On 09/24/2020 09:19 AM, Huacai Chen wrote:
Hi, Jiaxun,
On Wed, Sep 23, 2020 at 9:37 PM Jiaxun Yang <jiaxun.yang@xxxxxxxxxxx> wrote:
在 2020/9/23 10:30, Huacai Chen 写道:
Add kexec/kdump support for Loongson64 by:
1, Provide Loongson-specific kexec functions: loongson_kexec_prepare,
loongson_kexec_shutdown and loongson_crash_shutdown;
2, Provide Loongson-specific assembly code in kexec_smp_wait;
3, Clear mailbox in loongson3_smp_setup() since KEXEC bypass BIOS;
4, KEXEC always run at boot CPU, but KDUMP may triggered at non-boot
CPU. Loongson64 assume boot CPU is the first possible cpu, so fix
boot_cpu_id in prom_init_env();
To start Loongson64, The boot CPU needs 3 parameters:
fw_arg0: the number of arguments in cmdline (i.e., argc).
fw_arg1: structure holds cmdline such as "root=/dev/sda1 console=tty"
(i.e., argv).
fw_arg2: environment (i.e., envp, additional boot parameters from LEFI).
Non-boot CPUs do not need parameters at once. They query their own IPI
mailbox to get PC, SP and GP in a loop until boot CPU brings them up.
loongson_kexec_prepare(): Setup cmdline for kexec/kdump. The kexec/kdump
cmdline comes from kexec's "append" option string. This structure will
be parsed in fw_init_cmdline() of arch/mips/fw/lib/cmdline.c. Both image
->control_code_page and the cmdline need to be in a safe memory region
(memory allocated by the old kernel may be corrupted by the new kernel).
In order to maintain compatibility for the old firmware, the low 2MB is
reserverd and safe for Loongson. So let KEXEC_CTRL_CODE and KEXEC_ARGV_
ADDR be here. LEFI parameters may be corrupted at runtime, so backup it
at mips_reboot_setup(), and then restore it at loongson_kexec_shutdown()
/loongson_crash_shutdown().
loongson_kexec_shutdown(): Wake up all present CPUs and let them go to
reboot_code_buffer. Pass the kexec parameters to kexec_args.
loongson_crash_shutdown(): Pass the kdump parameters to kexec_args.
The assembly part in kexec_smp_wait provide a routine as BIOS does, in
order to keep secondary CPUs in a querying loop.
Cc: Eric Biederman <ebiederm@xxxxxxxxxxxx>
Signed-off-by: Huacai Chen <chenhc@xxxxxxxxxx>
Signed-off-by: Jinyang He <hejinyang@xxxxxxxxxxx>
Signed-off-by: Youling Tang <tangyouling@xxxxxxxxxxx>
---
arch/mips/kernel/relocate_kernel.S | 27 +++++++++
arch/mips/loongson64/env.c | 7 +++
arch/mips/loongson64/reset.c | 111 +++++++++++++++++++++++++++++++++++++
arch/mips/loongson64/smp.c | 5 ++
4 files changed, 150 insertions(+)
diff --git a/arch/mips/kernel/relocate_kernel.S b/arch/mips/kernel/relocate_kernel.S
index ac87089..91b2932 100644
--- a/arch/mips/kernel/relocate_kernel.S
+++ b/arch/mips/kernel/relocate_kernel.S
@@ -133,6 +133,33 @@ LEAF(kexec_smp_wait)
#else
sync
#endif
Hi Jingyang, Huacai,
Thanks a lot for what you have done on Loongson64 support.
Just add my cents here.
+
+#ifdef CONFIG_CPU_LOONGSON64
+#define MAILBOX_BASE 0x900000003ff01000
+ /* s0:prid s1:initfn */
+ /* t0:base t1:cpuid t2:node t9:count */
+ mfc0 t1, CP0_EBASE
+ andi t1, MIPS_EBASE_CPUNUM
+ dli t0, MAILBOX_BASE /* mailbox base */
+ dins t0, t1, 8, 2 /* insert core id*/
+ dext t2, t1, 2, 2
+ dins t0, t2, 44, 2 /* insert node id */
+ mfc0 s0, CP0_PRID
+ andi s0, s0, 0xf
+ blt s0, 0x6, 1f /* Loongson-3A1000 */
+ bgt s0, 0x7, 1f /* Loongson-3A2000/3A3000 */
+ dins t0, t2, 14, 2 /* Loongson-3B1000/3B1500 need bit 15~14 */
+1: li t9, 0x100 /* wait for init loop */
+2: addiu t9, -1 /* limit mailbox access */
+ bnez t9, 2b
+ ld s1, 0x20(t0) /* get PC via mailbox reg0 */
+ beqz s1, 1b
+ ld sp, 0x28(t0) /* get SP via mailbox reg1 */
+ ld gp, 0x30(t0) /* get GP via mailbox reg2 */
+ ld a1, 0x38(t0)
+ jr s1 /* jump to initial PC */
+#endif
+
I'm just a little bit uncomfortable with this kind of hardcoded address.
