Btrfs: don't do unnecessary delalloc flushes when relocating
authorFilipe Manana <fdmanana@suse.com>
Tue, 26 Apr 2016 14:39:32 +0000 (15:39 +0100)
committerFilipe Manana <fdmanana@suse.com>
Fri, 13 May 2016 00:59:16 +0000 (01:59 +0100)
commit9cfa3e34e20e6798a671236000d9e97c8aa5d318
tree34cfa1a9e68b0d1ab2a208c52334793855f05fba
parent578def7c50f236432ba140d35bb7ca4ef0a1b20b
Btrfs: don't do unnecessary delalloc flushes when relocating

Before we start the actual relocation process of a block group, we do
calls to flush delalloc of all inodes and then wait for ordered extents
to complete. However we do these flush calls just to make sure we don't
race with concurrent tasks that have actually already started to run
delalloc and have allocated an extent from the block group we want to
relocate, right before we set it to readonly mode, but have not yet
created the respective ordered extents. The flush calls make us wait
for such concurrent tasks because they end up calling
filemap_fdatawrite_range() (through btrfs_start_delalloc_roots() ->
__start_delalloc_inodes() -> btrfs_alloc_delalloc_work() ->
btrfs_run_delalloc_work()) which ends up serializing us with those tasks
due to attempts to lock the same pages (and the delalloc flush procedure
calls the allocator and creates the ordered extents before unlocking the
pages).

These flushing calls not only make us waste time (cpu, IO) but also reduce
the chances of writing larger extents (applications might be writing to
contiguous ranges and we flush before they finish dirtying the whole
ranges).

So make sure we don't flush delalloc and just wait for concurrent tasks
that have already started flushing delalloc and have allocated an extent
from the block group we are about to relocate.

This change also ends up fixing a race with direct IO writes that makes
relocation not wait for direct IO ordered extents. This race is
illustrated by the following diagram:

        CPU 1                                       CPU 2

 btrfs_relocate_block_group(bg X)

                                           starts direct IO write,
                                           target inode currently has no
                                           ordered extents ongoing nor
                                           dirty pages (delalloc regions),
                                           therefore the root for our inode
                                           is not in the list
                                           fs_info->ordered_roots

                                           btrfs_direct_IO()
                                             __blockdev_direct_IO()
                                               btrfs_get_blocks_direct()
                                                 btrfs_lock_extent_direct()
                                                   locks range in the io tree
                                                 btrfs_new_extent_direct()
                                                   btrfs_reserve_extent()
                                                     --> extent allocated
                                                         from bg X

   btrfs_inc_block_group_ro(bg X)

   btrfs_start_delalloc_roots()
     __start_delalloc_inodes()
       --> does nothing, no dealloc ranges
           in the inode's io tree so the
           inode's root is not in the list
           fs_info->delalloc_roots

   btrfs_wait_ordered_roots()
     --> does not find the inode's root in the
         list fs_info->ordered_roots

     --> ends up not waiting for the direct IO
         write started by the task at CPU 2

   relocate_block_group(rc->stage ==
     MOVE_DATA_EXTENTS)

     prepare_to_relocate()
       btrfs_commit_transaction()

     iterates the extent tree, using its
     commit root and moves extents into new
     locations

                                                   btrfs_add_ordered_extent_dio()
                                                     --> now a ordered extent is
                                                         created and added to the
                                                         list root->ordered_extents
                                                         and the root added to the
                                                         list fs_info->ordered_roots
                                                     --> this is too late and the
                                                         task at CPU 1 already
                                                         started the relocation

     btrfs_commit_transaction()

                                                   btrfs_finish_ordered_io()
                                                     btrfs_alloc_reserved_file_extent()
                                                       --> adds delayed data reference
                                                           for the extent allocated
                                                           from bg X

   relocate_block_group(rc->stage ==
     UPDATE_DATA_PTRS)

     prepare_to_relocate()
       btrfs_commit_transaction()
         --> delayed refs are run, so an extent
             item for the allocated extent from
             bg X is added to extent tree
         --> commit roots are switched, so the
             next scan in the extent tree will
             see the extent item

     sees the extent in the extent tree

When this happens the relocation produces the following warning when it
finishes:

