Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753322AbeAKDBc convert rfc822-to-8bit (ORCPT + 1 other); Wed, 10 Jan 2018 22:01:32 -0500 Received: from smtp.h3c.com ([60.191.123.56]:43418 "EHLO h3cmg01-ex.h3c.com" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S1753054AbeAKDBb (ORCPT ); Wed, 10 Jan 2018 22:01:31 -0500 From: Changwei Ge To: Gang He , "jlbec@evilplan.org" , "mfasheh@versity.com" CC: "ocfs2-devel@oss.oracle.com" , "linux-kernel@vger.kernel.org" Subject: Re: [Ocfs2-devel] [PATCH v2 2/2] ocfs2: add trimfs lock to avoid duplicated trims in cluster Thread-Topic: [Ocfs2-devel] [PATCH v2 2/2] ocfs2: add trimfs lock to avoid duplicated trims in cluster Thread-Index: AdOJ3PVpr6PIaeXlR1CPvv/Xoq4ERA== Date: Thu, 11 Jan 2018 03:00:31 +0000 Message-ID: <63ADC13FD55D6546B7DECE290D39E373F290EED7@H3CMLB12-EX.srv.huawei-3com.com> References: <1513228484-2084-1-git-send-email-ghe@suse.com> <1513228484-2084-2-git-send-email-ghe@suse.com> <63ADC13FD55D6546B7DECE290D39E373F290E6C5@H3CMLB12-EX.srv.huawei-3com.com> <5A5647C2020000F9000A2929@prv-mh.provo.novell.com> <63ADC13FD55D6546B7DECE290D39E373F290E878@H3CMLB12-EX.srv.huawei-3com.com> <5A5657D8020000F9000A2987@prv-mh.provo.novell.com> <63ADC13FD55D6546B7DECE290D39E373F290E930@H3CMLB12-EX.srv.huawei-3com.com> <5A57372A020000F9000A2E01@prv-mh.provo.novell.com> Accept-Language: en-US, zh-CN Content-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-originating-ip: [10.125.136.231] Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 8BIT MIME-Version: 1.0 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Return-Path: On 2018/1/11 10:07, Gang He wrote: > Hi Changwei, > > >>>> >> On 2018/1/10 18:14, Gang He wrote: >>> Hi Changwei, >>> >>> >>>>>> >>>> On 2018/1/10 17:05, Gang He wrote: >>>>> Hi Changwei, >>>>> >>>>> >>>>>>>> >>>>>> Hi Gang, >>>>>> >>>>>> On 2017/12/14 13:16, Gang He wrote: >>>>>>> As you know, ocfs2 has support trim the underlying disk via >>>>>>> fstrim command. But there is a problem, ocfs2 is a shared disk >>>>>>> cluster file system, if the user configures a scheduled fstrim >>>>>>> job on each file system node, this will trigger multiple nodes >>>>>>> trim a shared disk simultaneously, it is very wasteful for CPU >>>>>>> and IO consumption, also might negatively affect the lifetime >>>>>>> of poor-quality SSD devices. >>>>>>> Then, we introduce a trimfs dlm lock to communicate with each >>>>>>> other in this case, which will make only one fstrim command to >>>>>>> do the trimming on a shared disk among the cluster, the fstrim >>>>>>> commands from the other nodes should wait for the first fstrim >>>>>>> to finish and returned success directly, to avoid running a the >>>>>>> same trim on the shared disk again. >>>>>>> >>>>>>> Compare with first version, I change the fstrim commands' returned >>>>>>> value and behavior in case which meets a fstrim command is running >>>>>>> on a shared disk. >>>>>>> >>>>>>> Signed-off-by: Gang He >>>>>>> --- >>>>>>> fs/ocfs2/alloc.c | 44 ++++++++++++++++++++++++++++++++++++++++++++ >>>>>>> 1 file changed, 44 insertions(+) >>>>>>> >>>>>>> diff --git a/fs/ocfs2/alloc.c b/fs/ocfs2/alloc.c >>>>>>> index ab5105f..5c9c3e2 100644 >>>>>>> --- a/fs/ocfs2/alloc.c >>>>>>> +++ b/fs/ocfs2/alloc.c >>>>>>> @@ -7382,6 +7382,7 @@ int ocfs2_trim_fs(struct super_block *sb, struct >>>>>> fstrim_range *range) >>>>>>> struct buffer_head *gd_bh = NULL; >>>>>>> struct ocfs2_dinode *main_bm; >>>>>>> struct ocfs2_group_desc *gd = NULL; >>>>>>> + struct