[v3,5/5] LoongArch: Add generic ex-handler unwind in prologue unwinder

Message ID 20230112004048.26191-1-hejinyang@loongson.cn
State New
Headers
Series LoongArch: Some fix and new features for unwinders |

Commit Message

Jinyang He Jan. 12, 2023, 12:40 a.m. UTC
  When exception is triggered, code flow go handle_\exception in some
cases. One of stackframe in this case as follows,

high -> +-------+
        | REGS  |  <- a pt_regs
        |       |
        |       |  <- ex trigger
        | REGS  |  <- ex pt_regs   <-+
        |       |                    |
        |       |                    |
low  -> +-------+           ->unwind-+

When unwinder unwind to handler_\exception it cannot go on prologue
analysis. It is asynchronous code flow, we should get the next frame
PC from regs->csr_era but not from regs->regs[1]. And we copy the
handler codes to eentry in the early time and copy the handler codes
to NUMA-relative memory named pcpu_handlers if NUMA is enabled. Thus,
unwinder cannot unwind normally. Therefore, try to give some hint in
handler_\exception and fixup it in unwind_next_frame.

Reported-by: Qing Zhang <zhangqing@loongson.cn>
Signed-off-by: Jinyang He <hejinyang@loongson.cn>
---
 arch/loongarch/include/asm/unwind.h     |   2 +-
 arch/loongarch/kernel/genex.S           |   3 +
 arch/loongarch/kernel/unwind_prologue.c | 100 +++++++++++++++++++++---
 arch/loongarch/mm/tlb.c                 |   2 +-
 4 files changed, 92 insertions(+), 15 deletions(-)
  

Comments

Huacai Chen Jan. 12, 2023, 12:59 a.m. UTC | #1
Hi, Jinyang,

Have you tested this patch on NUMA systems?  I don't say it cannot
work, I just concern about it, because NUMA systems use different
exception tables.

Huacai

On Thu, Jan 12, 2023 at 8:41 AM Jinyang He <hejinyang@loongson.cn> wrote:
>
> When exception is triggered, code flow go handle_\exception in some
> cases. One of stackframe in this case as follows,
>
> high -> +-------+
>         | REGS  |  <- a pt_regs
>         |       |
>         |       |  <- ex trigger
>         | REGS  |  <- ex pt_regs   <-+
>         |       |                    |
>         |       |                    |
> low  -> +-------+           ->unwind-+
>
> When unwinder unwind to handler_\exception it cannot go on prologue
> analysis. It is asynchronous code flow, we should get the next frame
> PC from regs->csr_era but not from regs->regs[1]. And we copy the
> handler codes to eentry in the early time and copy the handler codes
> to NUMA-relative memory named pcpu_handlers if NUMA is enabled. Thus,
> unwinder cannot unwind normally. Therefore, try to give some hint in
> handler_\exception and fixup it in unwind_next_frame.
>
> Reported-by: Qing Zhang <zhangqing@loongson.cn>
> Signed-off-by: Jinyang He <hejinyang@loongson.cn>
> ---
>  arch/loongarch/include/asm/unwind.h     |   2 +-
>  arch/loongarch/kernel/genex.S           |   3 +
>  arch/loongarch/kernel/unwind_prologue.c | 100 +++++++++++++++++++++---
>  arch/loongarch/mm/tlb.c                 |   2 +-
>  4 files changed, 92 insertions(+), 15 deletions(-)
>
> diff --git a/arch/loongarch/include/asm/unwind.h b/arch/loongarch/include/asm/unwind.h
> index cb428e1b19af..a38eb152aefb 100644
> --- a/arch/loongarch/include/asm/unwind.h
> +++ b/arch/loongarch/include/asm/unwind.h
> @@ -22,7 +22,7 @@ struct unwind_state {
>         char type; /* UNWINDER_XXX */
>         struct stack_info stack_info;
>         struct task_struct *task;
> -       bool first, error, is_ftrace;
> +       bool first, error, reset;
>         int graph_idx;
>         unsigned long sp, pc, ra;
>  };
> diff --git a/arch/loongarch/kernel/genex.S b/arch/loongarch/kernel/genex.S
> index 75e5be807a0d..7e5c293ed89f 100644
> --- a/arch/loongarch/kernel/genex.S
> +++ b/arch/loongarch/kernel/genex.S
> @@ -67,14 +67,17 @@ SYM_FUNC_END(except_vec_cex)
>         .macro  BUILD_HANDLER exception handler prep
>         .align  5
>         SYM_FUNC_START(handle_\exception)
> +       666:
>         BACKUP_T0T1
>         SAVE_ALL
>         build_prep_\prep
>         move    a0, sp
>         la.abs  t0, do_\handler
>         jirl    ra, t0, 0
> +       668:
>         RESTORE_ALL_AND_RET
>         SYM_FUNC_END(handle_\exception)
> +       SYM_DATA(unwind_hint_\exception, .word 668b - 666b)
>         .endm
>
>         BUILD_HANDLER ade ade badv
> diff --git a/arch/loongarch/kernel/unwind_prologue.c b/arch/loongarch/kernel/unwind_prologue.c
> index e6c3f2ee507c..e52c18c2b604 100644
> --- a/arch/loongarch/kernel/unwind_prologue.c
> +++ b/arch/loongarch/kernel/unwind_prologue.c
> @@ -2,21 +2,100 @@
>  /*
>   * Copyright (C) 2022 Loongson Technology Corporation Limited
>   */
> +#include <linux/cpumask.h>
>  #include <linux/ftrace.h>
>  #include <linux/kallsyms.h>
>
>  #include <asm/inst.h>
> +#include <asm/loongson.h>
>  #include <asm/ptrace.h>
> +#include <asm/setup.h>
>  #include <asm/unwind.h>
>
> -static inline void unwind_state_fixup(struct unwind_state *state)
> +extern const int unwind_hint_ade;
> +extern const int unwind_hint_ale;
> +extern const int unwind_hint_bp;
> +extern const int unwind_hint_fpe;
> +extern const int unwind_hint_fpu;
> +extern const int unwind_hint_lsx;
> +extern const int unwind_hint_lasx;
> +extern const int unwind_hint_lbt;
> +extern const int unwind_hint_ri;
> +extern const int unwind_hint_watch;
> +extern unsigned long eentry;
> +#ifdef CONFIG_NUMA
> +extern unsigned long pcpu_handlers[NR_CPUS];
> +#endif
> +
> +static inline bool scan_handler(unsigned long entry_offset)
>  {
> -#ifdef CONFIG_DYNAMIC_FTRACE
> -       static unsigned long ftrace = (unsigned long)ftrace_call + 4;
> +       int idx, offset;
>
> -       if (state->pc == ftrace)
> -               state->is_ftrace = true;
> +       if (entry_offset >= EXCCODE_INT_START * VECSIZE)
> +               return false;
> +
> +       idx = entry_offset / VECSIZE;
> +       offset = entry_offset % VECSIZE;
> +       switch (idx) {
> +       case EXCCODE_ADE:
> +               return offset == unwind_hint_ade;
> +       case EXCCODE_ALE:
> +               return offset == unwind_hint_ale;
> +       case EXCCODE_BP:
> +               return offset == unwind_hint_bp;
> +       case EXCCODE_FPE:
> +               return offset == unwind_hint_fpe;
> +       case EXCCODE_FPDIS:
> +               return offset == unwind_hint_fpu;
> +       case EXCCODE_LSXDIS:
> +               return offset == unwind_hint_lsx;
> +       case EXCCODE_LASXDIS:
> +               return offset == unwind_hint_lasx;
> +       case EXCCODE_BTDIS:
> +               return offset == unwind_hint_lbt;
> +       case EXCCODE_INE:
> +               return offset == unwind_hint_ri;
> +       case EXCCODE_WATCH:
> +               return offset == unwind_hint_watch;
> +       default:
> +               return false;
> +       }
> +}
> +
> +static inline bool fix_exceptions(unsigned long pc)
> +{
> +#ifdef CONFIG_NUMA
> +       int cpu;
> +
> +       for_each_possible_cpu(cpu) {
> +               if (!pcpu_handlers[cpu])
> +                       continue;
> +               if (scan_handler(pc - pcpu_handlers[cpu]))
> +                       return true;
> +       }
>  #endif
> +       return scan_handler(pc - eentry);
> +}
> +
> +/*
> + * As we meet ftrace_regs_entry, reset first flag like first doing
> + * tracing. Prologue analysis will stop soon because PC is at entry.
> + */
> +static inline bool fix_ftrace(unsigned long pc)
> +{
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +       return pc == (unsigned long)ftrace_call + LOONGARCH_INSN_SIZE;
> +#else
> +       return false;
> +#endif
> +}
> +
> +static inline bool unwind_state_fixup(struct unwind_state *state)
> +{
> +       if (!fix_exceptions(state->pc) && !fix_ftrace(state->pc))
> +               return false;
> +       state->reset = true;
> +       return true;
>  }
>
>  /*
> @@ -39,14 +118,10 @@ static bool unwind_by_prologue(struct unwind_state *state)
>         if (state->sp >= info->end || state->sp < info->begin)
>                 return false;
>
> -       if (state->is_ftrace) {
> -               /*
> -                * As we meet ftrace_regs_entry, reset first flag like first doing
> -                * tracing. Prologue analysis will stop soon because PC is at entry.
> -                */
> +       if (state->reset) {
>                 regs = (struct pt_regs *)state->sp;
>                 state->first = true;
> -               state->is_ftrace = false;
> +               state->reset = false;
>                 state->pc = regs->csr_era;
>                 state->ra = regs->regs[1];
>                 state->sp = regs->regs[3];
> @@ -111,8 +186,7 @@ static bool unwind_by_prologue(struct unwind_state *state)
>
>  out:
>         state->first = false;
> -       unwind_state_fixup(state);
> -       return !!__kernel_text_address(state->pc);
> +       return unwind_state_fixup(state) || __kernel_text_address(state->pc);
>  }
>
>  static bool next_frame(struct unwind_state *state)
> diff --git a/arch/loongarch/mm/tlb.c b/arch/loongarch/mm/tlb.c
> index da3681f131c8..8bad6b0cff59 100644
> --- a/arch/loongarch/mm/tlb.c
> +++ b/arch/loongarch/mm/tlb.c
> @@ -251,7 +251,7 @@ static void output_pgtable_bits_defines(void)
>  }
>
>  #ifdef CONFIG_NUMA
> -static unsigned long pcpu_handlers[NR_CPUS];
> +unsigned long pcpu_handlers[NR_CPUS];
>  #endif
>  extern long exception_handlers[VECSIZE * 128 / sizeof(long)];
>
> --
> 2.34.3
>
>
  
