platform/kernel/linux-rpi.git
4 years agobtrfs: rename struct btrfs_clone_extent_info to a more generic name
Filipe Manana [Tue, 8 Sep 2020 10:27:22 +0000 (11:27 +0100)]
btrfs: rename struct btrfs_clone_extent_info to a more generic name

Now that we can use btrfs_clone_extent_info to convey information for a
new prealloc extent as well, and not just for existing extents that are
being cloned, rename it to btrfs_replace_extent_info, which reflects the
fact that this is now more generic and it is used to replace all existing
extents in a file range with the extent described by the structure.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: remove item_size member of struct btrfs_clone_extent_info
Filipe Manana [Tue, 8 Sep 2020 10:27:21 +0000 (11:27 +0100)]
btrfs: remove item_size member of struct btrfs_clone_extent_info

The value of item_size of struct btrfs_clone_extent_info is always set to
the size of a non-inline file extent item, and in fact the infrastructure
that uses this structure (btrfs_punch_hole_range()) does not work with
inline file extents at all (and it is not supposed to).

So just remove that field from the structure and use directly
sizeof(struct btrfs_file_extent_item) instead. Also assert that the
file extent type is not inline at btrfs_insert_clone_extent().

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: fix metadata reservation for fallocate that leads to transaction aborts
Filipe Manana [Tue, 8 Sep 2020 10:27:20 +0000 (11:27 +0100)]
btrfs: fix metadata reservation for fallocate that leads to transaction aborts

When doing an fallocate(), specially a zero range operation, we assume
that reserving 3 units of metadata space is enough, that at most we touch
one leaf in subvolume/fs tree for removing existing file extent items and
inserting a new file extent item. This assumption is generally true for
most common use cases. However when we end up needing to remove file extent
items from multiple leaves, we can end up failing with -ENOSPC and abort
the current transaction, turning the filesystem to RO mode. When this
happens a stack trace like the following is dumped in dmesg/syslog:

[ 1500.620934] ------------[ cut here ]------------
[ 1500.620938] BTRFS: Transaction aborted (error -28)
[ 1500.620973] WARNING: CPU: 2 PID: 30807 at fs/btrfs/inode.c:9724 __btrfs_prealloc_file_range+0x512/0x570 [btrfs]
[ 1500.620974] Modules linked in: btrfs intel_rapl_msr intel_rapl_common kvm_intel (...)
[ 1500.621010] CPU: 2 PID: 30807 Comm: xfs_io Tainted: G        W         5.9.0-rc3-btrfs-next-67 #1
[ 1500.621012] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014
[ 1500.621023] RIP: 0010:__btrfs_prealloc_file_range+0x512/0x570 [btrfs]
[ 1500.621026] Code: 8b 40 50 f0 48 (...)
[ 1500.621028] RSP: 0018:ffffb05fc8803ca0 EFLAGS: 00010286
[ 1500.621030] RAX: 0000000000000000 RBX: ffff9608af276488 RCX: 0000000000000000
[ 1500.621032] RDX: 0000000000000001 RSI: 0000000000000027 RDI: 00000000ffffffff
[ 1500.621033] RBP: ffffb05fc8803d90 R08: 0000000000000001 R09: 0000000000000001
[ 1500.621035] R10: 0000000000000000 R11: 0000000000000000 R12: 0000000003200000
[ 1500.621037] R13: 00000000ffffffe4 R14: ffff9608af275fe8 R15: ffff9608af275f60
[ 1500.621039] FS:  00007fb5b2368ec0(0000) GS:ffff9608b6600000(0000) knlGS:0000000000000000
[ 1500.621041] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 1500.621043] CR2: 00007fb5b2366fb8 CR3: 0000000202d38005 CR4: 00000000003706e0
[ 1500.621046] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ 1500.621047] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[ 1500.621049] Call Trace:
[ 1500.621076]  btrfs_prealloc_file_range+0x10/0x20 [btrfs]
[ 1500.621087]  btrfs_fallocate+0xccd/0x1280 [btrfs]
[ 1500.621108]  vfs_fallocate+0x14d/0x290
[ 1500.621112]  ksys_fallocate+0x3a/0x70
[ 1500.621117]  __x64_sys_fallocate+0x1a/0x20
[ 1500.621120]  do_syscall_64+0x33/0x80
[ 1500.621123]  entry_SYSCALL_64_after_hwframe+0x44/0xa9
[ 1500.621126] RIP: 0033:0x7fb5b248c477
[ 1500.621128] Code: 89 7c 24 08 (...)
[ 1500.621130] RSP: 002b:00007ffc7bee9060 EFLAGS: 00000293 ORIG_RAX: 000000000000011d
[ 1500.621132] RAX: ffffffffffffffda RBX: 0000000000000002 RCX: 00007fb5b248c477
[ 1500.621134] RDX: 0000000000000000 RSI: 0000000000000010 RDI: 0000000000000003
[ 1500.621136] RBP: 0000557718faafd0 R08: 0000000000000000 R09: 0000000000000000
[ 1500.621137] R10: 0000000003200000 R11: 0000000000000293 R12: 0000000000000010
[ 1500.621139] R13: 0000557718faafb0 R14: 0000557718faa480 R15: 0000000000000003
[ 1500.621151] irq event stamp: 1026217
[ 1500.621154] hardirqs last  enabled at (1026223): [<ffffffffba965570>] console_unlock+0x500/0x5c0
[ 1500.621156] hardirqs last disabled at (1026228): [<ffffffffba9654c7>] console_unlock+0x457/0x5c0
[ 1500.621159] softirqs last  enabled at (1022486): [<ffffffffbb6003dc>] __do_softirq+0x3dc/0x606
[ 1500.621161] softirqs last disabled at (1022477): [<ffffffffbb4010b2>] asm_call_on_stack+0x12/0x20
[ 1500.621162] ---[ end trace 2955b08408d8b9d4 ]---
[ 1500.621167] BTRFS: error (device sdj) in __btrfs_prealloc_file_range:9724: errno=-28 No space left

When we use fallocate() internally, for reserving an extent for a space
cache, inode cache or relocation, we can't hit this problem since either
there aren't any file extent items to remove from the subvolume tree or
there is at most one.

When using plain fallocate() it's very unlikely, since that would require
having many file extent items representing holes for the target range and
crossing multiple leafs - we attempt to increase the range (merge) of such
file extent items when punching holes, so at most we end up with 2 file
extent items for holes at leaf boundaries.

However when using the zero range operation of fallocate() for a large
range (100+ MiB for example) that's fairly easy to trigger. The following
example reproducer triggers the issue:

  $ cat reproducer.sh
  #!/bin/bash

  umount /dev/sdj &> /dev/null
  mkfs.btrfs -f -n 16384 -O ^no-holes /dev/sdj > /dev/null
  mount /dev/sdj /mnt/sdj

  # Create a 100M file with many file extent items. Punch a hole every 8K
  # just to speedup the file creation - we could do 4K sequential writes
  # followed by fsync (or O_SYNC) as well, but that takes a lot of time.
  file_size=$((100 * 1024 * 1024))
  xfs_io -f -c "pwrite -S 0xab -b 10M 0 $file_size" /mnt/sdj/foobar
  for ((i = 0; i < $file_size; i += 8192)); do
      xfs_io -c "fpunch $i 4096" /mnt/sdj/foobar
  done

  # Force a transaction commit, so the zero range operation will be forced
  # to COW all metadata extents it need to touch.
  sync

  xfs_io -c "fzero 0 $file_size" /mnt/sdj/foobar

  umount /mnt/sdj

  $ ./reproducer.sh
  wrote 104857600/104857600 bytes at offset 0
  100 MiB, 10 ops; 0.0669 sec (1.458 GiB/sec and 149.3117 ops/sec)
  fallocate: No space left on device

  $ dmesg
  <shows the same stack trace pasted before>

To fix this use the existing infrastructure that hole punching and
extent cloning use for replacing a file range with another extent. This
deals with doing the removal of file extent items and inserting the new
one using an incremental approach, reserving more space when needed and
always ensuring we don't leave an implicit hole in the range in case
we need to do multiple iterations and a crash happens between iterations.

A test case for fstests will follow up soon.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: remove unused function calc_global_rsv_need_space()
YueHaibing [Wed, 9 Sep 2020 13:51:42 +0000 (21:51 +0800)]
btrfs: remove unused function calc_global_rsv_need_space()

It is not used since commit 0096420adb03 ("btrfs: do not
account global reserve in can_overcommit").

Reviewed-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: move btrfs_dev_replace_update_device_in_mapping_tree to drop declaration
Anand Jain [Fri, 4 Sep 2020 17:34:36 +0000 (01:34 +0800)]
btrfs: move btrfs_dev_replace_update_device_in_mapping_tree to drop declaration

The function is short and simple, we can get rid of the declaration as
it's not necessary for a static function. Move it before its first
caller.  No functional changes.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: simplify gotos in open_seed_device
Anand Jain [Fri, 4 Sep 2020 17:34:35 +0000 (01:34 +0800)]
btrfs: simplify gotos in open_seed_device

The function does not have a common exit block and returns immediatelly
so there's no point having the goto. Remove the two cases.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: remove unnecessary tmp variable in btrfs_assign_next_active_device()
Anand Jain [Fri, 4 Sep 2020 17:34:34 +0000 (01:34 +0800)]
btrfs: remove unnecessary tmp variable in btrfs_assign_next_active_device()

We can check the argument value directly, no need for the temporary
variable.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: remove tmp variable for list traversal in btrfs_init_dev_replace_tgtdev
Anand Jain [Fri, 4 Sep 2020 17:34:32 +0000 (01:34 +0800)]
btrfs: remove tmp variable for list traversal in btrfs_init_dev_replace_tgtdev

In the function btrfs_init_dev_replace_tgtdev(), the local variable
devices is used only once, we can remove it.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: use sprout device_list_mutex in btrfs_init_devices_late
Anand Jain [Fri, 4 Sep 2020 17:34:31 +0000 (01:34 +0800)]
btrfs: use sprout device_list_mutex in btrfs_init_devices_late

On a mounted sprout filesystem, all threads now are using the
sprout::device_list_mutex, and this is the only code using the
seed::device_list_mutex. This patch converts to use the sprouts
fs_info->fs_devices->device_list_mutex.

The same reasoning holds true here, that device delete is holding
the sprout::device_list_mutex.

Signed-off-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: reada: lock all seed/sprout devices in __reada_start_machine
Anand Jain [Fri, 4 Sep 2020 17:34:30 +0000 (01:34 +0800)]
btrfs: reada: lock all seed/sprout devices in __reada_start_machine

On an fs mounted using a sprout device, the seed fs_devices are
maintained in a linked list under fs_info->fs_devices. Each seeds
fs_devices also has device_list_mutex initialized to protect against the
potential race with delete threads. But the delete thread (at
btrfs_rm_device()) is holding the fs_info::fs_devices::device_list_mutex
mutex which belongs to sprout device_list_mutex instead of seed
device_list_mutex. Moreover, there aren't any significient benefits in
using the seed::device_list_mutex instead of sprout::device_list_mutex.