ocfs2_trim_fs_info info, *pinfo = NULL; >>>>>> >>>>>> I think *pinfo* is not necessary. >>>>> This pointer is necessary, since it can be NULL or non-NULL depend on the >>>> code logic. >>>> >>>> This point is OK for me. >>>> >>>>> >>>>>>> >>>>>>> start = range->start >> osb->s_clustersize_bits; >>>>>>> len = range->len >> osb->s_clustersize_bits; >>>>>>> @@ -7419,6 +7420,42 @@ int ocfs2_trim_fs(struct super_block *sb, struct >>>>>> fstrim_range *range) >>>>>>> >>>>>>> trace_ocfs2_trim_fs(start, len, minlen); >>>>>>> >>>>>>> + ocfs2_trim_fs_lock_res_init(osb); >>>>>>> + ret = ocfs2_trim_fs_lock(osb, NULL, 1); >>>>>> >>>>>> I don't get why try to lock here and if fails, acquire the same lock again >>>>>> later but wait until granted. >>>>> Please think about the user case, the patch is only used to handle this >>>> case. >>>>> When the administer configures a fstrim schedule task on each node, then >>>> each node will trigger a fstrim on shared disks concurrently. >>>>> In this case, we should avoid duplicated fstrim on a shared disk since this >>>> will waste CPU/IO resources and affect SSD lifetime sometimes. >>>> >>>> I'm not worrying about that trimfs will affect SSD's lifetime quite a lot, >>>> since physical-logical address converting table resides in RAM while SSD is >>>> working. >>>> And that table won't be at a big scale. My point here is not affecting this >>>> patch. Just a tip here. >>> This depend on SSD firmware implementation, but for secure-trim, it really >> possibly affect SSD lifetime. >>> >>>>> Firstly, we use try_lock to get fstrim dlm lock to identify if there is any >>>> other node which is doing fstrim on the disk. >>>>> If not, this node is the first one, this node should do fstrim operation on >>>> the disk. >>>>> If yes, this node is not the first one, this node should wait until the >>>> first node is done for fstrim operation, then return the result from DLM >>>> lock's value. >>>>> >>>>>> Can it just acquire the _trimfs_ lock as a blocking one directly here? >>>>> We can not do a blocking lock directly, since we need to identify if there >>>> is any other node has being do fstrim operation when this node start to do >>>> fstrim. >>>> >>>> Thanks for your elaboration. >>>> >>>> Well how about the third node trying to trimming fs too? >>>> It needs LVB from the second node. >>>> But it seems that the second node can't provide a valid LVB. >>>> So the third node will perform trimfs once more. >>> No, the second node does not change DLM lock's value, but the DLM lock's >> value is still valid. >>> The third node also refer to this DLM lock's value, then do the same logic >> like the second node. >> >> Hi Gang, >> I don't see any places where ocfs2_lock_res::ocfs2_lock_res_ops::set_lvb is >> set while flag LOCK_TYPE_USES_LVB is added. >> >> Are you sure below code path can work well? > Yes, have done a full testing on two and three nodes. > >> ocfs2_process_blocked_lock >> ocfs2_unblock_lock >> Reference to ::set_lvb since LOCK_TYPE_USES_LVB is set. >> > the set_lvb callback function is not necessary, if we update DLM lock value by ourselves before unlock. I think this may relates to *LOCK_TYPE_REQUIRES_REFRESH* flag. Actually, I don't see why this flag is necessary to _orphan scan_. Why can't _orphan scan_ also set LVB during ocfs2_process_blocked_lock->ocfs2_unblock_lock? And it seems that _orphan scan_ also doesn't need to persist any stuff in LVB into disk. Thanks, Changwei > By the way, the code