Jinyang He Jan. 12, 2023, 5:55 a.m. UTC | #2
On 2023-01-12 08:59, Huacai Chen wrote:

> Hi, Jinyang,
>
> Have you tested this patch on NUMA systems?  I don't say it cannot
> work, I just concern about it, because NUMA systems use different
> exception tables.

Sorry, I do not have machine except 3A5000. When CONFIG_NUMA is enabled,
the non-boot CPU sets its CSR.EENTRY values pcpu_handlers[cpu]. So I think
if CONFIG_NUMA is enabled, the other machines should perform as 3A5000
doing. (The boot CPU use eentry as the exception entry, and the non-boot
CPU use pcpu_handlers[cpu] as the exception entry). If that makes sense,
I think the patch will also work well on other NUMA machines.


Thanks,

Jinyang

>
> Huacai
>
> On Thu, Jan 12, 2023 at 8:41 AM Jinyang He <hejinyang@loongson.cn> wrote:
>> When exception is triggered, code flow go handle_\exception in some
>> cases. One of stackframe in this case as follows,
>>
>> high -> +-------+
>>          | REGS  |  <- a pt_regs
>>          |       |
>>          |       |  <- ex trigger
>>          | REGS  |  <- ex pt_regs   <-+
>>          |       |                    |
>>          |       |                    |
>> low  -> +-------+           ->unwind-+
>>
>> When unwinder unwind to handler_\exception it cannot go on prologue
>> analysis. It is asynchronous code flow, we should get the next frame
>> PC from regs->csr_era but not from regs->regs[1]. And we copy the
>> handler codes to eentry in the early time and copy the handler codes
>> to NUMA-relative memory named pcpu_handlers if NUMA is enabled. Thus,
>> unwinder cannot unwind normally. Therefore, try to give some hint in
>> handler_\exception and fixup it in unwind_next_frame.
>>
>> Reported-by: Qing Zhang <zhangqing@loongson.cn>
>> Signed-off-by: Jinyang He <hejinyang@loongson.cn>
>> ---
>>   arch/loongarch/include/asm/unwind.h     |   2 +-
>>   arch/loongarch/kernel/genex.S           |   3 +
>>   arch/loongarch/kernel/unwind_prologue.c | 100 +++++++++++++++++++++---
>>   arch/loongarch/mm/tlb.c                 |   2 +-
>>   4 files changed, 92 insertions(+), 15 deletions(-)
>>
>> diff --git a/arch/loongarch/include/asm/unwind.h b/arch/loongarch/include/asm/unwind.h
>> index cb428e1b19af..a38eb152aefb 100644
>> --- a/arch/loongarch/include/asm/unwind.h
>> +++ b/arch/loongarch/include/asm/unwind.h
>> @@ -22,7 +22,7 @@ struct unwind_state {
>>          char type; /* UNWINDER_XXX */
>>          struct stack_info stack_info;
>>          struct task_struct *task;
>> -       bool first, error, is_ftrace;
>> +       bool first, error, reset;
>>          int graph_idx;
>>          unsigned long sp, pc, ra;
>>   };
>> diff --git a/arch/loongarch/kernel/genex.S b/arch/loongarch/kernel/genex.S
>> index 75e5be807a0d..7e5c293ed89f 100644
>> --- a/arch/loongarch/kernel/genex.S
>> +++ b/arch/loongarch/kernel/genex.S
>> @@ -67,14 +67,17 @@ SYM_FUNC_END(except_vec_cex)
>>          .macro  BUILD_HANDLER exception handler prep
>>          .align  5
>>          SYM_FUNC_START(handle_\exception)
>> +       666:
>>          BACKUP_T0T1
>>          SAVE_ALL
>>          build_prep_\prep
>>          move    a0, sp
>>          la.abs  t0, do_\handler
>>          jirl    ra, t0, 0
>> +       668:
>>          RESTORE_ALL_AND_RET
>>          SYM_FUNC_END(handle_\exception)
>> +       SYM_DATA(unwind_hint_\exception, .word 668b - 666b)
>>          .endm
>>
>>          BUILD_HANDLER ade ade badv
>> diff --git a/arch/loongarch/kernel/unwind_prologue.c b/arch/loongarch/kernel/unwind_prologue.c
>> index e6c3f2ee507c..e52c18c2b604 100644
>> --- a/arch/loongarch/kernel/unwind_prologue.c
>> +++ b/arch/loongarch/kernel/unwind_prologue.c
>> @@ -2,21 +2,100 @@
>>   /*
>>    * Copyright (C) 2022 Loongson Technology Corporation Limited
>>    */
>> +#include <linux/cpumask.h>
>>   #include <linux/ftrace.h>
>>   #include <linux/kallsyms.h>
>>
>>   #include <asm/inst.h>
>> +#include <asm/loongson.h>
>>   #include <asm/ptrace.h>
>> +#include <asm/setup.h>
>>   #include <asm/unwind.h>
>>
>> -static inline void unwind_state_fixup(struct unwind_state *state)
>> +extern const int unwind_hint_ade;
>> +extern const int unwind_hint_ale;
>> +extern const int unwind_hint_bp;
>> +extern const int unwind_hint_fpe;
>> +extern const int unwind_hint_fpu;
>> +extern const int unwind_hint_lsx;
>> +extern const int unwind_hint_lasx;
>> +extern const int unwind_hint_lbt;
>> +extern const int unwind_hint_ri;
>> +extern const int unwind_hint_watch;
>> +extern unsigned long eentry;
>> +#ifdef CONFIG_NUMA
>> +extern unsigned long pcpu_handlers[NR_CPUS];
>> +#endif
>> +
>> +static inline bool scan_handler(unsigned long entry_offset)
>>   {
>> -#ifdef CONFIG_DYNAMIC_FTRACE
>> -       static unsigned long ftrace = (unsigned long)ftrace_call + 4;
>> +       int idx, offset;
>>
>> -       if (state->pc == ftrace)
>> -               state->is_ftrace = true;
>> +       if (entry_offset >= EXCCODE_INT_START * VECSIZE)
>> +               return false;
>> +
>> +       idx = entry_offset / VECSIZE;
>> +       offset = entry_offset % VECSIZE;
>> +       switch (idx) {
>> +       case EXCCODE_ADE:
>> +               return offset == unwind_hint_ade;
>> +       case EXCCODE_ALE:
>> +               return offset == unwind_hint_ale;
>> +       case EXCCODE_BP:
>> +               return offset == unwind_hint_bp;
>> +       case EXCCODE_FPE:
>> +               return offset == unwind_hint_fpe;
>> +       case EXCCODE_FPDIS:
>> +               return offset == unwind_hint_fpu;
>> +       case EXCCODE_LSXDIS:
>> +               return offset == unwind_hint_lsx;
>> +       case EXCCODE_LASXDIS:
>> +               return offset == unwind_hint_lasx;
>> +       case EXCCODE_BTDIS:
>> +               return offset == unwind_hint_lbt;
>> +       case EXCCODE_INE:
>> +               return offset == unwind_hint_ri;
>> +       case EXCCODE_WATCH:
>> +               return offset == unwind_hint_watch;
>> +       default:
>> +               return false;
>> +       }
>> +}
>> +
>> +static inline bool fix_exceptions(unsigned long pc)
>> +{
>> +#ifdef CONFIG_NUMA
>> +       int cpu;
>> +
>> +       for_each_possible_cpu(cpu) {
>> +               if (!pcpu_handlers[cpu])
>> +                       continue;
>> +               if (scan_handler(pc - pcpu_handlers[cpu]))
>> +                       return true;
>> +       }
>>   #endif
>> +       return scan_handler(pc - eentry);
>> +}
>> +
>> +/*
>> + * As we meet ftrace_regs_entry, reset first flag like first doing
>> + * tracing. Prologue analysis will stop soon because PC is at entry.
>> + */
>> +static inline bool fix_ftrace(unsigned long pc)
>> +{
>> +#ifdef CONFIG_DYNAMIC_FTRACE
>> +       return pc == (unsigned long)ftrace_call + LOONGARCH_INSN_SIZE;
>> +#else
>> +       return false;
>> +#endif
>> +}
>> +
>> +static inline bool unwind_state_fixup(struct unwind_state *state)
>> +{
>> +       if (!fix_exceptions(state->pc) && !fix_ftrace(state->pc))
>> +               return false;
>> +       state->reset = true;
>> +       return true;
>>   }
>>
>>   /*
>> @@ -39,14 +118,10 @@ static bool unwind_by_prologue(struct unwind_state *state)
>>          if (state->sp >= info->end || state->sp < info->begin)
>>                  return false;
>>
>> -       if (state->is_ftrace) {
>> -               /*
>> -                * As we meet ftrace_regs_entry, reset first flag like first doing
>> -                * tracing. Prologue analysis will stop soon because PC is at entry.
>> -                */
>> +       if (state->reset) {
>>                  regs = (struct pt_regs *)state->sp;
>>                  state->first = true;
>> -               state->is_ftrace = false;
>> +               state->reset = false;
>>                  state->pc = regs->csr_era;
>>                  state->ra = regs->regs[1];
>>                  state->sp = regs->regs[3];
>> @@ -111,8 +186,7 @@ static bool unwind_by_prologue(struct unwind_state *state)
>>
>>   out:
>>          state->first = false;
>> -       unwind_state_fixup(state);
>> -       return !!__kernel_text_address(state->pc);
>> +       return unwind_state_fixup(state) || __kernel_text_address(state->pc);
>>   }
>>
>>   static bool next_frame(struct unwind_state *state)
>> diff --git a/arch/loongarch/mm/tlb.c b/arch/loongarch/mm/tlb.c
>> index da3681f131c8..8bad6b0cff59 100644
>> --- a/arch/loongarch/mm/tlb.c
>> +++ b/arch/loongarch/mm/tlb.c
>> @@ -251,7 +251,7 @@ static void output_pgtable_bits_defines(void)
>>   }
>>
>>   #ifdef CONFIG_NUMA
>> -static unsigned long pcpu_handlers[NR_CPUS];
>> +unsigned long pcpu_handlers[NR_CPUS];
>>   #endif
>>   extern long exception_handlers[VECSIZE * 128 / sizeof(long)];
>>
>> --
>> 2.34.3
>>
>>
  