So this patch converts them of using the seed::device_list_mutex to
sprout::device_list_mutex.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: handle errors in btrfs_sysfs_add_fs_devices
Anand Jain [Fri, 4 Sep 2020 17:34:29 +0000 (01:34 +0800)]
btrfs: handle errors in btrfs_sysfs_add_fs_devices

btrfs_sysfs_add_fs_devices() is called by btrfs_sysfs_add_mounted().
btrfs_sysfs_add_mounted() assumes that btrfs_sysfs_add_fs_devices() will
either add sysfs entries for all the devices or none. So this patch keeps up
to its caller expecatation and cleans up the created sysfs entries if it
has to fail at some device in the list.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: initialize sysfs devid and device link for seed device
Anand Jain [Fri, 4 Sep 2020 17:34:28 +0000 (01:34 +0800)]
btrfs: initialize sysfs devid and device link for seed device

We don't initialize the sysfs devid kobject and device-link yet for the
seed devices in an sprouted filesystem.
So this patch initializes the seed device devid kobject and the device
link in the sysfs.

Signed-off-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: split and refactor btrfs_sysfs_remove_devices_dir
Anand Jain [Fri, 4 Sep 2020 17:34:27 +0000 (01:34 +0800)]
btrfs: split and refactor btrfs_sysfs_remove_devices_dir

Similar to btrfs_sysfs_add_devices_dir()'s refactoring, split
btrfs_sysfs_remove_devices_dir() so that we don't have to use the device
argument to indicate whether to free all devices or just one device.

Export btrfs_sysfs_remove_device() as device operations outside of
sysfs.c now calls this instead of btrfs_sysfs_remove_devices_dir().

btrfs_sysfs_remove_devices_dir() is renamed to
btrfs_sysfs_remove_fs_devices() to suite its new role.

Now, no one outside of sysfs.c calls btrfs_sysfs_remove_fs_devices()
so it is redeclared s static. And the same function had to be moved
before its first caller.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: simplify parameters of btrfs_sysfs_add_devices_dir
Anand Jain [Fri, 4 Sep 2020 17:34:26 +0000 (01:34 +0800)]
btrfs: simplify parameters of btrfs_sysfs_add_devices_dir

When we add a device we need to add it to sysfs, so instead of using the
btrfs_sysfs_add_devices_dir() fs_devices argument to specify whether to
add a device or all of fs_devices, call the helper function directly
btrfs_sysfs_add_device() and thus make it non-static.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make btrfs_sysfs_remove_devices_dir return void
Anand Jain [Fri, 4 Sep 2020 17:34:25 +0000 (01:34 +0800)]
btrfs: make btrfs_sysfs_remove_devices_dir return void

btrfs_sysfs_remove_devices_dir() return value is unused declare it as
void.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: add btrfs_sysfs_remove_device helper
Anand Jain [Fri, 4 Sep 2020 17:34:24 +0000 (01:34 +0800)]
btrfs: add btrfs_sysfs_remove_device helper

btrfs_sysfs_remove_devices_dir() removes device link and devid kobject
(sysfs entries) for a device or all the devices in the btrfs_fs_devices.
In preparation to remove these sysfs entries for the seed as well, add
a btrfs_sysfs_remove_device() helper function and avoid code
duplication.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: add btrfs_sysfs_add_device helper
Anand Jain [Fri, 4 Sep 2020 17:34:23 +0000 (01:34 +0800)]
btrfs: add btrfs_sysfs_add_device helper

btrfs_sysfs_add_devices_dir() adds device link and devid kobject
(sysfs entries) for a device or all the devices in the btrfs_fs_devices.
In preparation to add these sysfs entries for the seed as well, add
a btrfs_sysfs_add_device() helper function and avoid code duplication.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: fix replace of seed device
Anand Jain [Fri, 4 Sep 2020 17:34:22 +0000 (01:34 +0800)]
btrfs: fix replace of seed device

If you replace a seed device in a sprouted fs, it appears to have
successfully replaced the seed device, but if you look closely, it
didn't.  Here is an example.

  $ mkfs.btrfs /dev/sda
  $ btrfstune -S1 /dev/sda
  $ mount /dev/sda /btrfs
  $ btrfs device add /dev/sdb /btrfs
  $ umount /btrfs
  $ btrfs device scan --forget
  $ mount -o device=/dev/sda /dev/sdb /btrfs
  $ btrfs replace start -f /dev/sda /dev/sdc /btrfs
  $ echo $?
  0

  BTRFS info (device sdb): dev_replace from /dev/sda (devid 1) to /dev/sdc started
  BTRFS info (device sdb): dev_replace from /dev/sda (devid 1) to /dev/sdc finished

  $ btrfs fi show
  Label: none  uuid: ab2c88b7-be81-4a7e-9849-c3666e7f9f4f
  Total devices 2 FS bytes used 256.00KiB
  devid    1 size 3.00GiB used 520.00MiB path /dev/sdc
  devid    2 size 3.00GiB used 896.00MiB path /dev/sdb

  Label: none  uuid: 10bd3202-0415-43af-96a8-d5409f310a7e
  Total devices 1 FS bytes used 128.00KiB
  devid    1 size 3.00GiB used 536.00MiB path /dev/sda

So as per the replace start command and kernel log replace was successful.
Now let's try to clean mount.

  $ umount /btrfs
  $ btrfs device scan --forget

  $ mount -o device=/dev/sdc /dev/sdb /btrfs
  mount: /btrfs: wrong fs type, bad option, bad superblock on /dev/sdb, missing codepage or helper program, or other error.

  [  636.157517] BTRFS error (device sdc): failed to read chunk tree: -2
  [  636.180177] BTRFS error (device sdc): open_ctree failed

That's because per dev items it is still looking for the original seed
device.

 $ btrfs inspect-internal dump-tree -d /dev/sdb

item 0 key (DEV_ITEMS DEV_ITEM 1) itemoff 16185 itemsize 98
devid 1 total_bytes 3221225472 bytes_used 545259520
io_align 4096 io_width 4096 sector_size 4096 type 0
generation 6 start_offset 0 dev_group 0
seek_speed 0 bandwidth 0
uuid 59368f50-9af2-4b17-91da-8a783cc418d4  <--- seed uuid
fsid 10bd3202-0415-43af-96a8-d5409f310a7e  <--- seed fsid
item 1 key (DEV_ITEMS DEV_ITEM 2) itemoff 16087 itemsize 98
devid 2 total_bytes 3221225472 bytes_used 939524096
io_align 4096 io_width 4096 sector_size 4096 type 0
generation 0 start_offset 0 dev_group 0
seek_speed 0 bandwidth 0
uuid 56a0a6bc-4630-4998-8daf-3c3030c4256a  <- sprout uuid
fsid ab2c88b7-be81-4a7e-9849-c3666e7f9f4f <- sprout fsid

But the replaced target has the following uuid+fsid in its superblock
which doesn't match with the expected uuid+fsid in its devitem.

  $ btrfs in dump-super /dev/sdc | egrep '^generation|dev_item.uuid|dev_item.fsid|devid'
  generation 20
  dev_item.uuid 59368f50-9af2-4b17-91da-8a783cc418d4
  dev_item.fsid ab2c88b7-be81-4a7e-9849-c3666e7f9f4f [match]
  dev_item.devid 1

So if you provide the original seed device the mount shall be
successful.  Which so long happening in the test case btrfs/163.

  $ btrfs device scan --forget
  $ mount -o device=/dev/sda /dev/sdb /btrfs

Fix in this patch:
If a seed is not sprouted then there is no replacement of it, because of
its read-only filesystem with a read-only device. Similarly, in the case
of a sprouted filesystem, the seed device is still read only. So, mark
it as you can't replace a seed device, you can only add a new device and
then delete the seed device. If replace is attempted then returns
-EINVAL.

Signed-off-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: improve device scanning messages
Anand Jain [Thu, 3 Sep 2020 13:30:12 +0000 (21:30 +0800)]
btrfs: improve device scanning messages

Systems booting without the initramfs seems to scan an unusual kind
of device path (/dev/root). And at a later time, the device is updated
to the correct path. We generally print the process name and PID of the
process scanning the device but we don't capture the same information if
the device path is rescanned with a different pathname.

The current message is too long, so drop the unnecessary UUID and add
process name and PID.

While at this also update the duplicate device warning to include the
process name and PID so the messages are consistent

CC: stable@vger.kernel.org # 4.19+
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=89721
Signed-off-by: Anand Jain <anand.jain@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: pretty print leaked root name
Josef Bacik [Thu, 3 Sep 2020 18:29:51 +0000 (14:29 -0400)]
btrfs: pretty print leaked root name

I'm a actual human being so am incapable of converting u64 to s64 in my
head, so add a helper to get the pretty name of a root objectid and use
that helper to spit out the name for any special roots for leaked roots,
so I don't have to scratch my head and figure out which root I messed up
the refs for.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: sysfs: export currently running exclusive operation
Goldwyn Rodrigues [Tue, 25 Aug 2020 15:02:33 +0000 (10:02 -0500)]
btrfs: sysfs: export currently running exclusive operation

/sys/fs/<fsid>/exclusive_operation contains the currently executing
exclusive operation. Add a sysfs_notify() when operation end, so
userspace can be notified of exclusive operation is finished.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: enumerate the type of exclusive operation in progress
Goldwyn Rodrigues [Tue, 25 Aug 2020 15:02:32 +0000 (10:02 -0500)]
btrfs: enumerate the type of exclusive operation in progress

Instead of using a flag bit for exclusive operation, use a variable to
store which exclusive operation is being performed.  Introduce an API
to start and finish an exclusive operation.

This would enable another way for tools to check which operation is
running on why starting an exclusive operation failed. The followup
patch adds a sysfs_notify() to alert userspace when the state changes, so
userspace can perform select() on it to get notified of the change.

This would enable us to enqueue a command which will wait for current
exclusive operation to complete before issuing the next exclusive
operation. This has been done synchronously as opposed to a background
process, or else error collection (if any) will become difficult.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ update comments ]
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: sysfs: init devices outside of the chunk_mutex
Josef Bacik [Tue, 1 Sep 2020 12:09:01 +0000 (08:09 -0400)]
btrfs: sysfs: init devices outside of the chunk_mutex

While running btrfs/061, btrfs/073, btrfs/078, or btrfs/178 we hit the
following lockdep splat:

  ======================================================
  WARNING: possible circular locking dependency detected
  5.9.0-rc3+ #4 Not tainted
  ------------------------------------------------------
  kswapd0/100 is trying to acquire lock:
  ffff96ecc22ef4a0 (&delayed_node->mutex){+.+.}-{3:3}, at: __btrfs_release_delayed_node.part.0+0x3f/0x330

  but task is already holding lock:
  ffffffff8dd74700 (fs_reclaim){+.+.}-{0:0}, at: __fs_reclaim_acquire+0x5/0x30

  which lock already depends on the new lock.

  the existing dependency chain (in reverse order) is:

  -> #3 (fs_reclaim){+.+.}-{0:0}:
 fs_reclaim_acquire+0x65/0x80
 slab_pre_alloc_hook.constprop.0+0x20/0x200
 kmem_cache_alloc+0x37/0x270
 alloc_inode+0x82/0xb0
 iget_locked+0x10d/0x2c0
 kernfs_get_inode+0x1b/0x130
 kernfs_get_tree+0x136/0x240
 sysfs_get_tree+0x16/0x40
 vfs_get_tree+0x28/0xc0
 path_mount+0x434/0xc00
 __x64_sys_mount+0xe3/0x120
 do_syscall_64+0x33/0x40
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

