From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753869Ab3BDHRS (ORCPT ); Mon, 4 Feb 2013 02:17:18 -0500 Received: from mail-pa0-f43.google.com ([209.85.220.43]:47112 "EHLO mail-pa0-f43.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752746Ab3BDHRQ (ORCPT ); Mon, 4 Feb 2013 02:17:16 -0500 From: Michel Lespinasse To: Andrea Arcangeli , Rik van Riel , Mel Gorman , Hugh Dickins , Andrew Morton , linux-mm@kvack.org Cc: linux-kernel@vger.kernel.org Subject: [PATCH v2 0/3] fixes for large mm_populate() and munlock() operations Date: Sun, 3 Feb 2013 23:17:09 -0800 Message-Id: <1359962232-20811-1-git-send-email-walken@google.com> X-Mailer: git-send-email 1.8.1 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org These 3 changes are to improve the handling of large mm_populate and munlock operations. They apply on top of mmotm (in particular, they depend on both my prior mm_populate work and Kirill's "thp: avoid dumping huge zero page" change). - Patch 1 fixes an integer overflow issue when populating 2^32 pages. The nr_pages argument to get_user_pages would overflow, resulting in 0 pages being processed per iteration. I am proposing to simply convert the nr_pages argument to a long. - Patch 2 accelerates populating regions with THP pages. get_user_pages() can increment the address by a huge page size in this case instead of a small page size, and avoid repeated mm->page_table_lock acquisitions. This fixes an issue reported by Roman Dubtsov where populating regions via mmap MAP_POPULATE was significantly slower than doing so by touching pages from userspace. - Patch 3 is a similar acceleration for the munlock case. I would actually like to get Andrea's attention on this one, as I can't explain how munlock_vma_page() is safe against racing with split_huge_page(). Note that patches 1-2 are logically independent of patch 3, so if the discussion of patch 3 takes too long I would ask Andrew to consider merging patches 1-2 first. Changes since v1: - Andrew accepted patch 1 into his -mm tree but suggested the nr_pages argument type should actually be unsigned long; I am sending this as a "fix" for the previous patch 1 to be collapsed over the previous one. - In patch 2, I am adding a separate follow_page_mask() function so that the callers to the original follow_page() don't have to be modified to ignore the returned page_mask (following another suggestion from Andrew). Also the page_mask argument type was changed to unsigned int. - In patch 3, I similarly changed the page_mask values to unsigned int. Michel Lespinasse (3): fix mm: use long type for page counts in mm_populate() and get_user_pages() mm: accelerate mm_populate() treatment of THP pages mm: accelerate munlock() treatment of THP pages include/linux/hugetlb.h | 2 +- include/linux/mm.h | 24 +++++++++++++++++------- mm/hugetlb.c | 8 ++++---- mm/internal.h | 2 +- mm/memory.c | 43 +++++++++++++++++++++++++++++-------------- mm/mlock.c | 34 ++++++++++++++++++++++------------ mm/nommu.c | 6 ++++-- 7 files changed, 78 insertions(+), 41 deletions(-) -- 1.8.1