Huacai Chen Jan. 17, 2023, 6:25 a.m. UTC | #3
On Thu, Jan 12, 2023 at 8:41 AM Jinyang He <hejinyang@loongson.cn> wrote:
>
> When exception is triggered, code flow go handle_\exception in some
> cases. One of stackframe in this case as follows,
>
> high -> +-------+
>         | REGS  |  <- a pt_regs
>         |       |
>         |       |  <- ex trigger
>         | REGS  |  <- ex pt_regs   <-+
>         |       |                    |
>         |       |                    |
> low  -> +-------+           ->unwind-+
>
> When unwinder unwind to handler_\exception it cannot go on prologue
> analysis. It is asynchronous code flow, we should get the next frame
> PC from regs->csr_era but not from regs->regs[1]. And we copy the
> handler codes to eentry in the early time and copy the handler codes
> to NUMA-relative memory named pcpu_handlers if NUMA is enabled. Thus,
> unwinder cannot unwind normally. Therefore, try to give some hint in
> handler_\exception and fixup it in unwind_next_frame.
>
> Reported-by: Qing Zhang <zhangqing@loongson.cn>
> Signed-off-by: Jinyang He <hejinyang@loongson.cn>
> ---
>  arch/loongarch/include/asm/unwind.h     |   2 +-
>  arch/loongarch/kernel/genex.S           |   3 +
>  arch/loongarch/kernel/unwind_prologue.c | 100 +++++++++++++++++++++---
>  arch/loongarch/mm/tlb.c                 |   2 +-
>  4 files changed, 92 insertions(+), 15 deletions(-)
>
> diff --git a/arch/loongarch/include/asm/unwind.h b/arch/loongarch/include/asm/unwind.h
> index cb428e1b19af..a38eb152aefb 100644
> --- a/arch/loongarch/include/asm/unwind.h
> +++ b/arch/loongarch/include/asm/unwind.h
> @@ -22,7 +22,7 @@ struct unwind_state {
>         char type; /* UNWINDER_XXX */
>         struct stack_info stack_info;
>         struct task_struct *task;
> -       bool first, error, is_ftrace;
> +       bool first, error, reset;
Hi, Qing,
Do you think is_reset is better than reset here?

>         int graph_idx;
>         unsigned long sp, pc, ra;
>  };
> diff --git a/arch/loongarch/kernel/genex.S b/arch/loongarch/kernel/genex.S
> index 75e5be807a0d..7e5c293ed89f 100644
> --- a/arch/loongarch/kernel/genex.S
> +++ b/arch/loongarch/kernel/genex.S
> @@ -67,14 +67,17 @@ SYM_FUNC_END(except_vec_cex)
>         .macro  BUILD_HANDLER exception handler prep
>         .align  5
>         SYM_FUNC_START(handle_\exception)
> +       666:
>         BACKUP_T0T1
>         SAVE_ALL
>         build_prep_\prep
>         move    a0, sp
>         la.abs  t0, do_\handler
>         jirl    ra, t0, 0
> +       668:
>         RESTORE_ALL_AND_RET
I think the 668 label should be after RESTORE_ALL_AND_RET, right?

Huacai

>         SYM_FUNC_END(handle_\exception)
> +       SYM_DATA(unwind_hint_\exception, .word 668b - 666b)
>         .endm
>
>         BUILD_HANDLER ade ade badv
> diff --git a/arch/loongarch/kernel/unwind_prologue.c b/arch/loongarch/kernel/unwind_prologue.c
> index e6c3f2ee507c..e52c18c2b604 100644
> --- a/arch/loongarch/kernel/unwind_prologue.c
> +++ b/arch/loongarch/kernel/unwind_prologue.c
> @@ -2,21 +2,100 @@
>  /*
>   * Copyright (C) 2022 Loongson Technology Corporation Limited
>   */
> +#include <linux/cpumask.h>
>  #include <linux/ftrace.h>
>  #include <linux/kallsyms.h>
>
>  #include <asm/inst.h>
> +#include <asm/loongson.h>
>  #include <asm/ptrace.h>
> +#include <asm/setup.h>
>  #include <asm/unwind.h>
>
> -static inline void unwind_state_fixup(struct unwind_state *state)
> +extern const int unwind_hint_ade;
> +extern const int unwind_hint_ale;
> +extern const int unwind_hint_bp;
> +extern const int unwind_hint_fpe;
> +extern const int unwind_hint_fpu;
> +extern const int unwind_hint_lsx;
> +extern const int unwind_hint_lasx;
> +extern const int unwind_hint_lbt;
> +extern const int unwind_hint_ri;
> +extern const int unwind_hint_watch;
> +extern unsigned long eentry;
> +#ifdef CONFIG_NUMA
> +extern unsigned long pcpu_handlers[NR_CPUS];
> +#endif
> +
> +static inline bool scan_handler(unsigned long entry_offset)
>  {
> -#ifdef CONFIG_DYNAMIC_FTRACE
> -       static unsigned long ftrace = (unsigned long)ftrace_call + 4;
> +       int idx, offset;
>
> -       if (state->pc == ftrace)
> -               state->is_ftrace = true;
> +       if (entry_offset >= EXCCODE_INT_START * VECSIZE)
> +               return false;
> +
> +       idx = entry_offset / VECSIZE;
> +       offset = entry_offset % VECSIZE;
> +       switch (idx) {
> +       case EXCCODE_ADE:
> +               return offset == unwind_hint_ade;
> +       case EXCCODE_ALE:
> +               return offset == unwind_hint_ale;
> +       case EXCCODE_BP:
> +               return offset == unwind_hint_bp;
> +       case EXCCODE_FPE:
> +               return offset == unwind_hint_fpe;
> +       case EXCCODE_FPDIS:
> +               return offset == unwind_hint_fpu;
> +       case EXCCODE_LSXDIS:
> +               return offset == unwind_hint_lsx;
> +       case EXCCODE_LASXDIS:
> +               return offset == unwind_hint_lasx;
> +       case EXCCODE_BTDIS:
> +               return offset == unwind_hint_lbt;
> +       case EXCCODE_INE:
> +               return offset == unwind_hint_ri;
> +       case EXCCODE_WATCH:
> +               return offset == unwind_hint_watch;
> +       default:
> +               return false;
> +       }
> +}
> +
> +static inline bool fix_exceptions(unsigned long pc)
> +{
> +#ifdef CONFIG_NUMA
> +       int cpu;
> +
> +       for_each_possible_cpu(cpu) {
> +               if (!pcpu_handlers[cpu])
> +                       continue;
> +               if (scan_handler(pc - pcpu_handlers[cpu]))
> +                       return true;
> +       }
>  #endif
> +       return scan_handler(pc - eentry);
> +}
> +
> +/*
> + * As we meet ftrace_regs_entry, reset first flag like first doing
> + * tracing. Prologue analysis will stop soon because PC is at entry.
> + */
> +static inline bool fix_ftrace(unsigned long pc)
> +{
> +#ifdef CONFIG_DYNAMIC_FTRACE
> +       return pc == (unsigned long)ftrace_call + LOONGARCH_INSN_SIZE;
> +#else
> +       return false;
> +#endif
> +}
> +
> +static inline bool unwind_state_fixup(struct unwind_state *state)
> +{
> +       if (!fix_exceptions(state->pc) && !fix_ftrace(state->pc))
> +               return false;
> +       state->reset = true;
> +       return true;
>  }
>
>  /*
> @@ -39,14 +118,10 @@ static bool unwind_by_prologue(struct unwind_state *state)
>         if (state->sp >= info->end || state->sp < info->begin)
>                 return false;
>
> -       if (state->is_ftrace) {
> -               /*
> -                * As we meet ftrace_regs_entry, reset first flag like first doing
> -                * tracing. Prologue analysis will stop soon because PC is at entry.
> -                */
> +       if (state->reset) {
>                 regs = (struct pt_regs *)state->sp;
>                 state->first = true;
> -               state->is_ftrace = false;
> +               state->reset = false;
>                 state->pc = regs->csr_era;
>                 state->ra = regs->regs[1];
>                 state->sp = regs->regs[3];
> @@ -111,8 +186,7 @@ static bool unwind_by_prologue(struct unwind_state *state)
>
>  out:
>         state->first = false;
> -       unwind_state_fixup(state);
> -       return !!__kernel_text_address(state->pc);
> +       return unwind_state_fixup(state) || __kernel_text_address(state->pc);
>  }
>
>  static bool next_frame(struct unwind_state *state)
> diff --git a/arch/loongarch/mm/tlb.c b/arch/loongarch/mm/tlb.c
> index da3681f131c8..8bad6b0cff59 100644
> --- a/arch/loongarch/mm/tlb.c
> +++ b/arch/loongarch/mm/tlb.c
> @@ -251,7 +251,7 @@ static void output_pgtable_bits_defines(void)
>  }
>
>  #ifdef CONFIG_NUMA
> -static unsigned long pcpu_handlers[NR_CPUS];
> +unsigned long pcpu_handlers[NR_CPUS];
>  #endif
>  extern long exception_handlers[VECSIZE * 128 / sizeof(long)];
>
> --
> 2.34.3
>
>
  