Is it possible to generate kexec_smp_wait with uasm, or pass the SMP
base as a parameter of this function?
This is very difficult, and moreover, uasm wrap the assembly in C
functions, can it be more beautiful? In my opinion, uasm is only used
for performance-critical routines, not for beautiful code. But anyway,
0x900000003ff01000 should be improved.
Also I do think we can handle kexec_flag in Loongson's platform SMP,
or even generic MIPS SMP code just like what cavium did so this kind
of platform quirk can be avoided.
I doubt this can be done, every CPU has its own method of SMP bringup.
j s1
END(kexec_smp_wait)
#endif
diff --git a/arch/mips/loongson64/env.c b/arch/mips/loongson64/env.c
index 134cb8e..e937f31 100644
--- a/arch/mips/loongson64/env.c
+++ b/arch/mips/loongson64/env.c
@@ -120,6 +120,13 @@ void __init prom_init_env(void)
loongson_sysconf.nr_cpus = ecpu->nr_cpus;
loongson_sysconf.boot_cpu_id = ecpu->cpu_startup_core_id;
loongson_sysconf.reserved_cpus_mask = ecpu->reserved_cores_mask;
+#ifdef CONFIG_KEXEC
+ loongson_sysconf.boot_cpu_id = get_ebase_cpunum();
+ loongson_sysconf.reserved_cpus_mask |=
+ (1 << loongson_sysconf.boot_cpu_id) - 1;
+ pr_info("Boot CPU ID is being fixed from %d to %d\n",
+ ecpu->cpu_startup_core_id, loongson_sysconf.boot_cpu_id);
+#endif
This could be done unconditionally and split to another patch,
"MIPS: Loongson64: Quirk inaccurate bootcore from the firmware"
This will be wholly removed.
if (ecpu->nr_cpus > NR_CPUS || ecpu->nr_cpus == 0)
loongson_sysconf.nr_cpus = NR_CPUS;
loongson_sysconf.nr_nodes = (loongson_sysconf.nr_cpus +
diff --git a/arch/mips/loongson64/reset.c b/arch/mips/loongson64/reset.c
index 3bb8a1e..b1e71f37 100644
--- a/arch/mips/loongson64/reset.c
+++ b/arch/mips/loongson64/reset.c
@@ -6,9 +6,14 @@
* Copyright (C) 2009 Lemote, Inc.
* Author: Zhangjin Wu, wuzhangjin@xxxxxxxxx
*/
+#include <linux/cpu.h>
+#include <linux/delay.h>
#include <linux/init.h>
+#include <linux/kexec.h>
#include <linux/pm.h>
+#include <linux/slab.h>
+#include <asm/bootinfo.h>
#include <asm/idle.h>
#include <asm/reboot.h>
@@ -47,12 +52,118 @@ static void loongson_halt(void)
}
}
+#ifdef CONFIG_KEXEC
+
+/* 0X80000000~0X80200000 is safe */
+#define MAX_ARGS 64
+#define KEXEC_CTRL_CODE 0xFFFFFFFF80100000UL
+#define KEXEC_ARGV_ADDR 0xFFFFFFFF80108000UL
+#define KEXEC_ARGV_SIZE COMMAND_LINE_SIZE
+#define KEXEC_ENVP_SIZE 4800
I won't say it's safe.
Loongson-2K's PMON may place reboot vector here, also some
UEFI firmware may place their suspend stack here.
What if we allocate these buffer at runtime?
The layout of low 2MB in our design:
0x80000000 the first MB, the first 64K：Exception vector
0x80010000 the first MB, the second 64K：STR(suspend) data
0x80020000 the first MB, the third and fourth 64K：UEFI HOB
0x80040000 the first MB, the fifth 64K：RT-Thread for SMC
0x80100000 the second MB, the first 64K：KEXEC code
0x80108000 the second MB, the second 64K：KEXEC data
Hi, Huacai,
Thank you for your explanation. It's really a safe region.
What I want to do is similar to other architecture which find a hole
in Crash kernel. It's also safe if "mem=" limit Capture kernel RAM. "mem="
is useful and sorry for the wrong comment yesterday. "mem=" excluded
cmdline segment and elfcorehdr segment. I have an idea that add a segment
for Loongson64 platform control_code_page. Find the hole by top-down that
different from common bottom-up way.
As you said, it's really difficult. More work should be do not only kernel
but also kexec-tools. At this stage, I am in favor of you that
control_code_page should point to the safe ragion in low 2MB.
All allocated buffer from the old kernel is not safe, because the new
kernel may be larger than the old kernel. Even if the low 2MB is not a
perfect place, it is the best place we can choose.
Huacai
Thanks.
- Jiaxun