  -> #2 (kernfs_mutex){+.+.}-{3:3}:
 __mutex_lock+0x7e/0x7e0
 kernfs_add_one+0x23/0x150
 kernfs_create_link+0x63/0xa0
 sysfs_do_create_link_sd+0x5e/0xd0
 btrfs_sysfs_add_devices_dir+0x81/0x130
 btrfs_init_new_device+0x67f/0x1250
 btrfs_ioctl+0x1ef/0x2e20
 __x64_sys_ioctl+0x83/0xb0
 do_syscall_64+0x33/0x40
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

  -> #1 (&fs_info->chunk_mutex){+.+.}-{3:3}:
 __mutex_lock+0x7e/0x7e0
 btrfs_chunk_alloc+0x125/0x3a0
 find_free_extent+0xdf6/0x1210
 btrfs_reserve_extent+0xb3/0x1b0
 btrfs_alloc_tree_block+0xb0/0x310
 alloc_tree_block_no_bg_flush+0x4a/0x60
 __btrfs_cow_block+0x11a/0x530
 btrfs_cow_block+0x104/0x220
 btrfs_search_slot+0x52e/0x9d0
 btrfs_insert_empty_items+0x64/0xb0
 btrfs_insert_delayed_items+0x90/0x4f0
 btrfs_commit_inode_delayed_items+0x93/0x140
 btrfs_log_inode+0x5de/0x2020
 btrfs_log_inode_parent+0x429/0xc90
 btrfs_log_new_name+0x95/0x9b
 btrfs_rename2+0xbb9/0x1800
 vfs_rename+0x64f/0x9f0
 do_renameat2+0x320/0x4e0
 __x64_sys_rename+0x1f/0x30
 do_syscall_64+0x33/0x40
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

  -> #0 (&delayed_node->mutex){+.+.}-{3:3}:
 __lock_acquire+0x119c/0x1fc0
 lock_acquire+0xa7/0x3d0
 __mutex_lock+0x7e/0x7e0
 __btrfs_release_delayed_node.part.0+0x3f/0x330
 btrfs_evict_inode+0x24c/0x500
 evict+0xcf/0x1f0
 dispose_list+0x48/0x70
 prune_icache_sb+0x44/0x50
 super_cache_scan+0x161/0x1e0
 do_shrink_slab+0x178/0x3c0
 shrink_slab+0x17c/0x290
 shrink_node+0x2b2/0x6d0
 balance_pgdat+0x30a/0x670
 kswapd+0x213/0x4c0
 kthread+0x138/0x160
 ret_from_fork+0x1f/0x30

  other info that might help us debug this:

  Chain exists of:
    &delayed_node->mutex --> kernfs_mutex --> fs_reclaim

   Possible unsafe locking scenario:

 CPU0                    CPU1
 ----                    ----
    lock(fs_reclaim);
 lock(kernfs_mutex);
 lock(fs_reclaim);
    lock(&delayed_node->mutex);

   *** DEADLOCK ***

  3 locks held by kswapd0/100:
   #0: ffffffff8dd74700 (fs_reclaim){+.+.}-{0:0}, at: __fs_reclaim_acquire+0x5/0x30
   #1: ffffffff8dd65c50 (shrinker_rwsem){++++}-{3:3}, at: shrink_slab+0x115/0x290
   #2: ffff96ed2ade30e0 (&type->s_umount_key#36){++++}-{3:3}, at: super_cache_scan+0x38/0x1e0

  stack backtrace:
  CPU: 0 PID: 100 Comm: kswapd0 Not tainted 5.9.0-rc3+ #4
  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014
  Call Trace:
   dump_stack+0x8b/0xb8
   check_noncircular+0x12d/0x150
   __lock_acquire+0x119c/0x1fc0
   lock_acquire+0xa7/0x3d0
   ? __btrfs_release_delayed_node.part.0+0x3f/0x330
   __mutex_lock+0x7e/0x7e0
   ? __btrfs_release_delayed_node.part.0+0x3f/0x330
   ? __btrfs_release_delayed_node.part.0+0x3f/0x330
   ? lock_acquire+0xa7/0x3d0
   ? find_held_lock+0x2b/0x80
   __btrfs_release_delayed_node.part.0+0x3f/0x330
   btrfs_evict_inode+0x24c/0x500
   evict+0xcf/0x1f0
   dispose_list+0x48/0x70
   prune_icache_sb+0x44/0x50
   super_cache_scan+0x161/0x1e0
   do_shrink_slab+0x178/0x3c0
   shrink_slab+0x17c/0x290
   shrink_node+0x2b2/0x6d0
   balance_pgdat+0x30a/0x670
   kswapd+0x213/0x4c0
   ? _raw_spin_unlock_irqrestore+0x41/0x50
   ? add_wait_queue_exclusive+0x70/0x70
   ? balance_pgdat+0x670/0x670
   kthread+0x138/0x160
   ? kthread_create_worker_on_cpu+0x40/0x40
   ret_from_fork+0x1f/0x30

This happens because we are holding the chunk_mutex at the time of
adding in a new device.  However we only need to hold the
device_list_mutex, as we're going to iterate over the fs_devices
devices.  Move the sysfs init stuff outside of the chunk_mutex to get
rid of this lockdep splat.

CC: stable@vger.kernel.org # 4.4.x: f3cd2c58110dad14e: btrfs: sysfs, rename device_link add/remove functions
CC: stable@vger.kernel.org # 4.4.x
Reported-by: David Sterba <dsterba@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make extent_fiemap take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:49 +0000 (14:42 +0300)]
btrfs: make extent_fiemap take btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make btrfs_zero_range_check_range_boundary take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:48 +0000 (14:42 +0300)]
btrfs: make btrfs_zero_range_check_range_boundary take btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make copy_inline_to_page take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:47 +0000 (14:42 +0300)]
btrfs: make copy_inline_to_page take btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make btrfs_find_ordered_sum take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:46 +0000 (14:42 +0300)]
btrfs: make btrfs_find_ordered_sum take btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make get_extent_skip_holes take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:45 +0000 (14:42 +0300)]
btrfs: make get_extent_skip_holes take btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make btrfs_writepage_endio_finish_ordered btrfs_inode-centric
Nikolay Borisov [Mon, 31 Aug 2020 11:42:44 +0000 (14:42 +0300)]
btrfs: make btrfs_writepage_endio_finish_ordered btrfs_inode-centric

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make btrfs_invalidatepage work on btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:43 +0000 (14:42 +0300)]
btrfs: make btrfs_invalidatepage work on btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: convert btrfs_inode_sectorsize to take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:42 +0000 (14:42 +0300)]
btrfs: convert btrfs_inode_sectorsize to take btrfs_inode

It's counterintuitive to have a function named btrfs_inode_xxx which
takes a generic inode. Also move the function to btrfs_inode.h so that
it has access to the definition of struct btrfs_inode.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make btrfs_dec_test_ordered_pending take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:41 +0000 (14:42 +0300)]
btrfs: make btrfs_dec_test_ordered_pending take btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make ordered extent tracepoint take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:40 +0000 (14:42 +0300)]
btrfs: make ordered extent tracepoint take btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make btrfs_lookup_first_ordered_extent take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:39 +0000 (14:42 +0300)]
btrfs: make btrfs_lookup_first_ordered_extent take btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make inode_tree_del take btrfs_inode
Nikolay Borisov [Mon, 31 Aug 2020 11:42:38 +0000 (14:42 +0300)]
btrfs: make inode_tree_del take btrfs_inode

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: use BTRFS_NESTED_NEW_ROOT for double splits
Josef Bacik [Thu, 20 Aug 2020 15:46:08 +0000 (11:46 -0400)]
btrfs: use BTRFS_NESTED_NEW_ROOT for double splits

I've made this change separate since it requires both of the newly added
NESTED flags and I didn't want to slip it into one of those changes.

If we do a double split of a node we can end up doing a
BTRFS_NESTED_SPLIT on level 0, which throws lockdep off because it
appears as a double lock.  Since we're maxed out on subclasses, use
BTRFS_NESTED_NEW_ROOT if we had to do a double split.  This is OK
because we won't have to do a double split if we had to insert a new
root, and the new root would be at a higher level anyway.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: introduce BTRFS_NESTING_NEW_ROOT for adding new roots
Josef Bacik [Thu, 20 Aug 2020 15:46:07 +0000 (11:46 -0400)]
btrfs: introduce BTRFS_NESTING_NEW_ROOT for adding new roots

The way we add new roots is confusing from a locking perspective for
lockdep.  We generally have the rule that we lock things in order from
highest level to lowest, but in the case of adding a new level to the
tree we actually allocate a new block for the root, which makes the
locking go in reverse.  A similar issue exists for snapshotting, we cow
the original root for the root of a new tree, however they're at the
same level.  Address this by using BTRFS_NESTING_NEW_ROOT for these
operations.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: introduce BTRFS_NESTING_SPLIT for split blocks
Josef Bacik [Thu, 20 Aug 2020 15:46:06 +0000 (11:46 -0400)]
btrfs: introduce BTRFS_NESTING_SPLIT for split blocks

If we are splitting a leaf/node, we could do something like the
following

lock(leaf)  BTRFS_NESTING_NORMAL
  lock(left) BTRFS_NESTING_LEFT + BTRFS_NESTING_COW
    push from leaf -> left
      reset path to point to left
        split left
          allocate new block, lock block BTRFS_NESTING_SPLIT

at the new block point we need to have a different nesting level,
because we have already used either BTRFS_NESTING_LEFT or
BTRFS_NESTING_RIGHT when pushing items from the original leaf into the
adjacent leaves.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: introduce BTRFS_NESTING_LEFT/RIGHT_COW
Josef Bacik [Thu, 20 Aug 2020 15:46:05 +0000 (11:46 -0400)]
btrfs: introduce BTRFS_NESTING_LEFT/RIGHT_COW

For similar reasons as BTRFS_NESTING_COW, we need
BTRFS_NESTING_LEFT/RIGHT_COW.  The pattern is this

lock leaf -> BTRFS_NESTING_NORMAL
  cow leaf -> BTRFS_NESTING_COW
    split leaf
      lock left -> BTRFS_NESTING_LEFT
        cow left -> BTRFS_NESTING_LEFT_COW

We need this in order to indicate to lockdep that these locks are
discrete and are being taken in a safe order.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: introduce BTRFS_NESTING_LEFT/BTRFS_NESTING_RIGHT
Josef Bacik [Thu, 20 Aug 2020 15:46:04 +0000 (11:46 -0400)]
btrfs: introduce BTRFS_NESTING_LEFT/BTRFS_NESTING_RIGHT

Our lockdep maps are based on rootid+level, however in some cases we
will lock adjacent blocks on the same level, namely in searching forward
or in split/balance.  Because of this lockdep will complain, so we need
a separate subclass to indicate to lockdep that these are different
locks.

lock leaf -> BTRFS_NESTING_NORMAL
  cow leaf -> BTRFS_NESTING_COW
    split leaf
       lock left -> BTRFS_NESTING_LEFT
       lock right -> BTRFS_NESTING_RIGHT

The above graph illustrates the need for this new nesting subclass.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: introduce BTRFS_NESTING_COW for cow'ing blocks
Josef Bacik [Thu, 20 Aug 2020 15:46:03 +0000 (11:46 -0400)]
btrfs: introduce BTRFS_NESTING_COW for cow'ing blocks

When we COW a block we are holding a lock on the original block, and
then we lock the new COW block.  Because our lockdep maps are based on
root + level, this will make lockdep complain.  We need a way to
indicate a subclass for locking the COW'ed block, so plumb through our
btrfs_lock_nesting from btrfs_cow_block down to the btrfs_init_buffer,
and then introduce BTRFS_NESTING_COW to be used for cow'ing blocks.