Qing Zhang Jan. 17, 2023, 7:41 a.m. UTC | #4
Hi, Huacai

On 2023/1/17 下午2:25, Huacai Chen wrote:
> On Thu, Jan 12, 2023 at 8:41 AM Jinyang He <hejinyang@loongson.cn> wrote:
>>
>> When exception is triggered, code flow go handle_\exception in some
>> cases. One of stackframe in this case as follows,
>>
>> high -> +-------+
>>          | REGS  |  <- a pt_regs
>>          |       |
>>          |       |  <- ex trigger
>>          | REGS  |  <- ex pt_regs   <-+
>>          |       |                    |
>>          |       |                    |
>> low  -> +-------+           ->unwind-+
>>
>> When unwinder unwind to handler_\exception it cannot go on prologue
>> analysis. It is asynchronous code flow, we should get the next frame
>> PC from regs->csr_era but not from regs->regs[1]. And we copy the
>> handler codes to eentry in the early time and copy the handler codes
>> to NUMA-relative memory named pcpu_handlers if NUMA is enabled. Thus,
>> unwinder cannot unwind normally. Therefore, try to give some hint in
>> handler_\exception and fixup it in unwind_next_frame.
>>
>> Reported-by: Qing Zhang <zhangqing@loongson.cn>
>> Signed-off-by: Jinyang He <hejinyang@loongson.cn>
>> ---
>>   arch/loongarch/include/asm/unwind.h     |   2 +-
>>   arch/loongarch/kernel/genex.S           |   3 +
>>   arch/loongarch/kernel/unwind_prologue.c | 100 +++++++++++++++++++++---
>>   arch/loongarch/mm/tlb.c                 |   2 +-
>>   4 files changed, 92 insertions(+), 15 deletions(-)
>>
>> diff --git a/arch/loongarch/include/asm/unwind.h b/arch/loongarch/include/asm/unwind.h
>> index cb428e1b19af..a38eb152aefb 100644
>> --- a/arch/loongarch/include/asm/unwind.h
>> +++ b/arch/loongarch/include/asm/unwind.h
>> @@ -22,7 +22,7 @@ struct unwind_state {
>>          char type; /* UNWINDER_XXX */
>>          struct stack_info stack_info;
>>          struct task_struct *task;
>> -       bool first, error, is_ftrace;
>> +       bool first, error, reset;
> Hi, Qing,
> Do you think is_reset is better than reset here?
ok, It depends on you.

