Message-Id: <20180105125034.762581d16dc8572990a9e5b6@linux-foundation.org>
Date:   Fri, 5 Jan 2018 12:50:34 -0800
From:   Andrew Morton <akpm@...ux-foundation.org>
To:     "Gang He" <ghe@...e.com>
Cc:     <jlbec@...lplan.org>, <jiangqi903@...il.com>,
        <mfasheh@...sity.com>, <ocfs2-devel@....oracle.com>,
        <linux-kernel@...r.kernel.org>
Subject: Re: [Ocfs2-devel] [PATCH v2] ocfs2: try a blocking lock before
 return AOP_TRUNCATED_PAGE

On Thu, 04 Jan 2018 23:31:12 -0700 "Gang He" <ghe@...e.com> wrote:

> Happy new year.
> Could you help to pick up this patch, which is used to fix a old patch 1cce4df04f37.
> If we have not this patch, some multiple node test cases will trigger softlockup problems,
> also make HA communication daemon (e.g. corosync) timeout and the node will has to be fenced.

I have the below queued for 4.16-rc1.

Is the problem seriosu enough to push this into 4.15?  Should the fix
be backported into -stable kernels?

From: Gang He <ghe@...e.com>
Subject: ocfs2: try a blocking lock before return AOP_TRUNCATED_PAGE

If we can't get inode lock immediately in the function
ocfs2_inode_lock_with_page() when reading a page, we should not return
directly here, since this will lead to a softlockup problem when the
kernel is configured with CONFIG_PREEMPT is not set.  The method is to get
a blocking lock and immediately unlock before returning, this can avoid
CPU resource waste due to lots of retries, and benefits fairness in
getting lock among multiple nodes, increase efficiency in case modifying
the same file frequently from multiple nodes.

The softlockup crash (when set /proc/sys/kernel/softlockup_panic to 1)
looks like:

Kernel panic - not syncing: softlockup: hung tasks
CPU: 0 PID: 885 Comm: multi_mmap Tainted: G L 4.12.14-6.1-default #1
Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
Call Trace:
  <IRQ>
  dump_stack+0x5c/0x82
  panic+0xd5/0x21e
  watchdog_timer_fn+0x208/0x210
  ? watchdog_park_threads+0x70/0x70
  __hrtimer_run_queues+0xcc/0x200
  hrtimer_interrupt+0xa6/0x1f0
  smp_apic_timer_interrupt+0x34/0x50
  apic_timer_interrupt+0x96/0xa0
  </IRQ>
 RIP: 0010:unlock_page+0x17/0x30
 RSP: 0000:ffffaf154080bc88 EFLAGS: 00000246 ORIG_RAX: ffffffffffffff10
 RAX: dead000000000100 RBX: fffff21e009f5300 RCX: 0000000000000004
 RDX: dead0000000000ff RSI: 0000000000000202 RDI: fffff21e009f5300
 RBP: 0000000000000000 R08: 0000000000000000 R09: ffffaf154080bb00
 R10: ffffaf154080bc30 R11: 0000000000000040 R12: ffff993749a39518
 R13: 0000000000000000 R14: fffff21e009f5300 R15: fffff21e009f5300
  ocfs2_inode_lock_with_page+0x25/0x30 [ocfs2]
  ocfs2_readpage+0x41/0x2d0 [ocfs2]
  ? pagecache_get_page+0x30/0x200
  filemap_fault+0x12b/0x5c0
  ? recalc_sigpending+0x17/0x50
  ? __set_task_blocked+0x28/0x70
  ? __set_current_blocked+0x3d/0x60
  ocfs2_fault+0x29/0xb0 [ocfs2]
  __do_fault+0x1a/0xa0
  __handle_mm_fault+0xbe8/0x1090
  handle_mm_fault+0xaa/0x1f0
  __do_page_fault+0x235/0x4b0
  trace_do_page_fault+0x3c/0x110
  async_page_fault+0x28/0x30
 RIP: 0033:0x7fa75ded638e
 RSP: 002b:00007ffd6657db18 EFLAGS: 00010287
 RAX: 000055c7662fb700 RBX: 0000000000000001 RCX: 000055c7662fb700
 RDX: 0000000000001770 RSI: 00007fa75e909000 RDI: 000055c7662fb700
 RBP: 0000000000000003 R08: 000000000000000e R09: 0000000000000000
 R10: 0000000000000483 R11: 00007fa75ded61b0 R12: 00007fa75e90a770
 R13: 000000000000000e R14: 0000000000001770 R15: 0000000000000000

About performance improvement, we can see the testing time is reduced, and
CPU utilization decreases, the detailed data is as follows.  I ran
multi_mmap test case in ocfs2-test package in a three nodes cluster.

