Re: [PATCH v3 RFC 14/14] mm: speedup page alloc for MPOL_PREFERRED_MANY by adding a NO_SLOWPATH gfp bit

From: Feng Tang <feng.tang@intel.com>
To: Michal Hocko <mhocko@suse.com>
Cc: "linux-mm@kvack.org" <linux-mm@kvack.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	Andrew Morton <akpm@linux-foundation.org>,
	Andrea Arcangeli <aarcange@redhat.com>,
	David Rientjes <rientjes@google.com>,
	Mel Gorman <mgorman@techsingularity.net>,
	Mike Kravetz <mike.kravetz@oracle.com>,
	Randy Dunlap <rdunlap@infradead.org>,
	Vlastimil Babka <vbabka@suse.cz>,
	"Hansen, Dave" <dave.hansen@intel.com>,
	"Widawsky, Ben" <ben.widawsky@intel.com>,
	Andi leen <ak@linux.intel.com>,
	"Williams, Dan J" <dan.j.williams@intel.com>
Subject: Re: [PATCH v3 RFC 14/14] mm: speedup page alloc for MPOL_PREFERRED_MANY by adding a NO_SLOWPATH gfp bit
Date: Wed, 3 Mar 2021 21:46:44 +0800	[thread overview]
Message-ID: <20210303134644.GC78458@shbuild999.sh.intel.com> (raw)
In-Reply-To: <20210303131832.GB78458@shbuild999.sh.intel.com>

On Wed, Mar 03, 2021 at 09:18:32PM +0800, Tang, Feng wrote:
> On Wed, Mar 03, 2021 at 01:32:11PM +0100, Michal Hocko wrote:
> > On Wed 03-03-21 20:18:33, Feng Tang wrote:
> > > On Wed, Mar 03, 2021 at 08:07:17PM +0800, Tang, Feng wrote:
> > > > Hi Michal,
> > > > 
> > > > On Wed, Mar 03, 2021 at 12:39:57PM +0100, Michal Hocko wrote:
> > > > > On Wed 03-03-21 18:20:58, Feng Tang wrote:
> > > > > > When doing broader test, we noticed allocation slowness in one test
> > > > > > case that malloc memory with size which is slightly bigger than free
> > > > > > memory of targeted nodes, but much less then the total free memory
> > > > > > of system.
> > > > > > 
> > > > > > The reason is the code enters the slowpath of __alloc_pages_nodemask(),
> > > > > > which takes quite some time. As alloc_pages_policy() will give it a 2nd
> > > > > > try with NULL nodemask, so there is no need to enter the slowpath for
> > > > > > the first try. Add a new gfp bit to skip the slowpath, so that user cases
> > > > > > like this can leverage.
> > > > > > 
> > > > > > With it, the malloc in such case is much accelerated as it never enters
> > > > > > the slowpath.
> > > > > > 
> > > > > > Adding a new gfp_mask bit is generally not liked, and another idea is to
> > > > > > add another nodemask to struct 'alloc_context', so it has 2: 'preferred-nmask'
> > > > > > and 'fallback-nmask', and they will be tried in turn if not NULL, with
> > > > > > it we can call __alloc_pages_nodemask() only once.
> > > > > 
> > > > > Yes, it is very much disliked. Is there any reason why you cannot use
> > > > > GFP_NOWAIT for that purpose?
> > > > 
> > > > I did try that at the first place, but it didn't obviously change the slowness.
> > > > I assumed the direct claim was still involved as GFP_NOWAIT only impact kswapd
> > > > reclaim.
> > 
> > I assume you haven't really created gfp mask correctly. What was the
> > exact gfp mask you have used?
> 
> The testcase is a malloc with multi-preferred-node policy, IIRC, the gfp
> mask is HIGHUSER_MOVABLE originally, and code here ORs (__GFP_RETRY_MAYFAIL | __GFP_NOWARN).
> 
> As GFP_WAIT == __GFP_KSWAPD_RECLAIM, in this test case, the bit is already set.
> 
> > > 
> > > One thing I tried which can fix the slowness is:
> > > 
> > > +	gfp_mask &= ~(__GFP_DIRECT_RECLAIM | __GFP_KSWAPD_RECLAIM);
> > > 
> > > which explicitly clears the 2 kinds of reclaim. And I thought it's too
> > > hacky and didn't mention it in the commit log.
> > 
> > Clearing __GFP_DIRECT_RECLAIM would be the right way to achieve
> > GFP_NOWAIT semantic. Why would you want to exclude kswapd as well? 
> 
> When I tried gfp_mask &= ~__GFP_DIRECT_RECLAIM, the slowness couldn't
> be fixed.

I just double checked by rerun the test, 'gfp_mask &= ~__GFP_DIRECT_RECLAIM'
can also accelerate the allocation much! though is still a little slower than
this patch. Seems I've messed some of the tries, and sorry for the confusion!

Could this be used as the solution? or the adding another fallback_nodemask way?
but the latter will change the current API quite a bit.

Thanks,
Feng