Thanks
-Qing

>>          int graph_idx;
>>          unsigned long sp, pc, ra;
>>   };
>> diff --git a/arch/loongarch/kernel/genex.S b/arch/loongarch/kernel/genex.S
>> index 75e5be807a0d..7e5c293ed89f 100644
>> --- a/arch/loongarch/kernel/genex.S
>> +++ b/arch/loongarch/kernel/genex.S
>> @@ -67,14 +67,17 @@ SYM_FUNC_END(except_vec_cex)
>>          .macro  BUILD_HANDLER exception handler prep
>>          .align  5
>>          SYM_FUNC_START(handle_\exception)
>> +       666:
>>          BACKUP_T0T1
>>          SAVE_ALL
>>          build_prep_\prep
>>          move    a0, sp
>>          la.abs  t0, do_\handler
>>          jirl    ra, t0, 0
>> +       668:
>>          RESTORE_ALL_AND_RET
> I think the 668 label should be after RESTORE_ALL_AND_RET, right?
> 
> Huacai
> 
>>          SYM_FUNC_END(handle_\exception)
>> +       SYM_DATA(unwind_hint_\exception, .word 668b - 666b)
>>          .endm
>>
>>          BUILD_HANDLER ade ade badv
>> diff --git a/arch/loongarch/kernel/unwind_prologue.c b/arch/loongarch/kernel/unwind_prologue.c
>> index e6c3f2ee507c..e52c18c2b604 100644
>> --- a/arch/loongarch/kernel/unwind_prologue.c
>> +++ b/arch/loongarch/kernel/unwind_prologue.c
>> @@ -2,21 +2,100 @@
>>   /*
>>    * Copyright (C) 2022 Loongson Technology Corporation Limited
>>    */
>> +#include <linux/cpumask.h>
>>   #include <linux/ftrace.h>
>>   #include <linux/kallsyms.h>
>>
>>   #include <asm/inst.h>
>> +#include <asm/loongson.h>
>>   #include <asm/ptrace.h>
>> +#include <asm/setup.h>
>>   #include <asm/unwind.h>
>>
>> -static inline void unwind_state_fixup(struct unwind_state *state)
>> +extern const int unwind_hint_ade;
>> +extern const int unwind_hint_ale;
>> +extern const int unwind_hint_bp;
>> +extern const int unwind_hint_fpe;
>> +extern const int unwind_hint_fpu;
>> +extern const int unwind_hint_lsx;
>> +extern const int unwind_hint_lasx;
>> +extern const int unwind_hint_lbt;
>> +extern const int unwind_hint_ri;
>> +extern const int unwind_hint_watch;
>> +extern unsigned long eentry;
>> +#ifdef CONFIG_NUMA
>> +extern unsigned long pcpu_handlers[NR_CPUS];
>> +#endif
>> +
>> +static inline bool scan_handler(unsigned long entry_offset)
>>   {
>> -#ifdef CONFIG_DYNAMIC_FTRACE
>> -       static unsigned long ftrace = (unsigned long)ftrace_call + 4;
>> +       int idx, offset;
>>
>> -       if (state->pc == ftrace)
>> -               state->is_ftrace = true;
>> +       if (entry_offset >= EXCCODE_INT_START * VECSIZE)
>> +               return false;
>> +
>> +       idx = entry_offset / VECSIZE;
>> +       offset = entry_offset % VECSIZE;
>> +       switch (idx) {
>> +       case EXCCODE_ADE:
>> +               return offset == unwind_hint_ade;
>> +       case EXCCODE_ALE:
>> +               return offset == unwind_hint_ale;
>> +       case EXCCODE_BP:
>> +               return offset == unwind_hint_bp;
>> +       case EXCCODE_FPE:
>> +               return offset == unwind_hint_fpe;
>> +       case EXCCODE_FPDIS:
>> +               return offset == unwind_hint_fpu;
>> +       case EXCCODE_LSXDIS:
>> +               return offset == unwind_hint_lsx;
>> +       case EXCCODE_LASXDIS:
>> +               return offset == unwind_hint_lasx;
>> +       case EXCCODE_BTDIS:
>> +               return offset == unwind_hint_lbt;
>> +       case EXCCODE_INE:
>> +               return offset == unwind_hint_ri;
>> +       case EXCCODE_WATCH:
>> +               return offset == unwind_hint_watch;
>> +       default:
>> +               return false;
>> +       }
>> +}
>> +
>> +static inline bool fix_exceptions(unsigned long pc)
>> +{
>> +#ifdef CONFIG_NUMA
>> +       int cpu;
>> +
>> +       for_each_possible_cpu(cpu) {
>> +               if (!pcpu_handlers[cpu])
>> +                       continue;
>> +               if (scan_handler(pc - pcpu_handlers[cpu]))
>> +                       return true;
>> +       }
>>   #endif
>> +       return scan_handler(pc - eentry);
>> +}
>> +
>> +/*
>> + * As we meet ftrace_regs_entry, reset first flag like first doing
>> + * tracing. Prologue analysis will stop soon because PC is at entry.
>> + */
>> +static inline bool fix_ftrace(unsigned long pc)
>> +{
>> +#ifdef CONFIG_DYNAMIC_FTRACE
>> +       return pc == (unsigned long)ftrace_call + LOONGARCH_INSN_SIZE;
>> +#else
>> +       return false;
>> +#endif
>> +}
>> +
>> +static inline bool unwind_state_fixup(struct unwind_state *state)
>> +{
>> +       if (!fix_exceptions(state->pc) && !fix_ftrace(state->pc))
>> +               return false;