[ 7260.832836] ------------[ cut here ]------------
[ 7260.834653] WARNING: CPU: 5 PID: 6765 at fs/btrfs/relocation.c:4318 btrfs_relocate_block_group+0x245/0x2a1 [btrfs]()
[ 7260.838268] Modules linked in: btrfs crc32c_generic xor ppdev raid6_pq psmouse sg acpi_cpufreq evdev i2c_piix4 tpm_tis serio_raw tpm i2c_core pcspkr parport_pc
[ 7260.850935] CPU: 5 PID: 6765 Comm: btrfs Not tainted 4.5.0-rc6-btrfs-next-28+ #1
[ 7260.852998] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014
[ 7260.852998]  0000000000000000 ffff88020bf57bc0 ffffffff812648b3 0000000000000000
[ 7260.852998]  0000000000000009 ffff88020bf57bf8 ffffffff81051608 ffffffffa03c1b2d
[ 7260.852998]  ffff8800b2bbb800 0000000000000000 ffff8800b17bcc58 ffff8800399dd000
[ 7260.852998] Call Trace:
[ 7260.852998]  [<ffffffff812648b3>] dump_stack+0x67/0x90
[ 7260.852998]  [<ffffffff81051608>] warn_slowpath_common+0x99/0xb2
[ 7260.852998]  [<ffffffffa03c1b2d>] ? btrfs_relocate_block_group+0x245/0x2a1 [btrfs]
[ 7260.852998]  [<ffffffff810516d4>] warn_slowpath_null+0x1a/0x1c
[ 7260.852998]  [<ffffffffa03c1b2d>] btrfs_relocate_block_group+0x245/0x2a1 [btrfs]
[ 7260.852998]  [<ffffffffa039d9de>] btrfs_relocate_chunk.isra.29+0x66/0xdb [btrfs]
[ 7260.852998]  [<ffffffffa039f314>] btrfs_balance+0xde1/0xe4e [btrfs]
[ 7260.852998]  [<ffffffff8127d671>] ? debug_smp_processor_id+0x17/0x19
[ 7260.852998]  [<ffffffffa03a9583>] btrfs_ioctl_balance+0x255/0x2d3 [btrfs]
[ 7260.852998]  [<ffffffffa03ac96a>] btrfs_ioctl+0x11e0/0x1dff [btrfs]
[ 7260.852998]  [<ffffffff811451df>] ? handle_mm_fault+0x443/0xd63
[ 7260.852998]  [<ffffffff81491817>] ? _raw_spin_unlock+0x31/0x44
[ 7260.852998]  [<ffffffff8108b36a>] ? arch_local_irq_save+0x9/0xc
[ 7260.852998]  [<ffffffff811876ab>] vfs_ioctl+0x18/0x34
[ 7260.852998]  [<ffffffff81187cb2>] do_vfs_ioctl+0x550/0x5be
[ 7260.852998]  [<ffffffff81190c30>] ? __fget_light+0x4d/0x71
[ 7260.852998]  [<ffffffff81187d77>] SyS_ioctl+0x57/0x79
[ 7260.852998]  [<ffffffff81492017>] entry_SYSCALL_64_fastpath+0x12/0x6b
[ 7260.893268] ---[ end trace eb7803b24ebab8ad ]---

This is because at the end of the first stage, in relocate_block_group(),
we commit the current transaction, which makes delayed refs run, the
commit roots are switched and so the second stage will find the extent
item that the ordered extent added to the delayed refs. But this extent
was not moved (ordered extent completed after first stage finished), so
at the end of the relocation our block group item still has a positive
used bytes counter, triggering a warning at the end of
btrfs_relocate_block_group(). Later on when trying to read the extent
contents from disk we hit a BUG_ON() due to the inability to map a block
with a logical address that belongs to the block group we relocated and
is no longer valid, resulting in the following trace:

[ 7344.885290] BTRFS critical (device sdi): unable to find logical 12845056 len 4096
[ 7344.887518] ------------[ cut here ]------------
[ 7344.888431] kernel BUG at fs/btrfs/inode.c:1833!
[ 7344.888431] invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
[ 7344.888431] Modules linked in: btrfs crc32c_generic xor ppdev raid6_pq psmouse sg acpi_cpufreq evdev i2c_piix4 tpm_tis serio_raw tpm i2c_core pcspkr parport_pc
[ 7344.888431] CPU: 0 PID: 6831 Comm: od Tainted: G        W       4.5.0-rc6-btrfs-next-28+ #1
[ 7344.888431] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014
[ 7344.888431] task: ffff880215818600 ti: ffff880204684000 task.ti: ffff880204684000
[ 7344.888431] RIP: 0010:[<ffffffffa037c88c>]  [<ffffffffa037c88c>] btrfs_merge_bio_hook+0x54/0x6b [btrfs]
[ 7344.888431] RSP: 0018:ffff8802046878f0  EFLAGS: 00010282
[ 7344.888431] RAX: 00000000ffffffea RBX: 0000000000001000 RCX: 0000000000000001
[ 7344.888431] RDX: ffff88023ec0f950 RSI: ffffffff8183b638 RDI: 00000000ffffffff
[ 7344.888431] RBP: ffff880204687908 R08: 0000000000000001 R09: 0000000000000000
[ 7344.888431] R10: ffff880204687770 R11: ffffffff82f2d52d R12: 0000000000001000
[ 7344.888431] R13: ffff88021afbfee8 R14: 0000000000006208 R15: ffff88006cd199b0
[ 7344.888431] FS:  00007f1f9e1d6700(0000) GS:ffff88023ec00000(0000) knlGS:0000000000000000
[ 7344.888431] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 7344.888431] CR2: 00007f1f9dc8cb60 CR3: 000000023e3b6000 CR4: 00000000000006f0
[ 7344.888431] Stack:
[ 7344.888431]  0000000000001000 0000000000001000 ffff880204687b98 ffff880204687950
[ 7344.888431]  ffffffffa0395c8f ffffea0004d64d48 0000000000000000 0000000000001000
[ 7344.888431]  ffffea0004d64d48 0000000000001000 0000000000000000 0000000000000000
[ 7344.888431] Call Trace:
[ 7344.888431]  [<ffffffffa0395c8f>] submit_extent_page+0xf5/0x16f [btrfs]
[ 7344.888431]  [<ffffffffa03970ac>] __do_readpage+0x4a0/0x4f1 [btrfs]
[ 7344.888431]  [<ffffffffa039680d>] ? btrfs_create_repair_bio+0xcb/0xcb [btrfs]
[ 7344.888431]  [<ffffffffa037eeb4>] ? btrfs_writepage_start_hook+0xbc/0xbc [btrfs]
[ 7344.888431]  [<ffffffff8108df55>] ? trace_hardirqs_on+0xd/0xf
[ 7344.888431]  [<ffffffffa039728c>] __do_contiguous_readpages.constprop.26+0xc2/0xe4 [btrfs]
[ 7344.888431]  [<ffffffffa037eeb4>] ? btrfs_writepage_start_hook+0xbc/0xbc [btrfs]
[ 7344.888431]  [<ffffffffa039739b>] __extent_readpages.constprop.25+0xed/0x100 [btrfs]
[ 7344.888431]  [<ffffffff81129d24>] ? lru_cache_add+0xe/0x10
[ 7344.888431]  [<ffffffffa0397ea8>] extent_readpages+0x160/0x1aa [btrfs]
[ 7344.888431]  [<ffffffffa037eeb4>] ? btrfs_writepage_start_hook+0xbc/0xbc [btrfs]
[ 7344.888431]  [<ffffffff8115daad>] ? alloc_pages_current+0xa9/0xcd
[ 7344.888431]  [<ffffffffa037cdc9>] btrfs_readpages+0x1f/0x21 [btrfs]
[ 7344.888431]  [<ffffffff81128316>] __do_page_cache_readahead+0x168/0x1fc
[ 7344.888431]  [<ffffffff811285a0>] ondemand_readahead+0x1f6/0x207
[ 7344.888431]  [<ffffffff811285a0>] ? ondemand_readahead+0x1f6/0x207
[ 7344.888431]  [<ffffffff8111cf34>] ? pagecache_get_page+0x2b/0x154
[ 7344.888431]  [<ffffffff8112870e>] page_cache_sync_readahead+0x3d/0x3f
[ 7344.888431]  [<ffffffff8111dbf7>] generic_file_read_iter+0x197/0x4e1
[ 7344.888431]  [<ffffffff8117773a>] __vfs_read+0x79/0x9d
[ 7344.888431]  [<ffffffff81178050>] vfs_read+0x8f/0xd2
[ 7344.888431]  [<ffffffff81178a38>] SyS_read+0x50/0x7e
[ 7344.888431]  [<ffffffff81492017>] entry_SYSCALL_64_fastpath+0x12/0x6b
[ 7344.888431] Code: 8d 4d e8 45 31 c9 45 31 c0 48 8b 00 48 c1 e2 09 48 8b 80 80 fc ff ff 4c 89 65 e8 48 8b b8 f0 01 00 00 e8 1d 42 02 00 85 c0 79 02 <0f> 0b 4c 0
[ 7344.888431] RIP  [<ffffffffa037c88c>] btrfs_merge_bio_hook+0x54/0x6b [btrfs]
[ 7344.888431]  RSP <ffff8802046878f0>
[ 7344.970544] ---[ end trace eb7803b24ebab8ae ]---

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
fs/btrfs/ctree.h
fs/btrfs/extent-tree.c
fs/btrfs/inode.c
fs/btrfs/relocation.c