The reason I've added all this extra infrastructure is because there
will be need of different nesting classes in follow up patches.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: add nesting tags to the locking helpers
Josef Bacik [Thu, 20 Aug 2020 15:46:02 +0000 (11:46 -0400)]
btrfs: add nesting tags to the locking helpers

We will need these when we switch to an rwsem, so plumb in the
infrastructure here to use later on.  I violate the 80 character limit
some here because it'll be cleaned up later.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: introduce btrfs_path::recurse
Josef Bacik [Thu, 20 Aug 2020 15:46:01 +0000 (11:46 -0400)]
btrfs: introduce btrfs_path::recurse

Our current tree locking stuff allows us to recurse with read locks if
we're already holding the write lock.  This is necessary for the space
cache inode, as we could be holding a lock on the root_tree root when we
need to cache a block group, and thus need to be able to read down the
root_tree to read in the inode cache.

We can get away with this in our current locking, but we won't be able
to with a rwsem.  Handle this by purposefully annotating the places
where we require recursion, so that in the future we can maybe come up
with a way to avoid the recursion.  In the case of the free space inode,
this will be superseded by the free space tree.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: rename extent_buffer::lock_nested to extent_buffer::lock_recursed
Josef Bacik [Thu, 20 Aug 2020 15:46:00 +0000 (11:46 -0400)]
btrfs: rename extent_buffer::lock_nested to extent_buffer::lock_recursed

Nested locking with lockdep and everything else refers to lock hierarchy
within the same lock map.  This is how we indicate the same locks for
different objects are ok to take in a specific order, for our use case
that would be to take the lock on a leaf and then take a lock on an
adjacent leaf.

What ->lock_nested _actually_ refers to is if we happen to already be
holding the write lock on the extent buffer and we're allowing a read
lock to be taken on that extent buffer, which is recursion.  Rename this
so we don't get confused when we switch to a rwsem and have to start
using the _nested helpers.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: don't opencode sync_blockdev in btrfs_init_new_device
Nikolay Borisov [Wed, 22 Jul 2020 08:09:25 +0000 (11:09 +0300)]
btrfs: don't opencode sync_blockdev in btrfs_init_new_device

Instead of opencoding filemap_write_and_wait simply call syncblockdev as
it makes it abundantly clear what's going on and why this is used. No
semantics changes.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: remove redundant code from btrfs_free_stale_devices
Nikolay Borisov [Wed, 22 Jul 2020 08:09:24 +0000 (11:09 +0300)]
btrfs: remove redundant code from btrfs_free_stale_devices

Following the refactor of btrfs_free_stale_devices in
7bcb8164ad94 ("btrfs: use device_list_mutex when removing stale devices")
fs_devices are freed after they have been iterated by the inner
list_for_each so the use-after-free fixed by introducing the break in
fd649f10c3d2 ("btrfs: Fix use-after-free when cleaning up fs_devs with
a single stale device") is no longer necessary. Just remove it
altogether. No functional changes.

Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: refactor locked condition in btrfs_init_new_device
Nikolay Borisov [Wed, 22 Jul 2020 08:09:23 +0000 (11:09 +0300)]
btrfs: refactor locked condition in btrfs_init_new_device

Invert unlocked to locked and exploit the fact it can only ever be
modified if we are adding a new device to a seed filesystem. This allows
to simplify the check in error: label. No semantics changes.

Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: use RCU for quick device check in btrfs_init_new_device
Nikolay Borisov [Wed, 22 Jul 2020 08:09:22 +0000 (11:09 +0300)]
btrfs: use RCU for quick device check in btrfs_init_new_device

When adding a new device there's a mandatory check to see if a device is
being duplicated to the filesystem it's added to. Since this is a
read-only operations not necessary to take device_list_mutex and can simply
make do with an rcu-readlock.

Using just RCU is safe because there won't be another device add delete
running in parallel as btrfs_init_new_device is called only from
btrfs_ioctl_add_dev.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: ctree: check key order before merging tree blocks
Qu Wenruo [Wed, 19 Aug 2020 06:35:50 +0000 (14:35 +0800)]
btrfs: ctree: check key order before merging tree blocks

[BUG]
With a crafted image, btrfs can panic at btrfs_del_csums():

  kernel BUG at fs/btrfs/ctree.c:3188!
  invalid opcode: 0000 [#1] SMP PTI
  CPU: 0 PID: 1156 Comm: btrfs-transacti Not tainted 5.0.0-rc8+ #9
  RIP: 0010:btrfs_set_item_key_safe+0x16c/0x180
  RSP: 0018:ffff976141257ab8 EFLAGS: 00010202
  RAX: 0000000000000001 RBX: ffff898a6b890930 RCX: 0000000004b70000
  RDX: 0000000000000000 RSI: ffff976141257bae RDI: ffff976141257acf
  RBP: ffff976141257b10 R08: 0000000000001000 R09: ffff9761412579a8
  R10: 0000000000000000 R11: 0000000000000000 R12: ffff976141257abe
  R13: 0000000000000003 R14: ffff898a6a8be578 R15: ffff976141257bae
  FS: 0000000000000000(0000) GS:ffff898a77a00000(0000) knlGS:0000000000000000
  CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00007f779d9cd624 CR3: 000000022b2b4006 CR4: 00000000000206f0
  Call Trace:
  truncate_one_csum+0xac/0xf0
  btrfs_del_csums+0x24f/0x3a0
  __btrfs_free_extent.isra.72+0x5a7/0xbe0
  __btrfs_run_delayed_refs+0x539/0x1120
  btrfs_run_delayed_refs+0xdb/0x1b0
  btrfs_commit_transaction+0x52/0x950
  ? start_transaction+0x94/0x450
  transaction_kthread+0x163/0x190
  kthread+0x105/0x140
  ? btrfs_cleanup_transaction+0x560/0x560
  ? kthread_destroy_worker+0x50/0x50
  ret_from_fork+0x35/0x40
  Modules linked in:
  ---[ end trace 93bf9db00e6c374e ]---

[CAUSE]
This crafted image has a tricky key order corruption:

  checksum tree key (CSUM_TREE ROOT_ITEM 0)
  node 29741056 level 1 items 14 free 107 generation 19 owner CSUM_TREE
          ...
          key (EXTENT_CSUM EXTENT_CSUM 73785344) block 29757440 gen 19
          key (EXTENT_CSUM EXTENT_CSUM 77594624) block 29753344 gen 19
          ...

  leaf 29757440 items 5 free space 150 generation 19 owner CSUM_TREE
          item 0 key (EXTENT_CSUM EXTENT_CSUM 73785344) itemoff 2323 itemsize 1672
                  range start 73785344 end 75497472 length 1712128
          item 1 key (EXTENT_CSUM EXTENT_CSUM 75497472) itemoff 2319 itemsize 4
                  range start 75497472 end 75501568 length 4096
          item 2 key (EXTENT_CSUM EXTENT_CSUM 75501568) itemoff 579 itemsize 1740
                  range start 75501568 end 77283328 length 1781760
          item 3 key (EXTENT_CSUM EXTENT_CSUM 77283328) itemoff 575 itemsize 4
                  range start 77283328 end 77287424 length 4096
          item 4 key (EXTENT_CSUM EXTENT_CSUM 4120596480) itemoff 275 itemsize 300 <<<
                  range start 4120596480 end 4120903680 length 307200
  leaf 29753344 items 3 free space 1936 generation 19 owner CSUM_TREE
          item 0 key (18446744073457893366 EXTENT_CSUM 77594624) itemoff 2323 itemsize 1672
                  range start 77594624 end 79306752 length 1712128
          ...

Note the item 4 key of leaf 29757440, which is obviously too large, and
even larger than the first key of the next leaf.

However it still follows the key order in that tree block, thus tree
checker is unable to detect it at read time, since tree checker can only
work inside one leaf, thus such complex corruption can't be detected in
advance.

[FIX]
The next time to detect such problem is at tree block merge time,
which is in push_node_left(), balance_node_right(), push_leaf_left() or
push_leaf_right().

Now we check if the key order of the right-most key of the left node is
larger than the left-most key of the right node.

By this we don't need to call the full tree-checker, while still keeping
the key order correct as key order in each node is already checked by
tree checker thus we only need to check the above two slots.

Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202833
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: extent-tree: kill the BUG_ON() in insert_inline_extent_backref()
Qu Wenruo [Wed, 19 Aug 2020 06:35:49 +0000 (14:35 +0800)]
btrfs: extent-tree: kill the BUG_ON() in insert_inline_extent_backref()

[BUG]
With a crafted image, btrfs can panic at insert_inline_extent_backref():

  kernel BUG at fs/btrfs/extent-tree.c:1857!
  invalid opcode: 0000 [#1] SMP PTI
  CPU: 0 PID: 1117 Comm: btrfs-transacti Not tainted 5.0.0-rc8+ #9
  RIP: 0010:insert_inline_extent_backref+0xcc/0xe0
  RSP: 0018:ffffac4dc1287be8 EFLAGS: 00010293
  RAX: 0000000000000000 RBX: 0000000000000007 RCX: 0000000000000001
  RDX: 0000000000001000 RSI: 0000000000000000 RDI: 0000000000000000
  RBP: ffffac4dc1287c28 R08: ffffac4dc1287ab8 R09: ffffac4dc1287ac0
  R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000000
  R13: ffff8febef88a540 R14: ffff8febeaa7bc30 R15: 0000000000000000
  FS: 0000000000000000(0000) GS:ffff8febf7a00000(0000) knlGS:0000000000000000
  CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00007f663ace94c0 CR3: 0000000235698006 CR4: 00000000000206f0
  Call Trace:
  ? _cond_resched+0x1a/0x50
  __btrfs_inc_extent_ref.isra.64+0x7e/0x240
  ? btrfs_merge_delayed_refs+0xa5/0x330
  __btrfs_run_delayed_refs+0x653/0x1120
  btrfs_run_delayed_refs+0xdb/0x1b0
  btrfs_commit_transaction+0x52/0x950
  ? start_transaction+0x94/0x450
  transaction_kthread+0x163/0x190
  kthread+0x105/0x140
  ? btrfs_cleanup_transaction+0x560/0x560
  ? kthread_destroy_worker+0x50/0x50
  ret_from_fork+0x35/0x40
  Modules linked in:
  ---[ end trace 2ad8b3de903cf825 ]---

[CAUSE]
Due to extent tree corruption (still valid by itself, but bad cross
ref), we can allocate an extent which is still in extent tree.  The
offending tree block of that case is from csum tree.  The newly
allocated tree block is also for csum tree.

Then we will try to insert a tree block ref for the existing tree block
ref.

For a tree extent item, tree block can never be shared directly by the
same tree twice.  We have such BUG_ON() to prevent such problem, but
this is not a proper error handling.

[FIX]
Replace that BUG_ON() with proper error message and leaf dump for debug
build.

Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202829
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: extent-tree: kill BUG_ON() in __btrfs_free_extent()
Qu Wenruo [Wed, 19 Aug 2020 06:35:48 +0000 (14:35 +0800)]
btrfs: extent-tree: kill BUG_ON() in __btrfs_free_extent()

__btrfs_free_extent() is doing two things:

1. Reduce the refs number of an extent backref
   Either it's an inline extent backref (inside EXTENT/METADATA item) or
   a keyed extent backref (SHARED_* item).
   We only need to locate that backref line, either reduce the number or
   remove the backref line completely.

2. Update the refs count in EXTENT/METADATA_ITEM

During step 1), we will try to locate the EXTENT/METADATA_ITEM without
triggering another btrfs_search_slot() as fast path.

Only when we fail to locate that item, we will trigger another
btrfs_search_slot() to get that EXTENT/METADATA_ITEM after we
updated/deleted the backref line.

And we have a lot of strict checks on things like refs_to_drop against
extent refs and special case checks for single ref extents.

There are 7 BUG_ON()s, although they're doing correct checks, they can
be triggered by crafted images.

This patch improves the function:

- Introduce two examples to show what __btrfs_free_extent() is doing
  One inline backref case and one keyed case.  Should cover most cases.

- Kill all BUG_ON()s with proper error message and optional leaf dump

- Add comment to show the overall flow

Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202819
[ The report triggers one BUG_ON() in __btrfs_free_extent() ]
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: extent_io: do extra check for extent buffer read write functions
Qu Wenruo [Wed, 19 Aug 2020 06:35:47 +0000 (14:35 +0800)]
btrfs: extent_io: do extra check for extent buffer read write functions

Although we have start, len check for extent buffer reader/write (e.g.
read_extent_buffer()), these checks have limitations:

- No overflow check
  Values like start = 1024 len = -1024 can still pass the basic
   (start + len) > eb->len check.

- Checks are not consistent
  For read_extent_buffer() we only check (start + len) against eb->len.
  While for memcmp_extent_buffer() we also check start against eb->len.

- Different error reporting mechanism
  We use WARN() in read_extent_buffer() but BUG() in
  memcpy_extent_buffer().

- Still modify memory if the request is obviously wrong
  In read_extent_buffer() even we find (start + len) > eb->len, we still
  call memset(dst, 0, len), which can easily cause memory access error
  if start + len overflows.

To address above problems, this patch creates a new common function to
check such access, check_eb_range().

- Add overflow check
  This function checks start, start + len against eb->len and overflow
  check.

- Unified checks

- Unified error reports
  Will call WARN() if CONFIG_BTRFS_DEBUG is configured.
  And also do btrfs_warn() message for non-debug build.

- Exit ASAP if check fails
  No more possible memory corruption.

- Add extra comment for @start @len used in those functions as it's
  sometimes confused with the logical addressing instead of a range
  inside the eb space

Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=202817
[ Inspired by above report, the report itself is already addressed ]
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ use check_add_overflow ]
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: rework error detection in init_tree_roots
Nikolay Borisov [Wed, 12 Aug 2020 13:16:35 +0000 (16:16 +0300)]
btrfs: rework error detection in init_tree_roots

To avoid duplicating 3 lines of code the error detection logic in
init_tree_roots is somewhat quirky. It first checks for the presence of
any error condition, then checks for the specific condition to perform
any specific actions. That's spurious because directly checking for
each respective error condition and doing the necessary steps is more
obvious. While at it change the -EUCLEAN to -EIO in case the extent
buffer is not read correctly, this is in line with other sites which
return -EIO when the eb couldn't be read.

Additionally it results in smaller code and the code reads
more linearly:

add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-95 (-95)
Function                                     old     new   delta
open_ctree                                 17243   17148     -95
Total: Before=113104, After=113009, chg -0.08%

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: qgroup: fix qgroup meta rsv leak for subvolume operations
Qu Wenruo [Fri, 24 Jul 2020 06:46:10 +0000 (14:46 +0800)]
btrfs: qgroup: fix qgroup meta rsv leak for subvolume operations

[BUG]
When quota is enabled for TEST_DEV, generic/013 sometimes fails like this:

  generic/013 14s ... _check_dmesg: something found in dmesg (see xfstests-dev/results//generic/013.dmesg)

And with the following metadata leak:

  BTRFS warning (device dm-3): qgroup 0/1370 has unreleased space, type 2 rsv 49152
  ------------[ cut here ]------------
  WARNING: CPU: 2 PID: 47912 at fs/btrfs/disk-io.c:4078 close_ctree+0x1dc/0x323 [btrfs]
  Call Trace:
   btrfs_put_super+0x15/0x17 [btrfs]
   generic_shutdown_super+0x72/0x110
   kill_anon_super+0x18/0x30
   btrfs_kill_super+0x17/0x30 [btrfs]
   deactivate_locked_super+0x3b/0xa0
   deactivate_super+0x40/0x50
   cleanup_mnt+0x135/0x190
   __cleanup_mnt+0x12/0x20
   task_work_run+0x64/0xb0
   __prepare_exit_to_usermode+0x1bc/0x1c0
   __syscall_return_slowpath+0x47/0x230
   do_syscall_64+0x64/0xb0
   entry_SYSCALL_64_after_hwframe+0x44/0xa9
  ---[ end trace a6cfd45ba80e4e06 ]---
  BTRFS error (device dm-3): qgroup reserved space leaked
  BTRFS info (device dm-3): disk space caching is enabled
  BTRFS info (device dm-3): has skinny extents

[CAUSE]
The qgroup preallocated meta rsv operations of that offending root are:

  btrfs_delayed_inode_reserve_metadata: rsv_meta_prealloc root=1370 num_bytes=131072
  btrfs_delayed_inode_reserve_metadata: rsv_meta_prealloc root=1370 num_bytes=131072
  btrfs_subvolume_reserve_metadata: rsv_meta_prealloc root=1370 num_bytes=49152
  btrfs_delayed_inode_release_metadata: convert_meta_prealloc root=1370 num_bytes=-131072
  btrfs_delayed_inode_release_metadata: convert_meta_prealloc root=1370 num_bytes=-131072

It's pretty obvious that, we reserve qgroup meta rsv in
btrfs_subvolume_reserve_metadata(), but doesn't have corresponding
release/convert calls in btrfs_subvolume_release_metadata().

This leads to the leakage.

[FIX]
To fix this bug, we should follow what we're doing in
btrfs_delalloc_reserve_metadata(), where we reserve qgroup space, and
add it to block_rsv->qgroup_rsv_reserved.

And free the qgroup reserved metadata space when releasing the
block_rsv.

To do this, we need to change the btrfs_subvolume_release_metadata() to
accept btrfs_root, and record the qgroup_to_release number, and call
btrfs_qgroup_convert_reserved_meta() for it.

Fixes: 733e03a0b26a ("btrfs: qgroup: Split meta rsv type into meta_prealloc and meta_pertrans")
CC: stable@vger.kernel.org # 4.19+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: qgroup: fix wrong qgroup metadata reserve for delayed inode
Qu Wenruo [Fri, 24 Jul 2020 06:46:09 +0000 (14:46 +0800)]
btrfs: qgroup: fix wrong qgroup metadata reserve for delayed inode

For delayed inode facility, qgroup metadata is reserved for it, and
later freed.

However we're freeing more bytes than we reserved.
In btrfs_delayed_inode_reserve_metadata():

num_bytes = btrfs_calc_metadata_size(fs_info, 1);
...
ret = btrfs_qgroup_reserve_meta_prealloc(root,
fs_info->nodesize, true);
...
if (!ret) {
node->bytes_reserved = num_bytes;

But in btrfs_delayed_inode_release_metadata():

if (qgroup_free)
btrfs_qgroup_free_meta_prealloc(node->root,
node->bytes_reserved);
else
btrfs_qgroup_convert_reserved_meta(node->root,
node->bytes_reserved);

This means, we're always releasing more qgroup metadata rsv than we have
reserved.

This won't trigger selftest warning, as btrfs qgroup metadata rsv has
extra protection against cases like quota enabled half-way.

But we still need to fix this problem any way.

This patch will use the same num_bytes for qgroup metadata rsv so we
could handle it correctly.

Fixes: f218ea6c4792 ("btrfs: delayed-inode: Remove wrong qgroup meta reservation calls")
CC: stable@vger.kernel.org # 4.19+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: do not hold device_list_mutex when closing devices
Josef Bacik [Mon, 10 Aug 2020 15:42:28 +0000 (11:42 -0400)]
btrfs: do not hold device_list_mutex when closing devices

The following lockdep splat

======================================================
WARNING: possible circular locking dependency detected
5.8.0-rc7-00169-g87212851a027-dirty #929 Not tainted
------------------------------------------------------
fsstress/8739 is trying to acquire lock:
ffff88bfd0eb0c90 (&fs_info->reloc_mutex){+.+.}-{3:3}, at: btrfs_record_root_in_trans+0x43/0x70

but task is already holding lock:
ffff88bfbd16e538 (sb_pagefaults){.+.+}-{0:0}, at: btrfs_page_mkwrite+0x6a/0x4a0

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #10 (sb_pagefaults){.+.+}-{0:0}:
       __sb_start_write+0x129/0x210
       btrfs_page_mkwrite+0x6a/0x4a0
       do_page_mkwrite+0x4d/0xc0
       handle_mm_fault+0x103c/0x1730
       exc_page_fault+0x340/0x660
       asm_exc_page_fault+0x1e/0x30

-> #9 (&mm->mmap_lock#2){++++}-{3:3}:
       __might_fault+0x68/0x90
       _copy_to_user+0x1e/0x80
       perf_read+0x141/0x2c0
       vfs_read+0xad/0x1b0
       ksys_read+0x5f/0xe0
       do_syscall_64+0x50/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xa9

-> #8 (&cpuctx_mutex){+.+.}-{3:3}:
       __mutex_lock+0x9f/0x930
       perf_event_init_cpu+0x88/0x150
       perf_event_init+0x1db/0x20b
       start_kernel+0x3ae/0x53c
       secondary_startup_64+0xa4/0xb0

-> #7 (pmus_lock){+.+.}-{3:3}:
       __mutex_lock+0x9f/0x930
       perf_event_init_cpu+0x4f/0x150
       cpuhp_invoke_callback+0xb1/0x900
       _cpu_up.constprop.26+0x9f/0x130
       cpu_up+0x7b/0xc0
       bringup_nonboot_cpus+0x4f/0x60
       smp_init+0x26/0x71
       kernel_init_freeable+0x110/0x258
       kernel_init+0xa/0x103
       ret_from_fork+0x1f/0x30

-> #6 (cpu_hotplug_lock){++++}-{0:0}:
       cpus_read_lock+0x39/0xb0
       kmem_cache_create_usercopy+0x28/0x230
       kmem_cache_create+0x12/0x20
       bioset_init+0x15e/0x2b0
       init_bio+0xa3/0xaa
       do_one_initcall+0x5a/0x2e0
       kernel_init_freeable+0x1f4/0x258
       kernel_init+0xa/0x103
       ret_from_fork+0x1f/0x30

-> #5 (bio_slab_lock){+.+.}-{3:3}:
       __mutex_lock+0x9f/0x930
       bioset_init+0xbc/0x2b0
       __blk_alloc_queue+0x6f/0x2d0
       blk_mq_init_queue_data+0x1b/0x70
       loop_add+0x110/0x290 [loop]
       fq_codel_tcf_block+0x12/0x20 [sch_fq_codel]
       do_one_initcall+0x5a/0x2e0
       do_init_module+0x5a/0x220
       load_module+0x2459/0x26e0
       __do_sys_finit_module+0xba/0xe0
       do_syscall_64+0x50/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xa9