>> +       state->reset = true;
>> +       return true;
>>   }
>>
>>   /*
>> @@ -39,14 +118,10 @@ static bool unwind_by_prologue(struct unwind_state *state)
>>          if (state->sp >= info->end || state->sp < info->begin)
>>                  return false;
>>
>> -       if (state->is_ftrace) {
>> -               /*
>> -                * As we meet ftrace_regs_entry, reset first flag like first doing
>> -                * tracing. Prologue analysis will stop soon because PC is at entry.
>> -                */
>> +       if (state->reset) {
>>                  regs = (struct pt_regs *)state->sp;
>>                  state->first = true;
>> -               state->is_ftrace = false;
>> +               state->reset = false;
>>                  state->pc = regs->csr_era;
>>                  state->ra = regs->regs[1];
>>                  state->sp = regs->regs[3];
>> @@ -111,8 +186,7 @@ static bool unwind_by_prologue(struct unwind_state *state)
>>
>>   out:
>>          state->first = false;
>> -       unwind_state_fixup(state);
>> -       return !!__kernel_text_address(state->pc);
>> +       return unwind_state_fixup(state) || __kernel_text_address(state->pc);
>>   }
>>
>>   static bool next_frame(struct unwind_state *state)
>> diff --git a/arch/loongarch/mm/tlb.c b/arch/loongarch/mm/tlb.c
>> index da3681f131c8..8bad6b0cff59 100644
>> --- a/arch/loongarch/mm/tlb.c
>> +++ b/arch/loongarch/mm/tlb.c
>> @@ -251,7 +251,7 @@ static void output_pgtable_bits_defines(void)
>>   }
>>
>>   #ifdef CONFIG_NUMA
>> -static unsigned long pcpu_handlers[NR_CPUS];
>> +unsigned long pcpu_handlers[NR_CPUS];
>>   #endif
>>   extern long exception_handlers[VECSIZE * 128 / sizeof(long)];
>>
>> --
>> 2.34.3
>>
>>
  
Qing Zhang Jan. 17, 2023, 7:47 a.m. UTC | #5
On 2023/1/17 下午2:25, Huacai Chen wrote:
> On Thu, Jan 12, 2023 at 8:41 AM Jinyang He <hejinyang@loongson.cn> wrote:
>>
>> When exception is triggered, code flow go handle_\exception in some
>> cases. One of stackframe in this case as follows,
>>
>> high -> +-------+
>>          | REGS  |  <- a pt_regs
>>          |       |
>>          |       |  <- ex trigger
>>          | REGS  |  <- ex pt_regs   <-+
>>          |       |                    |
>>          |       |                    |
>> low  -> +-------+           ->unwind-+
>>
>> When unwinder unwind to handler_\exception it cannot go on prologue
>> analysis. It is asynchronous code flow, we should get the next frame
>> PC from regs->csr_era but not from regs->regs[1]. And we copy the
>> handler codes to eentry in the early time and copy the handler codes
>> to NUMA-relative memory named pcpu_handlers if NUMA is enabled. Thus,
>> unwinder cannot unwind normally. Therefore, try to give some hint in
>> handler_\exception and fixup it in unwind_next_frame.
>>
>> Reported-by: Qing Zhang <zhangqing@loongson.cn>
>> Signed-off-by: Jinyang He <hejinyang@loongson.cn>
>> ---
>>   arch/loongarch/include/asm/unwind.h     |   2 +-
>>   arch/loongarch/kernel/genex.S           |   3 +
>>   arch/loongarch/kernel/unwind_prologue.c | 100 +++++++++++++++++++++---
>>   arch/loongarch/mm/tlb.c                 |   2 +-
>>   4 files changed, 92 insertions(+), 15 deletions(-)
>>
>> diff --git a/arch/loongarch/include/asm/unwind.h b/arch/loongarch/include/asm/unwind.h
>> index cb428e1b19af..a38eb152aefb 100644
>> --- a/arch/loongarch/include/asm/unwind.h
>> +++ b/arch/loongarch/include/asm/unwind.h
>> @@ -22,7 +22,7 @@ struct unwind_state {
>>          char type; /* UNWINDER_XXX */
>>          struct stack_info stack_info;
>>          struct task_struct *task;
>> -       bool first, error, is_ftrace;
>> +       bool first, error, reset;
> Hi, Qing,
> Do you think is_reset is better than reset here?
> 
>>          int graph_idx;
>>          unsigned long sp, pc, ra;
>>   };
>> diff --git a/arch/loongarch/kernel/genex.S b/arch/loongarch/kernel/genex.S
>> index 75e5be807a0d..7e5c293ed89f 100644
>> --- a/arch/loongarch/kernel/genex.S
>> +++ b/arch/loongarch/kernel/genex.S
>> @@ -67,14 +67,17 @@ SYM_FUNC_END(except_vec_cex)
>>          .macro  BUILD_HANDLER exception handler prep
>>          .align  5
>>          SYM_FUNC_START(handle_\exception)
>> +       666:
>>          BACKUP_T0T1
>>          SAVE_ALL
>>          build_prep_\prep
>>          move    a0, sp
>>          la.abs  t0, do_\handler
>>          jirl    ra, t0, 0
>> +       668:
>>          RESTORE_ALL_AND_RET
> I think the 668 label should be after RESTORE_ALL_AND_RET, right?

