Date:   Wed, 16 May 2018 09:44:59 -0700
From:   Dave Hansen <dave.hansen@...el.com>
To:     Fenghua Yu <fenghua.yu@...el.com>
Cc:     Thomas Gleixner <tglx@...utronix.de>, Ingo Molnar <mingo@...e.hu>,
        "H. Peter Anvin" <hpa@...ux.intel.com>,
        Ashok Raj <ashok.raj@...el.com>,
        Ravi V Shankar <ravi.v.shankar@...el.com>,
        Tony Luck <tony.luck@...el.com>,
        Rafael Wysocki <rafael.j.wysocki@...el.com>,
        Arjan van de Ven <arjan@...radead.org>,
        Alan Cox <alan@...ux.intel.com>, x86 <x86@...nel.org>,
        linux-kernel <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 03/15] x86/split_lock: Handle #AC exception for split lock
 in kernel mode

On 05/15/2018 10:21 AM, Fenghua Yu wrote:
> On Tue, May 15, 2018 at 08:51:24AM -0700, Dave Hansen wrote:
>> On 05/14/2018 11:52 AM, Fenghua Yu wrote:
>>> +#define	delay_ms	1
>>
>> That seems like a dangerously-generic name that should not be a #define
>> anyway.
> 
> Sure. I will change it to
> #define split_lock_delay_ms	1

Why not:

static unsigned int reenable_split_lock_delay_ms = 1;

?

>>> +/* Will the faulting instruction be re-executed? */
>>> +static bool re_execute(struct pt_regs *regs)
>>> +{
>>> +	/*
>>> +	 * The only reason for generating #AC from kernel is because of
>>> +	 * split lock. The kernel faulting instruction will be re-executed.
>>> +	 */
>>> +	if (!user_mode(regs))
>>> +		return true;
>>> +
>>> +	return false;
>>> +}
>>
>> This helper with a single user is a bit unnecessary.  Just open-code
>> this and move the comments into the caller.
> 
> In this patch, this helper is only used for checking kernel mode.
> Then in patch #11, this helper will add checking user mode code.
> It would be better to have a helper defined and called.

Then introduce the helper later, or call this out in a comment or the
patch description, please.

>>> +/*
>>> + * #AC handler for kernel split lock is called by generic #AC handler.
>>> + *
>>> + * Disable #AC for split lock on this CPU so that the faulting instruction
>>> + * gets executed. The #AC for split lock is re-enabled later.
>>> + */
>>> +bool do_split_lock_exception(struct pt_regs *regs, unsigned long error_code)
>>> +{
>>> +	unsigned long delay = msecs_to_jiffies(delay_ms);
>>> +	unsigned long address = read_cr2(); /* Get the faulting address */
>>> +	int this_cpu = smp_processor_id();
>>
>> How does this end up working?  This seems to depend on this handler not
>> getting preempted.
> 
> Maybe change the handler to:
> 
> this_cpu = task_cpu(current);
> Then disable split lock on this_cpu.
> Re-enable split lock on this_cpu (already in this way).
> 
> Does this sound better?

Actually, as I look at it, interrupts *are* still disabled here, so you
are OK.  But, you can do a local_irq_enable() once you get all of the
per-cpu state settled and go to start handling the fault if you are
going to do anything that takes an appreciable amount of time.

>>> diff --git a/arch/x86/kernel/traps.c b/arch/x86/kernel/traps.c
>>> index 03f3d7695dac..c07b817bbbe9 100644
>>> --- a/arch/x86/kernel/traps.c
>>> +++ b/arch/x86/kernel/traps.c
>>> @@ -61,6 +61,7 @@
>>>  #include <asm/mpx.h>
>>>  #include <asm/vm86.h>
>>>  #include <asm/umip.h>
>>> +#include <asm/cpu.h>
>>>  
>>>  #ifdef CONFIG_X86_64
>>>  #include <asm/x86_init.h>
>>> @@ -286,10 +287,21 @@ static void do_error_trap(struct pt_regs *regs, long error_code, char *str,
>>>  			  unsigned long trapnr, int signr)
>>>  {
>>>  	siginfo_t info;
>>> +	int ret;
>>>  
>>>  	RCU_LOCKDEP_WARN(!rcu_is_watching(), "entry code didn't wake RCU");
>>>  
>>>  	/*
>>> +	 * #AC exception could be handled by split lock handler.
>>> +	 * If the handler can't handle the exception, go to generic #AC handler.
>>> +	 */
>>> +	if (trapnr == X86_TRAP_AC) {
>>> +		ret = do_split_lock_exception(regs, error_code);
>>> +		if (ret)
>>> +			return;
>>> +	}
>>
>> Why are you hooking into do_error_trap()?  Shouldn't you just be
>> installing do_split_lock_exception() as *the* #AC handler and put it in
>> the IDT?
> 
> Split lock is not the only reason that causes #AC. #AC can be caused
> by user turning on AC bit in EFLAGS, which is just cache line misalignment
> and is different from split lock.
> 
> So split lock is sharing the handler with another #AC case and can't
> be installed seperately from previous #AC handler, right?

There are lots of exceptions that use do_error_trap().  I'm suggesting
that you make an IDT entry for X86_TRAP_AC that does not use
do_error_trap() since you need something different in there now.

See:

> #define DO_ERROR(trapnr, signr, str, name)                              \
> dotraplinkage void do_##name(struct pt_regs *regs, long error_code)     \
> {                                                                       \
>         do_error_trap(regs, error_code, str, trapnr, signr);            \
> }
> 
> DO_ERROR(X86_TRAP_DE,     SIGFPE,  "divide error",              divide_error)
> DO_ERROR(X86_TRAP_OF,     SIGSEGV, "overflow",                  overflow)
> DO_ERROR(X86_TRAP_UD,     SIGILL,  "invalid opcode",            invalid_op)
> DO_ERROR(X86_TRAP_OLD_MF, SIGFPE,  "coprocessor segment overrun",coprocessor_segment_overrun)
> DO_ERROR(X86_TRAP_TS,     SIGSEGV, "invalid TSS",               invalid_TSS)
> DO_ERROR(X86_TRAP_NP,     SIGBUS,  "segment not present",       segment_not_present)
> DO_ERROR(X86_TRAP_SS,     SIGBUS,  "stack segment",             stack_segment)
> DO_ERROR(X86_TRAP_AC,     SIGBUS,  "alignment check",           alignment_check)

Look at do_general_protection(), for instance.

Powered by blists - more mailing lists