Before applying this patch:
  PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
 2754 ocfs2te+  20   0  170248   6980   4856 D 80.73 0.341   0:18.71 multi_mmap
 1505 root      rt   0  222236 123060  97224 S 2.658 6.015   0:01.44 corosync
    5 root      20   0       0      0      0 S 1.329 0.000   0:00.19 kworker/u8:0
   95 root      20   0       0      0      0 S 1.329 0.000   0:00.25 kworker/u8:1
 2728 root      20   0       0      0      0 S 0.997 0.000   0:00.24 jbd2/sda1-33
 2721 root      20   0       0      0      0 S 0.664 0.000   0:00.07 ocfs2dc-3C8CFD4
 2750 ocfs2te+  20   0  142976   4652   3532 S 0.664 0.227   0:00.28 mpirun

ocfs2test@...node2:~>multiple_run.sh -i ens3 -k ~/linux-4.4.21-69.tar.gz -o
~/ocfs2mullog -C hacluster -s pcmk -n tb-node2,tb-node1,tb-node3 -d
/dev/sda1 -b 4096 -c 32768 -t multi_mmap /mnt/shared
Tests with "-b 4096 -C 32768"
Thu Dec 28 14:44:52 CST 2017
multi_mmap..................................................Passed.
Runtime 783 seconds.

After apply this patch:

  PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
 2508 ocfs2te+  20   0  170248   6804   4680 R 54.00 0.333   0:55.37 multi_mmap
  155 root      20   0       0      0      0 S 2.667 0.000   0:01.20 kworker/u8:3
   95 root      20   0       0      0      0 S 2.000 0.000   0:01.58 kworker/u8:1
 2504 ocfs2te+  20   0  142976   4604   3480 R 1.667 0.225   0:01.65 mpirun
    5 root      20   0       0      0      0 S 1.000 0.000   0:01.36 kworker/u8:0
 2482 root      20   0       0      0      0 S 1.000 0.000   0:00.86 jbd2/sda1-33
  299 root       0 -20       0      0      0 S 0.333 0.000   0:00.13 kworker/2:1H
  335 root       0 -20       0      0      0 S 0.333 0.000   0:00.17 kworker/1:1H
  535 root      20   0   12140   7268   1456 S 0.333 0.355   0:00.34 haveged
 1282 root      rt   0  222284 123108  97224 S 0.333 6.017   0:01.33 corosync

ocfs2test@...node2:~>multiple_run.sh -i ens3 -k ~/linux-4.4.21-69.tar.gz -o
~/ocfs2mullog -C hacluster -s pcmk -n tb-node2,tb-node1,tb-node3 -d
/dev/sda1 -b 4096 -c 32768 -t multi_mmap /mnt/shared
Tests with "-b 4096 -C 32768"
Thu Dec 28 15:04:12 CST 2017
multi_mmap..................................................Passed.
Runtime 487 seconds.

Link: http://lkml.kernel.org/r/1514447305-30814-1-git-send-email-ghe@suse.com
Fixes: 1cce4df04f37 ("ocfs2: do not lock/unlock() inode DLM lock")
Signed-off-by: Gang He <ghe@...e.com>
Reviewed-by: Eric Ren <zren@...e.com>
Acked-by: alex chen <alex.chen@...wei.com>
Acked-by: piaojun <piaojun@...wei.com>
Cc: Mark Fasheh <mfasheh@...sity.com>
Cc: Joel Becker <jlbec@...lplan.org>
Cc: Junxiao Bi <junxiao.bi@...cle.com>
Cc: Joseph Qi <jiangqi903@...il.com>
Cc: Changwei Ge <ge.changwei@....com>
Signed-off-by: Andrew Morton <akpm@...ux-foundation.org>
---

 fs/ocfs2/dlmglue.c |    9 +++++++++
 1 file changed, 9 insertions(+)

diff -puN fs/ocfs2/dlmglue.c~ocfs2-try-a-blocking-lock-before-return-aop_truncated_page fs/ocfs2/dlmglue.c
--- a/fs/ocfs2/dlmglue.c~ocfs2-try-a-blocking-lock-before-return-aop_truncated_page
+++ a/fs/ocfs2/dlmglue.c
@@ -2529,6 +2529,15 @@ int ocfs2_inode_lock_with_page(struct in
 	ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK);
 	if (ret == -EAGAIN) {
 		unlock_page(page);
+		/*
+		 * If we can't get inode lock immediately, we should not return
+		 * directly here, since this will lead to a softlockup problem.
+		 * The method is to get a blocking lock and immediately unlock
+		 * before returning, this can avoid CPU resource waste due to
+		 * lots of retries, and benefits fairness in getting lock.
+		 */
+		if (ocfs2_inode_lock(inode, ret_bh, ex) == 0)
+			ocfs2_inode_unlock(inode, ex);
 		ret = AOP_TRUNCATED_PAGE;
 	}
 
_

Powered by blists - more mailing lists