Re: BUG ? exc_page_fault() was optimized out of fred_hwexc() by gcc with default kernel build option (-O2).

[Xin Li <xin@zytor.com>]

On 1/18/2025 5:50 AM, Ethan Zhao wrote:
> Hi, Xin, Peter
> 
>    While checking the asm code of arch/x86/entry/entry_fred.o about 
> function fred_hwexc(),
> found the code was generated as following :
> 
> 0000000000000200 <fred_hwexc.constprop.0>:
>   200:   0f b6 87 a4 00 00 00    movzbl 0xa4(%rdi),%eax
>   207:   3c 0e                   cmp    $0xe,%al /* match X86_TRAP_PF */
>   209:   75 05                   jne    210 <fred_hwexc.constprop.0+0x10>
>   20b:   e9 00 00 00 00          jmp    210 <fred_hwexc.constprop.0+0x10>
>   210:   3c 0b                   cmp    $0xb,%al
>   212:   74 6a                   je     27e <fred_hwexc.constprop.0+0x7e>
>   214:   77 17                   ja     22d <fred_hwexc.constprop.0+0x2d>
>   216:   3c 06                   cmp    $0x6,%al
>   218:   0f 84 83 00 00 00       je     2a1 <fred_hwexc.constprop.0+0xa1>
>   21e:   76 29                   jbe    249 <fred_hwexc.constprop.0+0x49>
>   220:   3c 08                   cmp    $0x8,%al
>   222:   74 78                   je     29c <fred_hwexc.constprop.0+0x9c>
>   224:   3c 0a                   cmp    $0xa,%al
>   226:   75 18                   jne    240 <fred_hwexc.constprop.0+0x40>
>   228:   e9 00 00 00 00          jmp    22d <fred_hwexc.constprop.0+0x2d>
>   22d:   3c 11                   cmp    $0x11,%al
>   22f:   74 66                   je     297 <fred_hwexc.constprop.0+0x97>
>   231:   76 2c                   jbe    25f <fred_hwexc.constprop.0+0x5f>
>   233:   3c 13                   cmp    $0x13,%al
>   235:   74 5b                   je     292 <fred_hwexc.constprop.0+0x92>
>   237:   3c 15                   cmp    $0x15,%al
>   239:   75 1b                   jne    256 <fred_hwexc.constprop.0+0x56>
>   23b:   e9 00 00 00 00          jmp    240 <fred_hwexc.constprop.0+0x40>
>   240:   3c 07                   cmp    $0x7,%al
>   242:   75 49                   jne    28d <fred_hwexc.constprop.0+0x8d>
>   244:   e9 00 00 00 00          jmp    249 <fred_hwexc.constprop.0+0x49>
>   249:   3c 01                   cmp    $0x1,%al
>   24b:   74 3b                   je     288 <fred_hwexc.constprop.0+0x88>
>   24d:   3c 05                   cmp    $0x5,%al
>   24f:   75 1b                   jne    26c <fred_hwexc.constprop.0+0x6c>
>   251:   e9 00 00 00 00          jmp    256 <fred_hwexc.constprop.0+0x56>
>   256:   3c 12                   cmp    $0x12,%al
>   258:   75 33                   jne    28d <fred_hwexc.constprop.0+0x8d>
>   25a:   e9 00 00 00 00          jmp    25f <fred_hwexc.constprop.0+0x5f>
> 
> seems the following calling to exc_page_fault() was optimized out from 
> fred_hwexc() by gcc,
> 
> if(likely(regs->fred_ss.vector==X86_TRAP_PF))
> returnexc_page_fault(regs,error_code);
> 
> gcc version 11.4.0 (Ubuntu 11.4.0-1ubuntu1~22.04)
> 
> GNU objdump (GNU Binutils) 2.43
> 
> 
> default kernel config.
> .config:CONFIG_X86_FRED=y
> 
> my understanding, -O2 is the default kernel KBUILD_CFLAGS
> So, Are there any workaround needed to make the kernel works with 
> default build ?
> or just as Peter said in another loop, manually loading some event bits 
> to make the
> over-smart gcc behave normally ？or fall back to -O(ption)0 ?
> 
> Any idea, much appreciated !

This is an optimization done in the original code:

static noinstr void fred_hwexc(struct pt_regs *regs, unsigned long 
error_code)
{
         /* Optimize for #PF. That's the only exception which matters 
performance wise */
         if (likely(regs->fred_ss.vector == X86_TRAP_PF))
                 return exc_page_fault(regs, error_code);

         switch (regs->fred_ss.vector) {