-> #4 (loop_ctl_mutex){+.+.}-{3:3}:
       __mutex_lock+0x9f/0x930
       lo_open+0x18/0x50 [loop]
       __blkdev_get+0xec/0x570
       blkdev_get+0xe8/0x150
       do_dentry_open+0x167/0x410
       path_openat+0x7c9/0xa80
       do_filp_open+0x93/0x100
       do_sys_openat2+0x22a/0x2e0
       do_sys_open+0x4b/0x80
       do_syscall_64+0x50/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xa9

-> #3 (&bdev->bd_mutex){+.+.}-{3:3}:
       __mutex_lock+0x9f/0x930
       blkdev_put+0x1d/0x120
       close_fs_devices.part.31+0x84/0x130
       btrfs_close_devices+0x44/0xb0
       close_ctree+0x296/0x2b2
       generic_shutdown_super+0x69/0x100
       kill_anon_super+0xe/0x30
       btrfs_kill_super+0x12/0x20
       deactivate_locked_super+0x29/0x60
       cleanup_mnt+0xb8/0x140
       task_work_run+0x6d/0xb0
       __prepare_exit_to_usermode+0x1cc/0x1e0
       do_syscall_64+0x5c/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xa9

-> #2 (&fs_devs->device_list_mutex){+.+.}-{3:3}:
       __mutex_lock+0x9f/0x930
       btrfs_run_dev_stats+0x49/0x480
       commit_cowonly_roots+0xb5/0x2a0
       btrfs_commit_transaction+0x516/0xa60
       sync_filesystem+0x6b/0x90
       generic_shutdown_super+0x22/0x100
       kill_anon_super+0xe/0x30
       btrfs_kill_super+0x12/0x20
       deactivate_locked_super+0x29/0x60
       cleanup_mnt+0xb8/0x140
       task_work_run+0x6d/0xb0
       __prepare_exit_to_usermode+0x1cc/0x1e0
       do_syscall_64+0x5c/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xa9

-> #1 (&fs_info->tree_log_mutex){+.+.}-{3:3}:
       __mutex_lock+0x9f/0x930
       btrfs_commit_transaction+0x4bb/0xa60
       sync_filesystem+0x6b/0x90
       generic_shutdown_super+0x22/0x100
       kill_anon_super+0xe/0x30
       btrfs_kill_super+0x12/0x20
       deactivate_locked_super+0x29/0x60
       cleanup_mnt+0xb8/0x140
       task_work_run+0x6d/0xb0
       __prepare_exit_to_usermode+0x1cc/0x1e0
       do_syscall_64+0x5c/0x90
       entry_SYSCALL_64_after_hwframe+0x44/0xa9

-> #0 (&fs_info->reloc_mutex){+.+.}-{3:3}:
       __lock_acquire+0x1272/0x2310
       lock_acquire+0x9e/0x360
       __mutex_lock+0x9f/0x930
       btrfs_record_root_in_trans+0x43/0x70
       start_transaction+0xd1/0x5d0
       btrfs_dirty_inode+0x42/0xd0
       file_update_time+0xc8/0x110
       btrfs_page_mkwrite+0x10c/0x4a0
       do_page_mkwrite+0x4d/0xc0
       handle_mm_fault+0x103c/0x1730
       exc_page_fault+0x340/0x660
       asm_exc_page_fault+0x1e/0x30

other info that might help us debug this:

Chain exists of:
  &fs_info->reloc_mutex --> &mm->mmap_lock#2 --> sb_pagefaults

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(sb_pagefaults);
                               lock(&mm->mmap_lock#2);
                               lock(sb_pagefaults);
  lock(&fs_info->reloc_mutex);

 *** DEADLOCK ***

3 locks held by fsstress/8739:
 #0: ffff88bee66eeb68 (&mm->mmap_lock#2){++++}-{3:3}, at: exc_page_fault+0x173/0x660
 #1: ffff88bfbd16e538 (sb_pagefaults){.+.+}-{0:0}, at: btrfs_page_mkwrite+0x6a/0x4a0
 #2: ffff88bfbd16e630 (sb_internal){.+.+}-{0:0}, at: start_transaction+0x3da/0x5d0

stack backtrace:
CPU: 17 PID: 8739 Comm: fsstress Kdump: loaded Not tainted 5.8.0-rc7-00169-g87212851a027-dirty #929
Hardware name: Quanta Tioga Pass Single Side 01-0030993006/Tioga Pass Single Side, BIOS F08_3A18 12/20/2018
Call Trace:
 dump_stack+0x78/0xa0
 check_noncircular+0x165/0x180
 __lock_acquire+0x1272/0x2310
 ? btrfs_get_alloc_profile+0x150/0x210
 lock_acquire+0x9e/0x360
 ? btrfs_record_root_in_trans+0x43/0x70
 __mutex_lock+0x9f/0x930
 ? btrfs_record_root_in_trans+0x43/0x70
 ? lock_acquire+0x9e/0x360
 ? join_transaction+0x5d/0x450
 ? find_held_lock+0x2d/0x90
 ? btrfs_record_root_in_trans+0x43/0x70
 ? join_transaction+0x3d5/0x450
 ? btrfs_record_root_in_trans+0x43/0x70
 btrfs_record_root_in_trans+0x43/0x70
 start_transaction+0xd1/0x5d0
 btrfs_dirty_inode+0x42/0xd0
 file_update_time+0xc8/0x110
 btrfs_page_mkwrite+0x10c/0x4a0
 ? handle_mm_fault+0x5e/0x1730
 do_page_mkwrite+0x4d/0xc0
 ? __do_fault+0x32/0x150
 handle_mm_fault+0x103c/0x1730
 exc_page_fault+0x340/0x660
 ? asm_exc_page_fault+0x8/0x30
 asm_exc_page_fault+0x1e/0x30
RIP: 0033:0x7faa6c9969c4

Was seen in testing.  The fix is similar to that of

  btrfs: open device without device_list_mutex

where we're holding the device_list_mutex and then grab the bd_mutex,
which pulls in a bunch of dependencies under the bd_mutex.  We only ever
call btrfs_close_devices() on mount failure or unmount, so we're save to
not have the device_list_mutex here.  We're already holding the
uuid_mutex which keeps us safe from any external modification of the
fs_devices.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: move btrfs_rm_dev_replace_free_srcdev outside of all locks
Josef Bacik [Thu, 20 Aug 2020 15:18:27 +0000 (11:18 -0400)]
btrfs: move btrfs_rm_dev_replace_free_srcdev outside of all locks

When closing and freeing the source device we could end up doing our
final blkdev_put() on the bdev, which will grab the bd_mutex.  As such
we want to be holding as few locks as possible, so move this call
outside of the dev_replace->lock_finishing_cancel_unmount lock.  Since
we're modifying the fs_devices we need to make sure we're holding the
uuid_mutex here, so take that as well.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: remove alloc_list splice in btrfs_prepare_sprout
Nikolay Borisov [Wed, 12 Aug 2020 13:26:46 +0000 (16:26 +0300)]
btrfs: remove alloc_list splice in btrfs_prepare_sprout

btrfs_prepare_sprout is called when the first rw device is added to a
seed filesystem. This means the filesystem can't have its alloc_list
be non-empty, since seed filesystems are read only. Simply remove the
code altogether.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: document some invariants of seed code
Nikolay Borisov [Wed, 12 Aug 2020 14:04:36 +0000 (17:04 +0300)]
btrfs: document some invariants of seed code

Without good understanding of how seed devices works it's hard to grok
some of what the code in open_seed_devices or btrfs_prepare_sprout does.

Add comments hopefully reducing some of the cognitive load.

Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: switch seed device to list api
Nikolay Borisov [Thu, 16 Jul 2020 07:25:33 +0000 (10:25 +0300)]
btrfs: switch seed device to list api

While this patch touches a bunch of files the conversion is
straighforward. Instead of using the implicit linked list anchored at
btrfs_fs_devices::seed the code is switched to using
list_for_each_entry.

Previous patches in the series already factored out code that processed
both main and seed devices so in those cases the factored out functions
are called on the main fs_devices and then on every seed dev inside
list_for_each_entry.

Using list api also allows to simplify deletion from the seed dev list
performed in btrfs_rm_device and btrfs_rm_dev_replace_free_srcdev by
substituting a while() loop with a simple list_del_init.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: simplify setting/clearing fs_info to btrfs_fs_devices
Nikolay Borisov [Wed, 15 Jul 2020 10:48:49 +0000 (13:48 +0300)]
btrfs: simplify setting/clearing fs_info to btrfs_fs_devices

It makes no sense to have sysfs-related routines be responsible for
properly initialising the fs_info pointer of struct btrfs_fs_device.
Instead this can be streamlined by making it the responsibility of
btrfs_init_devices_late to initialize it. That function already
initializes fs_info of every individual device in btrfs_fs_devices.

As far as clearing it is concerned it makes sense to move it to
close_fs_devices. That function is only called when struct
btrfs_fs_devices is no longer in use - either for holding seeds or
main devices for a mounted filesystem.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make close_fs_devices return void
Nikolay Borisov [Wed, 15 Jul 2020 10:48:48 +0000 (13:48 +0300)]
btrfs: make close_fs_devices return void

The return value of this function conveys absolutely no information.
All callers already check the state of fs_devices->opened to decide how
to proceed. So convert the function to returning void. While at it make
btrfs_close_devices also return void.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: factor out loop logic from btrfs_free_extra_devids
Nikolay Borisov [Thu, 16 Jul 2020 07:17:04 +0000 (10:17 +0300)]
btrfs: factor out loop logic from btrfs_free_extra_devids

This prepares the code to switching seeds devices to a proper list.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: factor out reada loop in __reada_start_machine
Nikolay Borisov [Wed, 15 Jul 2020 10:48:46 +0000 (13:48 +0300)]
btrfs: factor out reada loop in __reada_start_machine

This is in preparation for moving fs_devices to proper lists.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: remove err variable from btrfs_get_extent
Nikolay Borisov [Mon, 3 Aug 2020 09:58:46 +0000 (12:58 +0300)]
btrfs: remove err variable from btrfs_get_extent

There's no practical reason too use 'err' as a variable to convey
errors. In fact it's value is either set explicitly in the beginning of
the function or it simply takes the value of 'ret'. Not conforming to
the usual pattern of having ret be the only variable used to convey
errors makes the code more error prone to bugs. In fact one such bug
was introduced by 6bf9e4bd6a27 ("btrfs: inode: Verify inode mode toi
avoid NULL pointer dereference") by assigning the error value to 'ret'
and not 'err'.

Let's fix that issue and make the function less tricky by leaving only
ret to convey error values.

Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: dio iomap DSYNC workaround
Josef Bacik [Thu, 3 Sep 2020 15:16:51 +0000 (11:16 -0400)]
btrfs: dio iomap DSYNC workaround

iomap dio will run generic_write_sync() for us if the iocb is DSYNC.
This is problematic for us because of 2 reasons:

1. we hold the inode_lock() during this operation, and we take it in
   generic_write_sync()
2. we hold a read lock on the dio_sem but take the write lock in fsync

Since we don't want to rip out this code right now, but reworking the
locking is a bit much to do at this point, work around this problem with
this masterpiece of a patch.

