Re: [PATCH] mremap: Increase LATENCY_LIMIT of mremap to reduce the number of TLB shootdowns

From: Nadav Amit <nadav.amit@gmail.com>
To: Mel Gorman <mgorman@techsingularity.net>
Cc: Andrew Morton <akpm@linux-foundation.org>,
	Dave Hansen <dave.hansen@intel.com>,
	mhocko@kernel.org, vbabka@suse.cz, Aaron Lu <aaron.lu@intel.com>,
	linux-kernel@vger.kernel.org, linux-mm@kvack.org
Subject: Re: [PATCH] mremap: Increase LATENCY_LIMIT of mremap to reduce the number of TLB shootdowns
Date: Wed, 6 Jun 2018 11:20:08 -0700	[thread overview]
Message-ID: <70CFD0DC-FEBD-4B51-9AE9-0786AF66149F@gmail.com> (raw)
In-Reply-To: <20180606174723.bag3o55fvqp6nbvc@techsingularity.net>

Mel Gorman <mgorman@techsingularity.net> wrote:

> On Wed, Jun 06, 2018 at 08:55:15AM -0700, Nadav Amit wrote:
>>> -#define LATENCY_LIMIT	(64 * PAGE_SIZE)
>>> +#define LATENCY_LIMIT	(PMD_SIZE)
>>> 
>>> unsigned long move_page_tables(struct vm_area_struct *vma,
>>> 		unsigned long old_addr, struct vm_area_struct *new_vma,
>> 
>> This LATENCY_LIMIT is only used in move_page_tables() in the following
>> manner:
>> 
>>  next = (new_addr + PMD_SIZE) & PMD_MASK;
>>  if (extent > next - new_addr)
>>      extent = next - new_addr;
>>  if (extent > LATENCY_LIMIT)
>>      extent = LATENCY_LIMIT;
>> 
>> If LATENCY_LIMIT is to be changed to PMD_SIZE, then IIUC the last condition
>> is not required, and LATENCY_LIMIT can just be removed (assuming there is no
>> underflow case that hides somewhere).
> 
> I see no problem removing it other than we may forget that we ever limited
> PTE lock hold times for any reason. I'm skeptical it will matter unless
> mremap-intensive workloads are a lot more common than I believe.

I have no opinion regarding the behavior change. It is just that code with
no effect is oftentimes confusing. A comment (if needed) can replace the
code, and git history would provide how it was once supported.