When the kernel returns back to that process, we
change to FP=1 but with hardware checkpoint FP state not updated. If
this transaction is then rolled back, the FP registers now contain the
wrong state.
The process looks like this:
Userspace: Kernel
< -----
...
tbegin
bne
/* sees FP=0 */
/* sees FP=1 (Incorrect) */
< -----
TM rollback
reads FP junk
When returning from the hardware exception, tm_active_with_fp() is
incorrectly making restore_fp() call load_fp_state() which is setting
FP=1.
The fix is to remove tm_active_ with_fp( ).
tm_active_with_fp() is attempting to handle the case where FP state
has been changed inside a transaction. In this case the checkpointed
and transactional FP state is different and hence we must restore the
FP state (ie. we can't do lazy FP restore inside a transaction that's
used FP). It's safe to remove tm_active_with_fp() as this case is
handled by restore_tm_state(). restore_tm_state() detects if FP has
been using inside a transaction and will set load_fp and call
restore_math() to ensure the FP state (checkpoint and transaction) is
restored.
This is a data integrity problem for the current process as the FP
registers are corrupted. It's also a security problem as the FP
registers from one process may be leaked to another.
Similarly for VMX.
A simple testcase to replicate this will be posted to selftests/ powerpc/ tm/tm-poison. c
tools/testing/
This fixes CVE-2019-15031.
Fixes: a7771176b439 ("powerpc: Don't enable FP/Altivec if not checkpointed") /git.kernel. org/pub/ scm/linux/ kernel/ git/torvalds/ linux.git/ commit/ ?id=8205d5d98ef 7f155de211f5e2e b6ca03d95a5a60
Cc: <email address hidden> # 4.15+
Signed-off-by: Gustavo Romero <email address hidden>
Signed-off-by: Michael Neuling <email address hidden>
Signed-off-by: Michael Ellerman <email address hidden>
Link: https://<email address hidden>
.
2.
https:/
powerpc/tm: Fix FP/VMX unavailable exceptions inside a transaction
When we take an FP unavailable exception in a transaction we have to
account for the hardware FP TM checkpointed registers being
incorrect. In this case for this process we know the current and
checkpointed FP registers must be the same (since FP wasn't used
inside the transaction) hence in the thread_struct we copy the current
FP registers to the checkpointed ones.
This copy is done in tm_reclaim_ thread( ). We use thread- >ckpt_regs. msr >ckpt_regs. msr tm_restore_ required( ).
to determine if FP was on when in userspace. thread-
represents the state of the MSR when exiting userspace. This is setup
by check_if_
Unfortunatley there is an optimisation in giveup_all() which returns regs->msr (via local variable `usermsr`) has tm_restore_ required( ) is not called and hence >ckpt_regs. msr is not updated and will contain an old value.
early if tsk->thread.
FP=VEC=VSX=SPE=0. This optimisation means that
check_if_
thread-
This can happen if due to load_fp=255 we start a userspace process >ckpt_regs. msr will contain FP=1. If that same process is then >ckpt_regs. msr (the bug) and MSR >ckpt_regs. msr. tm_reclaim_thread()
with MSR FP=1 and then we are context switched out. In this case
thread-
context switched in and load_fp overflows, MSR will have FP=0. If that
process now enters a transaction and does an FP instruction, the FP
unavailable will not update thread-
FP=1 will be retained in thread-
will then not perform the required memcpy and the checkpointed FP regs
in the thread struct will contain the wrong values.
The code path for this happening is:
Userspace: Kernel
Start userspace
with MSR FP/VEC/VSX/SPE=0 TM=1
< -----
FP unavailable
---- >
fp_ unavailable_ tm()
tm_reclaim_ current( )
tm_reclaim_ thread( )
giveup_ all()
tm_ reclaim( )
/* Sees ckpt MSR FP=1 (Incorrect) */
tm_recheckpoint ()
. ...
< -----
Return to userspace
with MSR TM=1 FP=1
with junk in the FP TM checkpoint
...
tbegin
bne
fp instruction
return early since FP/VMX/VSX=0
/* ckpt MSR not updated (Incorrect) */
/* thread_struct ckpt FP regs contain junk (OK) */
no memcpy() performed
/* thread_struct ckpt FP regs not fixed (Incorrect) */
/* Put junk in hardware checkpoint FP regs */
TM rollback
reads FP junk
This is a data integrity problem for the current process as the FP
registers are corrupted. It's also a security problem as the FP
registers from one process may be leaked to another.
This patch moves up check_if_ tm_restore_ required( ) in giveup_all() to >ckpt_regs. msr is updated correctly.
ensure thread-
A simple testcase to replicate this will be posted to selftests/ powerpc/ tm/tm-poison. c
tools/testing/
Similarly for VMX.
This fixes CVE-2019-15030.
Fixes: f48e91e87e67 ("powerpc/tm: Fix FP and VMX register corruption")
Cc: <email address hidden> # 4.12+
Signed-off-by: Gustavo Romero <email address hidden>
Signed-off-by: Michael Neuling <email address hidden>
Signed-off-by: Michael Ellerman <email address hidden>
Link: https://<email address hidden>