This means that the value of ra after do_\handler is called, the 668 
label should come after jirl.

Thanks
-Qing
> 
> Huacai
> 
>>          SYM_FUNC_END(handle_\exception)
>> +       SYM_DATA(unwind_hint_\exception, .word 668b - 666b)
>>          .endm
>>
>>          BUILD_HANDLER ade ade badv
>> diff --git a/arch/loongarch/kernel/unwind_prologue.c b/arch/loongarch/kernel/unwind_prologue.c
>> index e6c3f2ee507c..e52c18c2b604 100644
>> --- a/arch/loongarch/kernel/unwind_prologue.c
>> +++ b/arch/loongarch/kernel/unwind_prologue.c
>> @@ -2,21 +2,100 @@
>>   /*
>>    * Copyright (C) 2022 Loongson Technology Corporation Limited
>>    */
>> +#include <linux/cpumask.h>
>>   #include <linux/ftrace.h>
>>   #include <linux/kallsyms.h>
>>
>>   #include <asm/inst.h>
>> +#include <asm/loongson.h>
>>   #include <asm/ptrace.h>
>> +#include <asm/setup.h>
>>   #include <asm/unwind.h>
>>
>> -static inline void unwind_state_fixup(struct unwind_state *state)
>> +extern const int unwind_hint_ade;
>> +extern const int unwind_hint_ale;
>> +extern const int unwind_hint_bp;
>> +extern const int unwind_hint_fpe;
>> +extern const int unwind_hint_fpu;
>> +extern const int unwind_hint_lsx;
>> +extern const int unwind_hint_lasx;
>> +extern const int unwind_hint_lbt;
>> +extern const int unwind_hint_ri;
>> +extern const int unwind_hint_watch;
>> +extern unsigned long eentry;
>> +#ifdef CONFIG_NUMA
>> +extern unsigned long pcpu_handlers[NR_CPUS];
>> +#endif
>> +
>> +static inline bool scan_handler(unsigned long entry_offset)
>>   {
>> -#ifdef CONFIG_DYNAMIC_FTRACE
>> -       static unsigned long ftrace = (unsigned long)ftrace_call + 4;
>> +       int idx, offset;
>>
>> -       if (state->pc == ftrace)
>> -               state->is_ftrace = true;
>> +       if (entry_offset >= EXCCODE_INT_START * VECSIZE)
>> +               return false;
>> +
>> +       idx = entry_offset / VECSIZE;
>> +       offset = entry_offset % VECSIZE;
>> +       switch (idx) {
>> +       case EXCCODE_ADE:
>> +               return offset == unwind_hint_ade;
>> +       case EXCCODE_ALE:
>> +               return offset == unwind_hint_ale;
>> +       case EXCCODE_BP:
>> +               return offset == unwind_hint_bp;
>> +       case EXCCODE_FPE:
>> +               return offset == unwind_hint_fpe;
>> +       case EXCCODE_FPDIS:
>> +               return offset == unwind_hint_fpu;
>> +       case EXCCODE_LSXDIS:
>> +               return offset == unwind_hint_lsx;
>> +       case EXCCODE_LASXDIS:
>> +               return offset == unwind_hint_lasx;
>> +       case EXCCODE_BTDIS:
>> +               return offset == unwind_hint_lbt;
>> +       case EXCCODE_INE:
>> +               return offset == unwind_hint_ri;
>> +       case EXCCODE_WATCH:
>> +               return offset == unwind_hint_watch;
>> +       default:
>> +               return false;
>> +       }
>> +}
>> +
>> +static inline bool fix_exceptions(unsigned long pc)
>> +{
>> +#ifdef CONFIG_NUMA
>> +       int cpu;
>> +
>> +       for_each_possible_cpu(cpu) {
>> +               if (!pcpu_handlers[cpu])
>> +                       continue;
>> +               if (scan_handler(pc - pcpu_handlers[cpu]))
>> +                       return true;
>> +       }
>>   #endif
>> +       return scan_handler(pc - eentry);
>> +}
>> +
>> +/*
>> + * As we meet ftrace_regs_entry, reset first flag like first doing
>> + * tracing. Prologue analysis will stop soon because PC is at entry.
>> + */
>> +static inline bool fix_ftrace(unsigned long pc)
>> +{
>> +#ifdef CONFIG_DYNAMIC_FTRACE
>> +       return pc == (unsigned long)ftrace_call + LOONGARCH_INSN_SIZE;
>> +#else
>> +       return false;
>> +#endif
>> +}
>> +
>> +static inline bool unwind_state_fixup(struct unwind_state *state)
>> +{
>> +       if (!fix_exceptions(state->pc) && !fix_ftrace(state->pc))
>> +               return false;
>> +       state->reset = true;
>> +       return true;
>>   }
>>
>>   /*
>> @@ -39,14 +118,10 @@ static bool unwind_by_prologue(struct unwind_state *state)
>>          if (state->sp >= info->end || state->sp < info->begin)
>>                  return false;
>>
>> -       if (state->is_ftrace) {
>> -               /*
>> -                * As we meet ftrace_regs_entry, reset first flag like first doing
>> -                * tracing. Prologue analysis will stop soon because PC is at entry.
>> -                */
>> +       if (state->reset) {
>>                  regs = (struct pt_regs *)state->sp;
>>                  state->first = true;
>> -               state->is_ftrace = false;
>> +               state->reset = false;
>>                  state->pc = regs->csr_era;
>>                  state->ra = regs->regs[1];
>>                  state->sp = regs->regs[3];
>> @@ -111,8 +186,7 @@ static bool unwind_by_prologue(struct unwind_state *state)
>>
>>   out:
>>          state->first = false;
>> -       unwind_state_fixup(state);
>> -       return !!__kernel_text_address(state->pc);
>> +       return unwind_state_fixup(state) || __kernel_text_address(state->pc);
>>   }
>>
>>   static bool next_frame(struct unwind_state *state)
>> diff --git a/arch/loongarch/mm/tlb.c b/arch/loongarch/mm/tlb.c
>> index da3681f131c8..8bad6b0cff59 100644
>> --- a/arch/loongarch/mm/tlb.c
>> +++ b/arch/loongarch/mm/tlb.c
>> @@ -251,7 +251,7 @@ static void output_pgtable_bits_defines(void)
>>   }
>>
>>   #ifdef CONFIG_NUMA
>> -static unsigned long pcpu_handlers[NR_CPUS];
>> +unsigned long pcpu_handlers[NR_CPUS];
>>   #endif
>>   extern long exception_handlers[VECSIZE * 128 / sizeof(long)];
>>
>> --
>> 2.34.3
>>
>>
  

