In case perf IRQ is the highest of the medium-level IRQs, and is alone on its level, it may be treated as NMI: - LOCKLEVEL is defined to be one level less than EXCM level, - IRQ masking never lowers current IRQ level, - new fake exception cause code, EXCCAUSE_MAPPED_NMI is assigned to that IRQ; new second level exception handler, do_nmi, assigned to it handles it as NMI, - atomic operations in configurations without s32c1i still need to mask all interrupts. Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
Double exceptions that happen during register window overflow/underflow are handled in the topmost stack frame, as if it was the only exception that occured. However unaligned access exception handler is special because it needs to analyze instruction that caused the exception, but the userspace instruction that triggered window exception is completely irrelevant. Unaligned data access is rather normal in the generic userspace code, but stack pointer manipulation must always be done by architecture-aware code and thus unaligned stack means a serious problem anyway. Use the default unaligned access handler that raises SIGBUS in case of unaligned access in window overflow/underflow handler. Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
There are two FIXMEs in the double exception handler 'for the extremely
unlikely case'. This case gets hit by gcc during kernel build once in
a few hours, resulting in an unrecoverable exception condition.
Provide missing fixup routine to handle this case. Double exception
literals now need 8 more bytes, add them to the linker script.
Also replace bbsi instructions with bbsi.l as we're branching depending
on 8th and 7th LSB-based bits of exception address.
This may be tested by adding the explicit DTLB invalidation to window
overflow handlers, like the following:
--- a/arch/xtensa/kernel/vectors.S
+++ b/arch/xtensa/kernel/vectors.S
@@ -592,6 +592,14 @@ ENDPROC(_WindowUnderflow4)
ENTRY_ALIGN64(_WindowOverflow8)
s32e a0, a9, -16
+ bbsi.l a9, 31, 1f
+ rsr a0, ccount
+ bbsi.l a0, 4, 1f
+ pdtlb a0, a9
+ idtlb a0
+ movi a0, 9
+ idtlb a0
+1:
l32e a0, a1, -12
s32e a2, a9, -8
s32e a1, a9, -12
Cc: stable@vger.kernel.org
Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
The original implementation could clobber registers under certain conditions. The Xtensa processor architecture uses windowed registers and the original implementation was using a4 as a temporary register, which under certain conditions could be register a0 of the oldest window frame, and didn't always restore the content correctly. By moving the _spill_registers routine inside the fast system call, it frees up one more register (the return address is not required anymore) for the spill routine. Signed-off-by: Chris Zankel <chris@zankel.net>
Based on the SMP patch by Joe Taylor and subsequent fixes. Preserve exception table pointer (normally stored in excsave1 SR) as it cannot be easily restored in SMP environment. Signed-off-by: Max Filippov <jcmvbkbc@gmail.com> Signed-off-by: Chris Zankel <chris@zankel.net>
IRQ handlers are expected to run with IRQs disabled. See e.g. http://lwn.net/Articles/380931/ for a longer story. This was overlooked in the commit 2d1c645 xtensa: dispatch medium-priority interrupts Revert to old behavior and simplify interrupt entry and exit code. Interrupt handler still honours IRQ priority. do_notify_resume/schedule must be called with interrupts enabled, enable interrupts if we return from user exception. Signed-off-by: Max Filippov <jcmvbkbc@gmail.com> Signed-off-by: Chris Zankel <chris@zankel.net>
MMUv3 comes out of reset with identity vaddr -> paddr mapping in the TLB
way 6:
Way 6 (512 MB)
Vaddr Paddr ASID Attr RWX Cache
---------- ---------- ---- ---- --- -------
0x00000000 0x00000000 0x01 0x03 RWX Bypass
0x20000000 0x20000000 0x01 0x03 RWX Bypass
0x40000000 0x40000000 0x01 0x03 RWX Bypass
0x60000000 0x60000000 0x01 0x03 RWX Bypass
0x80000000 0x80000000 0x01 0x03 RWX Bypass
0xa0000000 0xa0000000 0x01 0x03 RWX Bypass
0xc0000000 0xc0000000 0x01 0x03 RWX Bypass
0xe0000000 0xe0000000 0x01 0x03 RWX Bypass
This patch adds remapping code at the reset vector or at the kernel
_start (depending on CONFIG_INITIALIZE_XTENSA_MMU_INSIDE_VMLINUX) that
reconfigures MMUv3 as MMUv2:
Way 5 (128 MB)
Vaddr Paddr ASID Attr RWX Cache
---------- ---------- ---- ---- --- -------
0xd0000000 0x00000000 0x01 0x07 RWX WB
0xd8000000 0x00000000 0x01 0x03 RWX Bypass
Way 6 (256 MB)
Vaddr Paddr ASID Attr RWX Cache
---------- ---------- ---- ---- --- -------
0xe0000000 0xf0000000 0x01 0x07 RWX WB
0xf0000000 0xf0000000 0x01 0x03 RWX Bypass
Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
Signed-off-by: Chris Zankel <chris@zankel.net>
Add support for dispatching medium-priority interrupts, that is, interrupts of priority levels 2 to EXCM_LEVEL. IRQ handling may be preempted by higher priority IRQ. Signed-off-by: Marc Gauthier <marc@tensilica.com> Signed-off-by: Max Filippov <jcmvbkbc@gmail.com> Signed-off-by: Chris Zankel <chris@zankel.net>
Remove heading and trailing spaces, trim trailing lines, and wrap lines that are longer than 80 characters. Signed-off-by: Chris Zankel <chris@zankel.net>
Use ENDPROC() to mark the end of assembler functions. Signed-off-by: Chris Zankel <chris@zankel.net>