pnfsblock: fix writeback deadlock
Peng Tao [Fri, 23 Sep 2011 01:50:17 +0000 (21:50 -0400)]
commit 7542274519b3ba87555410c66e8356ac1e3bc9b3 upstream.

We should check if the sector is already initialized before
trying to grab the page from page cache. Otherwise when two
pages of the same block are written back by two threads each
calling from writepage_locked, it can cause deadlock like bellow.

 [ 1080.972099] INFO: task kswapd0:25 blocked for more than 120 seconds.
 [ 1080.972377] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
 [ 1080.972812] kswapd0         D ffff88000c4926c0     0    25      2 0x00000000
 [ 1080.972816]  ffff88000df276b0 0000000000000046 ffff88000df27640 ffffffff81013ba7
 [ 1080.972821]  ffff88000c492310 ffff88000df27fd8 ffff88000df27fd8 00000000001d3440
 [ 1080.972824]  ffff88000c378000 ffff88000c492310 ffff8800175d3d40 ffff880017fc75a8
 [ 1080.972828] Call Trace:
 [ 1080.972860]  [<ffffffff81013ba7>] ? read_tsc+0x9/0x19
 [ 1080.972877]  [<ffffffff810e0b23>] ? lock_page+0x2b/0x2b
 [ 1080.972899]  [<ffffffff81475a1d>] io_schedule+0x63/0x7e
 [ 1080.972902]  [<ffffffff810e0b31>] sleep_on_page+0xe/0x12
 [ 1080.972905]  [<ffffffff81475fe8>] __wait_on_bit_lock+0x46/0x8f
 [ 1080.972916]  [<ffffffff810822d7>] ? lock_release_holdtime.part.7+0x6b/0x72
 [ 1080.972919]  [<ffffffff810e0af6>] __lock_page+0x66/0x68
 [ 1080.972928]  [<ffffffff81072705>] ? autoremove_wake_function+0x3d/0x3d
 [ 1080.972932]  [<ffffffff810e0b1f>] lock_page+0x27/0x2b
 [ 1080.972934]  [<ffffffff810e0bcf>] find_lock_page+0x34/0x57
 [ 1080.972937]  [<ffffffff810e1738>] find_or_create_page+0x34/0x8a
 [ 1080.972947]  [<ffffffffa034245b>] bl_write_pagelist+0x205/0x6da [blocklayoutdriver]
 [ 1080.972951]  [<ffffffffa034145d>] ? bl_free_lseg+0x38/0x38 [blocklayoutdriver]
 [ 1080.972995]  [<ffffffffa02e27b9>] ? nfs_write_rpcsetup+0x118/0x123 [nfs]
 [ 1080.973033]  [<ffffffffa030246b>] pnfs_generic_pg_writepages+0x10b/0x1f4 [nfs]
 [ 1080.973089]  [<ffffffffa02deaae>] nfs_pageio_doio+0x1a/0x43 [nfs]
 [ 1080.973098]  [<ffffffffa02df035>] nfs_pageio_complete+0x16/0x2d [nfs]
 [ 1080.973108]  [<ffffffffa02e2d8f>] nfs_writepage_locked+0xa0/0xbf [nfs]
 [ 1080.973119]  [<ffffffffa02e36a1>] nfs_writepage+0x16/0x2b [nfs]
 [ 1080.973122]  [<ffffffff810e8762>] ? clear_page_dirty_for_io+0x87/0x9a
 [ 1080.973133]  [<ffffffff810efc5b>] shrink_page_list+0x39b/0x6c8
 [ 1080.973139]  [<ffffffff810f03bb>] shrink_inactive_list+0x22c/0x39e
 [ 1080.973144]  [<ffffffff810822d7>] ? lock_release_holdtime.part.7+0x6b/0x72
 [ 1080.973148]  [<ffffffff810f0c33>] shrink_zone+0x445/0x588
 [ 1080.973152]  [<ffffffff810f1a11>] balance_pgdat+0x2c2/0x56b
 [ 1080.973170]  [<ffffffff81254208>] ? __bitmap_weight+0x34/0x80
 [ 1080.973175]  [<ffffffff810f1f78>] kswapd+0x2be/0x2fa
 [ 1080.973179]  [<ffffffff810726c8>] ? __init_waitqueue_head+0x4b/0x4b
 [ 1080.973183]  [<ffffffff810f1cba>] ? balance_pgdat+0x56b/0x56b
 [ 1080.973187]  [<ffffffff81071f69>] kthread+0xa8/0xb0
 [ 1080.973200]  [<ffffffff814806b4>] kernel_thread_helper+0x4/0x10
 [ 1080.973205]  [<ffffffff81071ec1>] ? __init_kthread_worker+0x5a/0x5a
 [ 1080.973210]  [<ffffffff814806b0>] ? gs_change+0x13/0x13
 [ 1080.973213] no locks held by kswapd0/25.

Signed-off-by: Peng Tao <peng_tao@emc.com>
Signed-off-by: Jim Rees <rees@umich.edu>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

fs/nfs/blocklayout/blocklayout.c

index 323be71..e3c381d 100644 (file)
@@ -533,6 +533,11 @@ bl_write_pagelist(struct nfs_write_data *wdata, int sync)
 fill_invalid_ext:
                dprintk("%s need to zero %d pages\n", __func__, npg_zero);
                for (;npg_zero > 0; npg_zero--) {
+                       if (bl_is_sector_init(be->be_inval, isect)) {
+                               dprintk("isect %llu already init\n",
+                                       (unsigned long long)isect);
+                               goto next_page;
+                       }
                        /* page ref released in bl_end_io_write_zero */
                        index = isect >> PAGE_CACHE_SECTOR_SHIFT;
                        dprintk("%s zero %dth page: index %lu isect %llu\n",
@@ -552,8 +557,7 @@ fill_invalid_ext:
                         * PageUptodate: It was read before
                         * sector_initialized: already written out
                         */
-                       if (PageDirty(page) || PageWriteback(page) ||
-                           bl_is_sector_init(be->be_inval, isect)) {
+                       if (PageDirty(page) || PageWriteback(page)) {
                                print_page(page);
                                unlock_page(page);
                                page_cache_release(page);