Patch

diff --git a/arch/loongarch/include/asm/unwind.h b/arch/loongarch/include/asm/unwind.h
index cb428e1b19af..a38eb152aefb 100644
--- a/arch/loongarch/include/asm/unwind.h
+++ b/arch/loongarch/include/asm/unwind.h
@@ -22,7 +22,7 @@  struct unwind_state {
 	char type; /* UNWINDER_XXX */
 	struct stack_info stack_info;
 	struct task_struct *task;
-	bool first, error, is_ftrace;
+	bool first, error, reset;
 	int graph_idx;
 	unsigned long sp, pc, ra;
 };
diff --git a/arch/loongarch/kernel/genex.S b/arch/loongarch/kernel/genex.S
index 75e5be807a0d..7e5c293ed89f 100644
--- a/arch/loongarch/kernel/genex.S
+++ b/arch/loongarch/kernel/genex.S
@@ -67,14 +67,17 @@  SYM_FUNC_END(except_vec_cex)
 	.macro	BUILD_HANDLER exception handler prep
 	.align	5
 	SYM_FUNC_START(handle_\exception)
+	666:
 	BACKUP_T0T1
 	SAVE_ALL
 	build_prep_\prep
 	move	a0, sp
 	la.abs	t0, do_\handler
 	jirl	ra, t0, 0
+	668:
 	RESTORE_ALL_AND_RET
 	SYM_FUNC_END(handle_\exception)
+	SYM_DATA(unwind_hint_\exception, .word 668b - 666b)
 	.endm
 
 	BUILD_HANDLER ade ade badv
diff --git a/arch/loongarch/kernel/unwind_prologue.c b/arch/loongarch/kernel/unwind_prologue.c
index e6c3f2ee507c..e52c18c2b604 100644
--- a/arch/loongarch/kernel/unwind_prologue.c
+++ b/arch/loongarch/kernel/unwind_prologue.c
@@ -2,21 +2,100 @@ 
 /*
  * Copyright (C) 2022 Loongson Technology Corporation Limited
  */
+#include <linux/cpumask.h>
 #include <linux/ftrace.h>
 #include <linux/kallsyms.h>
 
 #include <asm/inst.h>
+#include <asm/loongson.h>
 #include <asm/ptrace.h>
+#include <asm/setup.h>
 #include <asm/unwind.h>
 
-static inline void unwind_state_fixup(struct unwind_state *state)
+extern const int unwind_hint_ade;
+extern const int unwind_hint_ale;
+extern const int unwind_hint_bp;
+extern const int unwind_hint_fpe;
+extern const int unwind_hint_fpu;
+extern const int unwind_hint_lsx;
+extern const int unwind_hint_lasx;
+extern const int unwind_hint_lbt;
+extern const int unwind_hint_ri;
+extern const int unwind_hint_watch;
+extern unsigned long eentry;
+#ifdef CONFIG_NUMA
+extern unsigned long pcpu_handlers[NR_CPUS];
+#endif
+
+static inline bool scan_handler(unsigned long entry_offset)
 {
-#ifdef CONFIG_DYNAMIC_FTRACE
-	static unsigned long ftrace = (unsigned long)ftrace_call + 4;
+	int idx, offset;
 
-	if (state->pc == ftrace)
-		state->is_ftrace = true;
+	if (entry_offset >= EXCCODE_INT_START * VECSIZE)
+		return false;
+
+	idx = entry_offset / VECSIZE;
+	offset = entry_offset % VECSIZE;
+	switch (idx) {
+	case EXCCODE_ADE:
+		return offset == unwind_hint_ade;
+	case EXCCODE_ALE:
+		return offset == unwind_hint_ale;
+	case EXCCODE_BP:
+		return offset == unwind_hint_bp;
+	case EXCCODE_FPE:
+		return offset == unwind_hint_fpe;
+	case EXCCODE_FPDIS:
+		return offset == unwind_hint_fpu;
+	case EXCCODE_LSXDIS:
+		return offset == unwind_hint_lsx;
+	case EXCCODE_LASXDIS:
+		return offset == unwind_hint_lasx;
+	case EXCCODE_BTDIS:
+		return offset == unwind_hint_lbt;
+	case EXCCODE_INE:
+		return offset == unwind_hint_ri;
+	case EXCCODE_WATCH:
+		return offset == unwind_hint_watch;
+	default:
+		return false;
+	}
+}
+
+static inline bool fix_exceptions(unsigned long pc)
+{
+#ifdef CONFIG_NUMA
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		if (!pcpu_handlers[cpu])
+			continue;
+		if (scan_handler(pc - pcpu_handlers[cpu]))
+			return true;
+	}
 #endif
+	return scan_handler(pc - eentry);
+}
+
+/*
+ * As we meet ftrace_regs_entry, reset first flag like first doing
+ * tracing. Prologue analysis will stop soon because PC is at entry.
+ */
+static inline bool fix_ftrace(unsigned long pc)
+{
+#ifdef CONFIG_DYNAMIC_FTRACE
+	return pc == (unsigned long)ftrace_call + LOONGARCH_INSN_SIZE;
+#else
+	return false;
+#endif
+}
+
+static inline bool unwind_state_fixup(struct unwind_state *state)
+{
+	if (!fix_exceptions(state->pc) && !fix_ftrace(state->pc))
+		return false;
+	state->reset = true;
+	return true;
 }
 
 /*
@@ -39,14 +118,10 @@  static bool unwind_by_prologue(struct unwind_state *state)
 	if (state->sp >= info->end || state->sp < info->begin)
 		return false;
 
-	if (state->is_ftrace) {
-		/*
-		 * As we meet ftrace_regs_entry, reset first flag like first doing
-		 * tracing. Prologue analysis will stop soon because PC is at entry.
-		 */
+	if (state->reset) {
 		regs = (struct pt_regs *)state->sp;
 		state->first = true;
-		state->is_ftrace = false;
+		state->reset = false;
 		state->pc = regs->csr_era;
 		state->ra = regs->regs[1];
 		state->sp = regs->regs[3];
@@ -111,8 +186,7 @@  static bool unwind_by_prologue(struct unwind_state *state)
 
 out:
 	state->first = false;
-	unwind_state_fixup(state);
-	return !!__kernel_text_address(state->pc);
+	return unwind_state_fixup(state) || __kernel_text_address(state->pc);
 }
 
 static bool next_frame(struct unwind_state *state)
diff --git a/arch/loongarch/mm/tlb.c b/arch/loongarch/mm/tlb.c
index da3681f131c8..8bad6b0cff59 100644
--- a/arch/loongarch/mm/tlb.c
+++ b/arch/loongarch/mm/tlb.c
@@ -251,7 +251,7 @@  static void output_pgtable_bits_defines(void)
 }
 
 #ifdef CONFIG_NUMA
-static unsigned long pcpu_handlers[NR_CPUS];
+unsigned long pcpu_handlers[NR_CPUS];
 #endif
 extern long exception_handlers[VECSIZE * 128 / sizeof(long)];