Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1757307AbcK2W2m (ORCPT ); Tue, 29 Nov 2016 17:28:42 -0500 Received: from mail-lf0-f67.google.com ([209.85.215.67]:34380 "EHLO mail-lf0-f67.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1757237AbcK2W2d (ORCPT ); Tue, 29 Nov 2016 17:28:33 -0500 MIME-Version: 1.0 In-Reply-To: References: <20161115165538.878698352bd45e212751b57a@gmail.com> <20161115170038.75e127739b66f850e50d7fc1@gmail.com> From: Dan Streetman Date: Tue, 29 Nov 2016 17:27:46 -0500 X-Google-Sender-Auth: iWcH1u7NKkiWvO-TlvTs6cQxH6A Message-ID: Subject: Re: [PATCH 3/3] z3fold: discourage use of pages that weren't compacted To: Vitaly Wool Cc: Linux-MM , linux-kernel , Andrew Morton Content-Type: text/plain; charset=UTF-8 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 4155 Lines: 88 On Mon, Nov 28, 2016 at 9:14 AM, Vitaly Wool wrote: > On Fri, Nov 25, 2016 at 7:25 PM, Dan Streetman wrote: >> On Tue, Nov 15, 2016 at 11:00 AM, Vitaly Wool wrote: >>> If a z3fold page couldn't be compacted, we don't want it to be >>> used for next object allocation in the first place. >> >> why? !compacted can only mean 1) already compact or 2) middle chunks >> is mapped. #1 is as good compaction-wise as the page can get, so do >> you mean that if a page couldn't be compacted because of #2, we >> shouldn't use it for next allocation? if so, that isn't quite what >> this patch does. >> >>> It makes more >>> sense to add it to the end of the relevant unbuddied list. If that >>> page gets compacted later, it will be added to the beginning of >>> the list then. >>> >>> This simple idea gives 5-7% improvement in randrw fio tests and >>> about 10% improvement in fio sequential read/write. >> >> i don't understand why there is any improvement - the unbuddied lists >> are grouped by the amount of free chunks, so all pages in a specific >> unbuddied list should have exactly that number of free chunks >> available, and it shouldn't matter if a page gets put into the front >> or back...where is the performance improvement coming from? > > When the next attempt to compact this page comes, it's less likely > it's locked so the wait times are slightly lower in average. which wait time? for a page with the middle chunk mapped compact should exit immediately...? > > ~vitaly > >>> Signed-off-by: Vitaly Wool >>> --- >>> mm/z3fold.c | 22 +++++++++++++++++----- >>> 1 file changed, 17 insertions(+), 5 deletions(-) >>> >>> diff --git a/mm/z3fold.c b/mm/z3fold.c >>> index ffd9353..e282ba0 100644 >>> --- a/mm/z3fold.c >>> +++ b/mm/z3fold.c >>> @@ -539,11 +539,19 @@ static void z3fold_free(struct z3fold_pool *pool, unsigned long handle) >>> free_z3fold_page(zhdr); >>> atomic64_dec(&pool->pages_nr); >>> } else { >>> - z3fold_compact_page(zhdr); >>> + int compacted = z3fold_compact_page(zhdr); >>> /* Add to the unbuddied list */ >>> spin_lock(&pool->lock); >>> freechunks = num_free_chunks(zhdr); >>> - list_add(&zhdr->buddy, &pool->unbuddied[freechunks]); >>> + /* >>> + * If the page has been compacted, we want to use it >>> + * in the first place. >>> + */ >>> + if (compacted) >>> + list_add(&zhdr->buddy, &pool->unbuddied[freechunks]); >>> + else >>> + list_add_tail(&zhdr->buddy, >>> + &pool->unbuddied[freechunks]); >>> spin_unlock(&pool->lock); >>> z3fold_page_unlock(zhdr); >>> } >>> @@ -672,12 +680,16 @@ static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries) >>> spin_lock(&pool->lock); >>> list_add(&zhdr->buddy, &pool->buddied); >>> } else { >>> - z3fold_compact_page(zhdr); >>> + int compacted = z3fold_compact_page(zhdr); >>> /* add to unbuddied list */ >>> spin_lock(&pool->lock); >>> freechunks = num_free_chunks(zhdr); >>> - list_add(&zhdr->buddy, >>> - &pool->unbuddied[freechunks]); >>> + if (compacted) >>> + list_add(&zhdr->buddy, >>> + &pool->unbuddied[freechunks]); >>> + else >>> + list_add_tail(&zhdr->buddy, >>> + &pool->unbuddied[freechunks]); >>> } >>> } >>> >>> -- >>> 2.4.2