| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In kup.h we currently include kup-radix.h for all 64-bit builds, which
includes Book3S and Book3E. The latter doesn't make sense, Book3E
never uses the Radix MMU.
This has worked up until now, but almost by accident, and the recent
uaccess flush changes introduced a build breakage on Book3E because of
the bad structure of the code.
So disentangle things so that we only use kup-radix.h for Book3S. This
requires some more stubs in kup.h and fixing an include in
syscall_64.c.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
|
| |\
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
From Nick's cover letter:
Linux powerpc new system call instruction and ABI
System Call Vectored (scv) ABI
==============================
The scv instruction is introduced with POWER9 / ISA3, it comes with an
rfscv counter-part. The benefit of these instructions is
performance (trading slower SRR0/1 with faster LR/CTR registers, and
entering the kernel with MSR[EE] and MSR[RI] left enabled, which can
reduce MSR updates. The scv instruction has 128 levels (not enough to
cover the Linux system call space).
Assignment and advertisement
----------------------------
The proposal is to assign scv levels conservatively, and advertise
them with HWCAP feature bits as we add support for more.
Linux has not enabled FSCR[SCV] yet, so executing the scv instruction
will cause the kernel to log a "SCV facility unavilable" message, and
deliver a SIGILL with ILL_ILLOPC to the process. Linux has defined a
HWCAP2 bit PPC_FEATURE2_SCV for SCV support, but does not set it.
This change allocates the zero level ('scv 0'), advertised with
PPC_FEATURE2_SCV, which will be used to provide normal Linux system
calls (equivalent to 'sc').
Attempting to execute scv with other levels will cause a SIGILL to be
delivered the same as before, but will not log a "SCV facility
unavailable" message (because the processor facility is enabled).
Calling convention
------------------
The proposal is for scv 0 to provide the standard Linux system call
ABI with the following differences from sc convention[1]:
- LR is to be volatile across scv calls. This is necessary because the
scv instruction clobbers LR. From previous discussion, this should
be possible to deal with in GCC clobbers and CFI.
- cr1 and cr5-cr7 are volatile. This matches the C ABI and would allow
the kernel system call exit to avoid restoring the volatile cr
registers (although we probably still would anyway to avoid
information leaks).
- Error handling: The consensus among kernel, glibc, and musl is to
move to using negative return values in r3 rather than CR0[SO]=1 to
indicate error, which matches most other architectures, and is
closer to a function call.
Notes
-----
- r0,r4-r8 are documented as volatile in the ABI, but the kernel patch
as submitted currently preserves them. This is to leave room for
deciding which way to go with these. Some small benefit was found by
preserving them[1] but I'm not convinced it's worth deviating from
the C function call ABI just for this. Release code should follow
the ABI.
Previous discussions:
https://lists.ozlabs.org/pipermail/linuxppc-dev/2020-April/208691.html
https://lists.ozlabs.org/pipermail/linuxppc-dev/2020-April/209268.html
[1] https://github.com/torvalds/linux/blob/master/Documentation/powerpc/syscall64-abi.rst
[2] https://lists.ozlabs.org/pipermail/linuxppc-dev/2020-April/209263.html
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Add support for the scv instruction on POWER9 and later CPUs.
For now this implements the zeroth scv vector 'scv 0', as identical to
'sc' system calls, with the exception that LR is not preserved, nor
are volatile CR registers, and error is not indicated with CR0[SO],
but by returning a negative errno.
rfscv is implemented to return from scv type system calls. It can not
be used to return from sc system calls because those are defined to
preserve LR.
getpid syscall throughput on POWER9 is improved by 26% (428 to 318
cycles), largely due to reducing mtmsr and mtspr.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Fix ppc64e build]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200611081203.995112-3-npiggin@gmail.com
|
| |/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Before returning to user, if there are missing FP/VEC/VSX bits from the
user MSR then those registers had been saved and must be restored again
before use. restore_math will decide whether to restore immediately, or
skip the restore and let fp/vec/vsx unavailable faults demand load the
registers.
Each time restore_math restores one of the FP/VSX or VEC register sets
is loaded, an 8-bit counter is incremented (load_fp and load_vec). When
these wrap to zero, restore_math no longer restores that register set
until after they are next demand faulted.
It's quite usual for those counters to have different values, so if one
wraps to zero and restore_math no longer restores its registers or user
MSR bit but the other is not zero yet does not need to be restored
(because the kernel is not frequently using the FPU), then restore_math
will be called and it will also not return in the early exit check.
This causes msr_check_and_set to test and set the MSR at every kernel
exit despite having no work to do.
This can cause workloads (e.g., a NULL syscall microbenchmark) to run
fast for a time while both counters are non-zero, then slow down when
one of the counters reaches zero, then speed up again after the second
counter reaches zero. The cost is significant, about 10% slowdown on a
NULL syscall benchmark, and the jittery behaviour is very undesirable.
Fix this by having restore_math test all conditions first, and only
update MSR if we will be loading registers.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200623234139.2262227-2-npiggin@gmail.com
|
| |
|
|
|
|
|
|
|
|
|
| |
The AMR update is made conditional on AMR actually changing, which
should be the less common case on most workloads (though kernel page
faults on uaccess could be frequent, this doesn't significantly slow
down that case).
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200429065654.1677541-4-npiggin@gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The same complicated sequence for juggling EE, RI, soft mask, and
irq tracing is repeated 3 times, tidy these up into one function.
