zephyr/arch/arc/core/isr_wrapper.S

250 lines
8.2 KiB
ArmAsm

/*
* Copyright (c) 2014-2015 Wind River Systems, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/**
* @file
* @brief Wrapper around ISRs with logic for context switching
*
*
* Wrapper installed in vector table for handling dynamic interrupts that accept
* a parameter.
*/
#define _ASMLANGUAGE
#include <offsets.h>
#include <toolchain.h>
#include <sections.h>
#include <sw_isr_table.h>
#include <nano_private.h>
#include <arch/cpu.h>
GTEXT(_isr_enter)
GTEXT(_isr_demux)
#if defined(CONFIG_NANOKERNEL) && defined(CONFIG_TICKLESS_IDLE)
GTEXT(_power_save_idle_exit)
#endif
/*
The symbols in this file are not real functions, and neither are
_rirq_enter/_firq_enter: they are jump points.
The flow is the following:
ISR -> _isr_enter -- + -> _rirq_enter -> _isr_demux -> ISR -> _rirq_exit
|
+ -> _firq_enter -> _isr_demux -> ISR -> _firq_exit
Context switch explanation:
The context switch code is spread in these files:
isr_wrapper.s, swap.s, swap_macros.s, fast_irq.s, regular_irq.s
IRQ stack frame layout:
high address
status32
pc
[jli_base]
[ldi_base]
[ei_base]
lp_count
lp_start
lp_end
blink
r13
...
sp -> r0
low address
[registers not taken into account in the current implementation]
The context switch code adopts this standard so that it is easier to follow:
- r1 contains _nanokernel ASAP and is not overwritten over the lifespan of
the functions.
- r2 contains _nanokernel.current ASAP, and the incoming thread when we
transition from outgoing thread to incoming thread
Not loading _nanokernel into r0 allows loading _nanokernel without stomping on
the parameter in r0 in _Swap().
ARCv2 processors have two kinds of interrupts: fast (FIRQ) and regular. The
official documentation calls the regular interrupts 'IRQs', but the internals
of the kernel call them 'RIRQs' to differentiate from the 'irq' subsystem,
which is the interrupt API/layer of abstraction.
FIRQs can be used to allow ISRs to run without having to save any context,
since they work with a second register bank. However, they can be somewhat more
limited than RIRQs since the register bank does not copy every possible
register that is needed to implement all available instructions: an example is
that the 'loop' registers (lp_count, lp_end, lp_start) are not present in the
second bank. The kernel thus takes upon itself to save these extra registers,
if the FIRQ is made known to the kernel. It is possible for a FIRQ to operate
outside of the kernel, but care must be taken to only use instructions that
only use the banked registers. RIRQs must always use the kernel's interrupt
entry and exit mechanisms.
The kernel is able to handle transitions to and from FIRQ, RIRQ and threads
(fibers/task). The contexts are saved 'lazily': the minimum amount of work is
done upfront, and the rest is done when needed:
o RIRQ
All needed regisers to run C code in the ISR are saved automatically
on the outgoing thread's stack: loop, status32, pc, and the caller-
saved GPRs. That stack frame layout is pre-determined. If returning
to a fiber, the stack is popped and no registers have to be saved by
the kernel. If a context switch is required, the callee-saved GPRs
are then saved in the thread control structure (TCS).
o FIRQ
First, a FIRQ can be interrupting a lower-priority RIRQ: if this is the case,
the FIRQ does not take a scheduling decision and leaves it the RIRQ to
handle. This limits the amount of code that has to run at interrupt-level.
GPRs are banked, loop registers are saved in a global structure upon
interrupt entry. If returning to a fiber, loop registers are poppped and the
CPU switches back to bank 0 for the GPRs. If a context switch is
needed, at this point only are all the registers saved. First, a
stack frame with the same layout as the automatic RIRQ one is created
and then the callee-saved GPRs are saved in the TCS. status32_p0 and
ilink are saved in this case, not status32 and pc.
To create the stack frame, the FIRQ handling code must first go back to using
bank0 of registers, since that is where the registers containing the exiting
thread are saved. Care must be taken not to touch any register before saving
them: the only one usable at that point is the stack pointer.
o coop
When a coop context switch is done, the callee-saved registers are
saved in the TCS. The other GPRs do not need to be saved, since the
compiler has already placed them on the stack.
For restoring the contexts, there are six cases. In all cases, the
callee-saved registers of the incoming thread have to be restored. Then, there
are specifics for each case:
From coop:
o to coop
Restore interrupt lock level and do a normal function call return.
o to any irq
The incoming interrupted thread has an IRQ stack frame containing the
caller-saved registers that has to be popped. status32 has to be restored,
then we jump to the interrupted instruction.
From FIRQ:
The processor is back to using bank0, not bank1 anymore, because it had to
save the outgoing context from bank0, and now has to load the incoming one
into bank0.
o to coop
The address of the returning instruction from _Swap() is loaded in ilink and
the saved status32 in status32_p0, taking care to adjust the interrupt lock
state desired in status32_p0. The return value is put in r0.
o to any irq
The IRQ has saved the caller-saved registers in a stack frame, which must be
popped, and statu32 and pc loaded in status32_p0 and ilink.
From RIRQ:
o to coop
The interrupt return mechanism in the processor expects a stack frame, but
the outgoing context did not create one. A fake one is created here, with
only the relevant values filled in: pc, status32 and the return value in r0.
There is a discrepancy between the ABI from the ARCv2 docs, including the
way the processor pushes GPRs in pairs in the IRQ stack frame, and the ABI
GCC uses. r13 should be a callee-saved register, but GCC treats it as
caller-saved. This means that the processor pushes it in the stack frame
along with r12, but the compiler does not save it before entering a
function. So, it is saved as part of the callee-saved registers, and
restored there, but the processor restores it _a second time_ when popping
the IRQ stack frame. Thus, the correct value must also be put in the fake
stack frame when returning to a thread that context switched out
cooperatively.
o to any irq
Both types of IRQs already have an IRQ stack frame: simply return from
interrupt.
*/
SECTION_FUNC(TEXT, _isr_enter)
lr r0, [_ARC_V2_AUX_IRQ_ACT]
ffs r0, r0
cmp r0, 0
mov.z r3, _firq_exit
mov.z r2, _firq_enter
mov.nz r3, _rirq_exit
mov.nz r2, _rirq_enter
j_s.nd [r2]
#if defined(CONFIG_NANOKERNEL) && defined(CONFIG_TICKLESS_IDLE)
.macro exit_tickless_idle
clri r0 /* do not interrupt exiting tickless idle operations */
push_s r0
push_s blink
jl _power_save_idle_exit
pop_s blink
pop_s r0
seti r0
.endm
#else
#define exit_tickless_idle
#endif
/* when getting here, r3 contains the interrupt exit stub to call */
SECTION_FUNC(TEXT, _isr_demux)
push_s r3
/* cannot be done before this point because we must be able to run C */
/* r0 is available to be stomped here, and exit_tickless_idle uses it */
exit_tickless_idle
lr r0, [_ARC_V2_ICAUSE]
sub r0, r0, 16
mov r1, _sw_isr_table
add3 r0, r1, r0 /* table entries are 8-bytes wide */
ld r1, [r0, 4] /* ISR into r1 */
jl_s.d [r1]
ld_s r0, [r0] /* delay slot: ISR parameter into r0 */
/* back from ISR, jump to exit stub */
pop_s r3
j_s.nd [r3]
nop