is transparent to the underlying DLM stack (o2cb or pcmk). True. > > Thanks > Gang > >> Thanks, >> Changwei >> >>> >>>> >>>> IOW, three nodes are trying to trimming fs concurrently. Is your patch able >>>> to handle such a scenario? >>> Yes, the patch can handle this case. >>> >>>> >>>> Even the second lock request with QUEUE set just follows >>>> ocfs2_trim_fs_lock_res_uninit() will not get rid of concurrent trimfs. >>>> >>>>> >>>>>> >>>>>>> + if (ret < 0) { >>>>>>> + if (ret != -EAGAIN) { >>>>>>> + mlog_errno(ret); >>>>>>> + ocfs2_trim_fs_lock_res_uninit(osb); >>>>>>> + goto out_unlock; >>>>>>> + } >>>>>>> + >>>>>>> + mlog(ML_NOTICE, "Wait for trim on device (%s) to " >>>>>>> + "finish, which is running from another node.\n", >>>>>>> + osb->dev_str); >>>>>>> + ret = ocfs2_trim_fs_lock(osb, &info, 0); >>>>>>> + if (ret < 0) { >>>>>>> + mlog_errno(ret); >>>>>>> + ocfs2_trim_fs_lock_res_uninit(osb); >>>>>> >>>>>> In ocfs2_trim_fs_lock_res_uninit(), you drop lock. But it is never granted. >>>>>> Still need to drop lock resource? >>>>> Yes, we need to init/uninit fstrim dlm lock resource for each time. >>>>> Otherwise, trylock does not work, this is a little different from other dlm >>>> lock usage in ocfs2. >>>> >>>> This point is OK for now, too. >>>>> >>>>>> >>>>>>> + goto out_unlock; >>>>>>> + } >>>>>>> + >>>>>>> + if (info.tf_valid && info.tf_success && >>>>>>> + info.tf_start == start && info.tf_len == len && >>>>>>> + info.tf_minlen == minlen) { >>>>>>> + /* Avoid sending duplicated trim to a shared device */ >>>>>>> + mlog(ML_NOTICE, "The same trim on device (%s) was " >>>>>>> + "just done from node (%u), return.\n", >>>>>>> + osb->dev_str, info.tf_nodenum); >>>>>>> + range->len = info.tf_trimlen; >>>>>>> + goto out_trimunlock; >>>>>>> + } >>>>>>> + } >>>>>>> + >>>>>>> + info.tf_nodenum = osb->node_num; >>>>>>> + info.tf_start = start; >>>>>>> + info.tf_len = len; >>>>>>> + info.tf_minlen = minlen; >>>>>> >>>>>> If we faild during dong trimfs, I think we should not cache above info in >>>>>> LVB. >>>>> It is necessary, if the second node is waiting the first node, the first >>>> node fails to do fstrim, >>>>> the first node should update dlm lock's value, then the second node can get >>>> the latest dlm lock value (rather than the last time DLM lock value), >>>>> the second node will do the fstrim again, since the first node has failed. >>>> >>>> Yes, it makes scene. >>>>> >>>>>> BTW, it seems that this patch is on top of 'try lock' patches which you >>>>>> previously sent out. >>>>>> Are they related? >>>>> try lock patch is related to non-block aio support for ocfs2. >>>>> >>>>> Thanks >>>>> Gang >>>>>> >>>>>> Thanks, >>>>>> Changwei >>>>>> >>>>>>> + >>>>>>> /* Determine first and last group to examine based on start and len */ >>>>>>> first_group = ocfs2_which_cluster_group(main_bm_inode, start); >>>>>>> if (first_group == osb->first_cluster_group_blkno) >>>>>>> @@ -7463,6 +7500,13 @@ int ocfs2_trim_fs(struct super_block *sb, struct >>>>>> fstrim_range *range) >>>>>>> group += ocfs2_clusters_to_blocks(sb, osb->bitmap_cpg); >>>>>>> } >>>>>>> range->len = trimmed * sb->s_blocksize; >>>>>>> + >>>>>>> + info.tf_trimlen = range->len; >>>>>>> + info.tf_success = (ret ? 0 : 1); >>>>>>> + pinfo = &info; >>>>>>> +out_trimunlock: >>>>>>> + ocfs2_trim_fs_unlock(osb, pinfo); >>>>>>> + ocfs2_trim_fs_lock_res_uninit(osb); >>>>>>> out_unlock: >>>>>>> ocfs2_inode_unlock(main_bm_inode, 0); >>>>>>> brelse(main_bm_bh); >>>>>>> >>>>> >>> >