This differs qiute a bit between sub architectures, so this makes
the ppc32 port cleaner as well.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200429062421.1675400-1-npiggin@gmail.com
|
| |
|
|
|
|
|
|
|
|
| |
Any kind of WARN causes a program check that will crash with
unrecoverable exception if it occurs when RI is clear.
Fixes: 68b34588e202 ("powerpc/64/sycall: Implement syscall entry/exit logic in C")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200429065654.1677541-2-npiggin@gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Hugh reported that his trusty G5 crashed after a few hours under load
with an "Unrecoverable exception 380".
The crash is in interrupt_return() where we check lazy_irq_pending(),
which calls get_paca() and with CONFIG_DEBUG_PREEMPT=y that goes to
check_preemption_disabled() via debug_smp_processor_id().
As Nick explained on the list:
Problem is MSR[RI] is cleared here, ready to do the last few things
for interrupt return where we're not allowed to take any other
interrupts.
SLB interrupts can happen just about anywhere aside from kernel
text, global variables, and stack. When that hits, it appears to be
unrecoverable due to RI=0.
The problematic access is in preempt_count() which is:
return READ_ONCE(current_thread_info()->preempt_count);
Because of THREAD_INFO_IN_TASK, current_thread_info() just points to
current, so the access is to somewhere in kernel memory, but not on
the stack or in .data, which means it can cause an SLB miss. If we
take an SLB miss with RI=0 it is fatal.
The easiest solution is to add a version of lazy_irq_pending() that
doesn't do the preemption check and call it from the interrupt return
path.
Fixes: 68b34588e202 ("powerpc/64/sycall: Implement syscall entry/exit logic in C")
Reported-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200502143316.929341-1-mpe@ellerman.id.au
|
| |
|
|
|
|
|
|
|
| |
There are numerous references to 32bit functions in generic and 64bit
code so ifdef them out.
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/e5619617020ef3a1f54f0c076e7d74cb9ec9f3bf.1584699455.git.msuchanek@suse.de
|
| |
|
|
|
|
|
|
|
|
| |
This reconciles interrupts in the system call case like all other
interrupts. This allows system_call_common to be shared with the scv
system call implementation in a subsequent patch.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200225173541.1549955-31-npiggin@gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Implement the bulk of interrupt return logic in C. The asm return code
must handle a few cases: restoring full GPRs, and emulating stack
store.
The stack store emulation is significantly simplfied, rather than
creating a new return frame and switching to that before performing
the store, it uses the PACA to keep a scratch register around to
perform the store.
The asm return code is moved into 64e for now. The new logic has made
allowance for 64e, but I don't have a full environment that works well
to test it, and even booting in emulated qemu is not great for stress
testing. 64e shouldn't be too far off working with this, given a bit
more testing and auditing of the logic.
This is slightly faster on a POWER9 (page fault speed increases about
1.1%), probably due to reduced mtmsrd.
mpe: Includes fixes from Nick for _TIF_EMULATE_STACK_STORE
handling (including the fast_interrupt_return path), to remove
trace_hardirqs_on(), and fixes the interrupt-return part of the
MSR_VSX restore bug caught by tm-unavailable selftest.
mpe: Incorporate fix from Nick:
The return-to-kernel path has to replay any soft-pending interrupts if
it is returning to a context that had interrupts soft-enabled. It has
to do this carefully and avoid plain enabling interrupts if this is an
irq context, which can cause multiple nesting of interrupts on the
stack, and other unexpected issues.
The code which avoided this case got the soft-mask state wrong, and
marked interrupts as enabled before going around again to retry. This
seems to be mostly harmless except when PREEMPT=y, this calls
preempt_schedule_irq with irqs apparently enabled and runs into a BUG
in kernel/sched/core.c
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200225173541.1549955-29-npiggin@gmail.com
|
|
|
System call entry and particularly exit code is beyond the limit of
what is reasonable to implement in asm.
This conversion moves all conditional branches out of the asm code,
except for the case that all GPRs should be restored at exit.
Null syscall test is about 5% faster after this patch, because the
exit work is handled under local_irq_disable, and the hard mask and
pending interrupt replay is handled after that, which avoids games
with MSR.
mpe: Includes subsequent fixes from Nick:
This fixes 4 issues caught by TM selftests. First was a tm-syscall bug
that hit due to tabort_syscall being called after interrupts were
reconciled (in a subsequent patch), which led to interrupts being
enabled before tabort_syscall was called. Rather than going through an
un-reconciling interrupts for the return, I just go back to putting
the test early in asm, the C-ification of that wasn't a big win
anyway.
Second is the syscall return _TIF_USER_WORK_MASK check would go into
an infinite loop if _TIF_RESTORE_TM became set. The asm code uses
_TIF_USER_WORK_MASK to brach to slowpath which includes
restore_tm_state.
Third is system call return was not calling restore_tm_state, I missed
this completely (alhtough it's in the return from interrupt C
conversion because when the asm syscall code encountered problems it
would branch to the interrupt return code.
Fourth is MSR_VEC missing from restore_math, which was caught by
tm-unavailable selftest taking an unexpected facility unavailable
interrupt when testing VSX unavailble exception with MSR.FP=1
MSR.VEC=1. Fourth case also has a fixup in a subsequent patch.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200225173541.1549955-26-npiggin@gmail.com
|