First, we clear DSYNC on the iocb so that the iomap stuff doesn't know
that it needs to handle the sync.  We save this fact in
current->journal_info, because we need to see do special things once
we're in iomap_begin, and we have no way to pass private information
into iomap_dio_rw().

Next we specify a separate iomap_dio_ops for sync, which implements an
->end_io() callback that gets called when the dio completes.  This is
important for AIO, because we really do need to run generic_write_sync()
if we complete asynchronously.  However if we're still in the submitting
context when we enter ->end_io() we clear the flag so that the submitter
knows they're the ones that needs to run generic_write_sync().

This is meant to be temporary.  We need to work out how to eliminate the
inode_lock() and the dio_sem in our fsync and use another mechanism to
protect these operations.

Tested-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: switch to iomap for direct IO
Goldwyn Rodrigues [Mon, 17 Aug 2020 16:18:21 +0000 (11:18 -0500)]
btrfs: switch to iomap for direct IO

We're using direct io implementation based on buffer heads. This patch
switches to the new iomap infrastructure.

Switch from __blockdev_direct_IO() to iomap_dio_rw().  Rename
btrfs_get_blocks_direct() to btrfs_dio_iomap_begin() and use it as
iomap_begin() for iomap direct I/O functions. This function allocates
and locks all the blocks required for the I/O.  btrfs_submit_direct() is
used as the submit_io() hook for direct I/O ops.

Since we need direct I/O reads to go through iomap_dio_rw(), we change
file_operations.read_iter() to a btrfs_file_read_iter() which calls
btrfs_direct_IO() for direct reads and falls back to
generic_file_buffered_read() for incomplete reads and buffered reads.

We don't need address_space.direct_IO() anymore: set it to noop.

Similarly, we don't need flags used in __blockdev_direct_IO(). iomap is
capable of direct I/O reads from a hole, so we don't need to return
-ENOENT.

Btrfs direct I/O is now done under i_rwsem, shared in case of reads and
exclusive in case of writes. This guards against simultaneous truncates.

Use iomap->iomap_end() to check for failed or incomplete direct I/O:

  - for writes, call __endio_write_update_ordered()
  - for reads, unlock extents

btrfs_dio_data is now hooked in iomap->private and not
current->journal_info. It carries the reservation variable and the
amount of data submitted, so we can calculate the amount of data to call
__endio_write_update_ordered in case of an error.

This patch removes last use of struct buffer_head from btrfs.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: add owner and fs_info to alloc_state io_tree
Qu Wenruo [Thu, 20 Aug 2020 07:42:46 +0000 (15:42 +0800)]
btrfs: add owner and fs_info to alloc_state io_tree

Commit 1c11b63eff2a ("btrfs: replace pending/pinned chunks lists with io
tree") introduced btrfs_device::alloc_state extent io tree, but it
doesn't initialize the fs_info and owner member.

This means the following features are not properly supported:

- Fs owner report for insert_state() error
  Without fs_info initialized, although btrfs_err() won't panic, it
  won't output which fs is causing the error.

- Wrong owner for trace events
  alloc_state will get the owner as pinned extents.

Fix this by assiging proper fs_info and owner for
btrfs_device::alloc_state.

Fixes: 1c11b63eff2a ("btrfs: replace pending/pinned chunks lists with io tree")
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make read_block_group_item return void
Marcos Paulo de Souza [Mon, 17 Aug 2020 13:56:10 +0000 (10:56 -0300)]
btrfs: make read_block_group_item return void

Since it's inclusion on 9afc66498a0b ("btrfs: block-group: refactor how
we read one block group item") this function always returned 0, so there
is no need to check for the returned value.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Marcos Paulo de Souza <mpdesouza@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: sysfs: fix unused-but-set-variable warnings
Leon Romanovsky [Wed, 19 Aug 2020 14:16:28 +0000 (17:16 +0300)]
btrfs: sysfs: fix unused-but-set-variable warnings

The compilation with W=1 generates the following warnings:
 fs/btrfs/sysfs.c:1630:6: warning: variable 'ret' set but not used [-Wunused-but-set-variable]
  1630 |  int ret;
       |      ^~~
 fs/btrfs/sysfs.c:1629:6: warning: variable 'features' set but not used [-Wunused-but-set-variable]
  1629 |  u64 features;
       |      ^~~~~~~~

[ The unused variables are leftover from e410e34fad91 ("Revert "btrfs:
  synchronize incompat feature bits with sysfs files""), which needs
  to be properly fixed by moving feature bit manipulation from the sysfs
  context.  Silence the warning to save pepople time, we got several
  reports. ]

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make fast fsyncs wait only for writeback
Filipe Manana [Tue, 11 Aug 2020 11:43:58 +0000 (12:43 +0100)]
btrfs: make fast fsyncs wait only for writeback

Currently regardless of a full or a fast fsync we always wait for ordered
extents to complete, and then start logging the inode after that. However
for fast fsyncs we can just wait for the writeback to complete, we don't
need to wait for the ordered extents to complete since we use the list of
modified extents maps to figure out which extents we must log and we can
get their checksums directly from the ordered extents that are still in
flight, otherwise look them up from the checksums tree.

Until commit b5e6c3e170b770 ("btrfs: always wait on ordered extents at
fsync time"), for fast fsyncs, we used to start logging without even
waiting for the writeback to complete first, we would wait for it to
complete after logging, while holding a transaction open, which lead to
performance issues when using cgroups and probably for other cases too,
as wait for IO while holding a transaction handle should be avoided as
much as possible. After that, for fast fsyncs, we started to wait for
ordered extents to complete before starting to log, which adds some
latency to fsyncs and we even got at least one report about a performance
drop which bisected to that particular change:

https://lore.kernel.org/linux-btrfs/20181109215148.GF23260@techsingularity.net/

This change makes fast fsyncs only wait for writeback to finish before
starting to log the inode, instead of waiting for both the writeback to
finish and for the ordered extents to complete. This brings back part of
the logic we had that extracts checksums from in flight ordered extents,
which are not yet in the checksums tree, and making sure transaction
commits wait for the completion of ordered extents previously logged
(by far most of the time they have already completed by the time a
transaction commit starts, resulting in no wait at all), to avoid any
data loss if an ordered extent completes after the transaction used to
log an inode is committed, followed by a power failure.

When there are no other tasks accessing the checksums and the subvolume
btrees, the ordered extent completion is pretty fast, typically taking
100 to 200 microseconds only in my observations. However when there are
other tasks accessing these btrees, ordered extent completion can take a
lot more time due to lock contention on nodes and leaves of these btrees.
I've seen cases over 2 milliseconds, which starts to be significant. In
particular when we do have concurrent fsyncs against different files there
is a lot of contention on the checksums btree, since we have many tasks
writing the checksums into the btree and other tasks that already started
the logging phase are doing lookups for checksums in the btree.

This change also turns all ranged fsyncs into full ranged fsyncs, which
is something we already did when not using the NO_HOLES features or when
doing a full fsync. This is to guarantee we never miss checksums due to
writeback having been triggered only for a part of an extent, and we end
up logging the full extent but only checksums for the written range, which
results in missing checksums after log replay. Allowing ranged fsyncs to
operate again only in the original range, when using the NO_HOLES feature
and doing a fast fsync is doable but requires some non trivial changes to
the writeback path, which can always be worked on later if needed, but I
don't think they are a very common use case.

Several tests were performed using fio for different numbers of concurrent
jobs, each writing and fsyncing its own file, for both sequential and
random file writes. The tests were run on bare metal, no virtualization,
on a box with 12 cores (Intel i7-8700), 64Gb of RAM and a NVMe device,
with a kernel configuration that is the default of typical distributions
(debian in this case), without debug options enabled (kasan, kmemleak,
slub debug, debug of page allocations, lock debugging, etc).

The following script that calls fio was used:

  $ cat test-fsync.sh
  #!/bin/bash

  DEV=/dev/nvme0n1
  MNT=/mnt/btrfs
  MOUNT_OPTIONS="-o ssd -o space_cache=v2"
  MKFS_OPTIONS="-d single -m single"

  if [ $# -ne 5 ]; then
    echo "Use $0 NUM_JOBS FILE_SIZE FSYNC_FREQ BLOCK_SIZE [write|randwrite]"
    exit 1
  fi

  NUM_JOBS=$1
  FILE_SIZE=$2
  FSYNC_FREQ=$3
  BLOCK_SIZE=$4
  WRITE_MODE=$5

  if [ "$WRITE_MODE" != "write" ] && [ "$WRITE_MODE" != "randwrite" ]; then
    echo "Invalid WRITE_MODE, must be 'write' or 'randwrite'"
    exit 1
  fi

  cat <<EOF > /tmp/fio-job.ini
  [writers]
  rw=$WRITE_MODE
  fsync=$FSYNC_FREQ
  fallocate=none
  group_reporting=1
  direct=0
  bs=$BLOCK_SIZE
  ioengine=sync
  size=$FILE_SIZE
  directory=$MNT
  numjobs=$NUM_JOBS
  EOF

  echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

  echo
  echo "Using config:"
  echo
  cat /tmp/fio-job.ini
  echo

  umount $MNT &> /dev/null
  mkfs.btrfs -f $MKFS_OPTIONS $DEV
  mount $MOUNT_OPTIONS $DEV $MNT
  fio /tmp/fio-job.ini
  umount $MNT

The results were the following:

*************************
*** sequential writes ***
*************************

==== 1 job, 8GiB file, fsync frequency 1, block size 64KiB ====

Before patch:

WRITE: bw=36.6MiB/s (38.4MB/s), 36.6MiB/s-36.6MiB/s (38.4MB/s-38.4MB/s), io=8192MiB (8590MB), run=223689-223689msec

After patch:

WRITE: bw=40.2MiB/s (42.1MB/s), 40.2MiB/s-40.2MiB/s (42.1MB/s-42.1MB/s), io=8192MiB (8590MB), run=203980-203980msec
(+9.8%, -8.8% runtime)

==== 2 jobs, 4GiB files, fsync frequency 1, block size 64KiB ====

Before patch:

WRITE: bw=35.8MiB/s (37.5MB/s), 35.8MiB/s-35.8MiB/s (37.5MB/s-37.5MB/s), io=8192MiB (8590MB), run=228950-228950msec

After patch:

WRITE: bw=43.5MiB/s (45.6MB/s), 43.5MiB/s-43.5MiB/s (45.6MB/s-45.6MB/s), io=8192MiB (8590MB), run=188272-188272msec
(+21.5% throughput, -17.8% runtime)

==== 4 jobs, 2GiB files, fsync frequency 1, block size 64KiB ====

Before patch:

WRITE: bw=50.1MiB/s (52.6MB/s), 50.1MiB/s-50.1MiB/s (52.6MB/s-52.6MB/s), io=8192MiB (8590MB), run=163446-163446msec

After patch:

WRITE: bw=64.5MiB/s (67.6MB/s), 64.5MiB/s-64.5MiB/s (67.6MB/s-67.6MB/s), io=8192MiB (8590MB), run=126987-126987msec
(+28.7% throughput, -22.3% runtime)

==== 8 jobs, 1GiB files, fsync frequency 1, block size 64KiB ====

Before patch:

WRITE: bw=64.0MiB/s (68.1MB/s), 64.0MiB/s-64.0MiB/s (68.1MB/s-68.1MB/s), io=8192MiB (8590MB), run=126075-126075msec

After patch:

WRITE: bw=86.8MiB/s (91.0MB/s), 86.8MiB/s-86.8MiB/s (91.0MB/s-91.0MB/s), io=8192MiB (8590MB), run=94358-94358msec
(+35.6% throughput, -25.2% runtime)

==== 16 jobs, 512MiB files, fsync frequency 1, block size 64KiB ====

Before patch:

WRITE: bw=79.8MiB/s (83.6MB/s), 79.8MiB/s-79.8MiB/s (83.6MB/s-83.6MB/s), io=8192MiB (8590MB), run=102694-102694msec

After patch:

WRITE: bw=107MiB/s (112MB/s), 107MiB/s-107MiB/s (112MB/s-112MB/s), io=8192MiB (8590MB), run=76446-76446msec
(+34.1% throughput, -25.6% runtime)

==== 32 jobs, 512MiB files, fsync frequency 1, block size 64KiB ====

Before patch:

WRITE: bw=93.2MiB/s (97.7MB/s), 93.2MiB/s-93.2MiB/s (97.7MB/s-97.7MB/s), io=16.0GiB (17.2GB), run=175836-175836msec

After patch:

WRITE: bw=111MiB/s (117MB/s), 111MiB/s-111MiB/s (117MB/s-117MB/s), io=16.0GiB (17.2GB), run=147001-147001msec
(+19.1% throughput, -16.4% runtime)

==== 64 jobs, 512MiB files, fsync frequency 1, block size 64KiB ====

Before patch:

WRITE: bw=108MiB/s (114MB/s), 108MiB/s-108MiB/s (114MB/s-114MB/s), io=32.0GiB (34.4GB), run=302656-302656msec

After patch:

WRITE: bw=133MiB/s (140MB/s), 133MiB/s-133MiB/s (140MB/s-140MB/s), io=32.0GiB (34.4GB), run=246003-246003msec
(+23.1% throughput, -18.7% runtime)

************************
***   random writes  ***
************************

==== 1 job, 8GiB file, fsync frequency 16, block size 4KiB ====

Before patch:

WRITE: bw=11.5MiB/s (12.0MB/s), 11.5MiB/s-11.5MiB/s (12.0MB/s-12.0MB/s), io=8192MiB (8590MB), run=714281-714281msec

After patch:

WRITE: bw=11.6MiB/s (12.2MB/s), 11.6MiB/s-11.6MiB/s (12.2MB/s-12.2MB/s), io=8192MiB (8590MB), run=705959-705959msec
(+0.9% throughput, -1.7% runtime)

==== 2 jobs, 4GiB files, fsync frequency 16, block size 4KiB ====

Before patch:

WRITE: bw=12.8MiB/s (13.5MB/s), 12.8MiB/s-12.8MiB/s (13.5MB/s-13.5MB/s), io=8192MiB (8590MB), run=638101-638101msec

After patch:

WRITE: bw=13.1MiB/s (13.7MB/s), 13.1MiB/s-13.1MiB/s (13.7MB/s-13.7MB/s), io=8192MiB (8590MB), run=625374-625374msec
(+2.3% throughput, -2.0% runtime)

==== 4 jobs, 2GiB files, fsync frequency 16, block size 4KiB ====

Before patch:

WRITE: bw=15.4MiB/s (16.2MB/s), 15.4MiB/s-15.4MiB/s (16.2MB/s-16.2MB/s), io=8192MiB (8590MB), run=531146-531146msec

After patch:

WRITE: bw=17.8MiB/s (18.7MB/s), 17.8MiB/s-17.8MiB/s (18.7MB/s-18.7MB/s), io=8192MiB (8590MB), run=460431-460431msec
(+15.6% throughput, -13.3% runtime)

==== 8 jobs, 1GiB files, fsync frequency 16, block size 4KiB ====

Before patch:

WRITE: bw=19.9MiB/s (20.8MB/s), 19.9MiB/s-19.9MiB/s (20.8MB/s-20.8MB/s), io=8192MiB (8590MB), run=412664-412664msec

After patch:

WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=8192MiB (8590MB), run=368589-368589msec
(+11.6% throughput, -10.7% runtime)

==== 16 jobs, 512MiB files, fsync frequency 16, block size 4KiB ====

Before patch:

WRITE: bw=29.3MiB/s (30.7MB/s), 29.3MiB/s-29.3MiB/s (30.7MB/s-30.7MB/s), io=8192MiB (8590MB), run=279924-279924msec

After patch:

WRITE: bw=30.4MiB/s (31.9MB/s), 30.4MiB/s-30.4MiB/s (31.9MB/s-31.9MB/s), io=8192MiB (8590MB), run=269258-269258msec
(+3.8% throughput, -3.8% runtime)

==== 32 jobs, 512MiB files, fsync frequency 16, block size 4KiB ====

Before patch:

WRITE: bw=36.9MiB/s (38.7MB/s), 36.9MiB/s-36.9MiB/s (38.7MB/s-38.7MB/s), io=16.0GiB (17.2GB), run=443581-443581msec

After patch:

WRITE: bw=41.6MiB/s (43.6MB/s), 41.6MiB/s-41.6MiB/s (43.6MB/s-43.6MB/s), io=16.0GiB (17.2GB), run=394114-394114msec
(+12.7% throughput, -11.2% runtime)

==== 64 jobs, 512MiB files, fsync frequency 16, block size 4KiB ====

Before patch:

WRITE: bw=45.9MiB/s (48.1MB/s), 45.9MiB/s-45.9MiB/s (48.1MB/s-48.1MB/s), io=32.0GiB (34.4GB), run=714614-714614msec

After patch:

WRITE: bw=48.8MiB/s (51.1MB/s), 48.8MiB/s-48.8MiB/s (51.1MB/s-51.1MB/s), io=32.0GiB (34.4GB), run=672087-672087msec
(+6.3% throughput, -6.0% runtime)

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: do not commit logs and transactions during link and rename operations
Filipe Manana [Tue, 11 Aug 2020 11:43:48 +0000 (12:43 +0100)]
btrfs: do not commit logs and transactions during link and rename operations

Since commit d4682ba03ef618 ("Btrfs: sync log after logging new name") we
started to commit logs, and fallback to transaction commits when we failed
to log the new names or commit the logs, after link and rename operations
when the target inodes (or their parents) were previously logged in the
current transaction. This was to avoid losing directories despite an
explicit fsync on them when they are ancestors of some inode that got a
new named logged, due to a link or rename operation. However that adds the
cost of starting IO and waiting for it to complete, which can cause higher
latencies for applications.

Instead of doing that, just make sure that when we log a new name for an
inode we don't mark any of its ancestors as logged, so that if any one
does an fsync against any of them, without doing any other change on them,
the fsync commits the log. This way we only pay the cost of a log commit
(or a transaction commit if something goes wrong or a new block group was
created) if the application explicitly asks to fsync any of the parent
directories.

Using dbench, which mixes several filesystems operations including renames,
revealed some significant latency gains. The following script that uses
dbench was used to test this:

  #!/bin/bash

  DEV=/dev/nvme0n1
  MNT=/mnt/btrfs
  MOUNT_OPTIONS="-o ssd -o space_cache=v2"
  MKFS_OPTIONS="-m single -d single"
  THREADS=16

  echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
  mkfs.btrfs -f $MKFS_OPTIONS $DEV
  mount $MOUNT_OPTIONS $DEV $MNT

  dbench -t 300 -D $MNT $THREADS

  umount $MNT

The test was run on bare metal, no virtualization, on a box with 12 cores
(Intel i7-8700), 64Gb of RAM and using a NVMe device, with a kernel
configuration that is the default of typical distributions (debian in this
case), without debug options enabled (kasan, kmemleak, slub debug, debug
of page allocations, lock debugging, etc).

Results before this patch:

 Operation      Count    AvgLat    MaxLat
 ----------------------------------------
 NTCreateX    10750455     0.011   155.088
 Close         7896674     0.001     0.243
 Rename         455222     2.158  1101.947
 Unlink        2171189     0.067   121.638
 Deltree           256     2.425     7.816
 Mkdir             128     0.002     0.003
 Qpathinfo     9744323     0.006    21.370
 Qfileinfo     1707092     0.001     0.146
 Qfsinfo       1786756     0.001    11.228
 Sfileinfo      875612     0.003    21.263
 Find          3767281     0.025     9.617
 WriteX        5356924     0.011   211.390
 ReadX        16852694     0.003     9.442
 LockX           35008     0.002     0.119
 UnlockX         35008     0.001     0.138
 Flush          753458     4.252  1102.249

Throughput 1128.35 MB/sec  16 clients  16 procs  max_latency=1102.255 ms

Results after this patch:

16 clients, after

 Operation      Count    AvgLat    MaxLat
 ----------------------------------------
 NTCreateX    11471098     0.012   448.281
 Close         8426396     0.001     0.925
 Rename         485746     0.123   267.183
 Unlink        2316477     0.080    63.433
 Deltree           288     2.830    11.144
 Mkdir             144     0.003     0.010
 Qpathinfo    10397420     0.006    10.288
 Qfileinfo     1822039     0.001     0.169
 Qfsinfo       1906497     0.002    14.039
 Sfileinfo      934433     0.004     2.438
 Find          4019879     0.026    10.200
 WriteX        5718932     0.011   200.985
 ReadX        17981671     0.003    10.036
 LockX           37352     0.002     0.076
 UnlockX         37352     0.001     0.109
 Flush          804018     5.015   778.033

Throughput 1201.98 MB/sec  16 clients  16 procs  max_latency=778.036 ms
(+6.5% throughput, -29.4% max latency, -75.8% rename latency)

Test case generic/498 from fstests tests the scenario that the previously
mentioned commit fixed.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: do not take the log_mutex of the subvolume when pinning the log
Filipe Manana [Tue, 11 Aug 2020 11:43:37 +0000 (12:43 +0100)]
btrfs: do not take the log_mutex of the subvolume when pinning the log

During a rename we pin the log to make sure no one commits a log that
reflects an ongoing rename operation, as it might result in a committed
log where it recorded the unlink of the old name without having recorded
the new name. However we are taking the subvolume's log_mutex before
incrementing the log_writers counter, which is not necessary since that
counter is atomic and we only remove the old name from the log and add
the new name to the log after we have incremented log_writers, ensuring
that no one can commit the log after we have removed the old name from
the log and before we added the new name to the log.

By taking the log_mutex lock we are just adding unnecessary contention on
the lock, which can become visible for workloads that mix renames with
fsyncs, writes for files opened with O_SYNC and unlink operations (if the
inode or its parent were fsynced before in the current transaction).

So just remove the lock and unlock of the subvolume's log_mutex at
btrfs_pin_log_trans().

Using dbench, which mixes different types of operations that end up taking
that mutex (fsyncs, renames, unlinks and writes into files opened with
O_SYNC) revealed some small gains. The following script that calls dbench
was used:

  #!/bin/bash

  DEV=/dev/nvme0n1
  MNT=/mnt/btrfs
  MOUNT_OPTIONS="-o ssd -o space_cache=v2"
  MKFS_OPTIONS="-m single -d single"
  THREADS=32

  echo "performance" | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
  mkfs.btrfs -f $MKFS_OPTIONS $DEV
  mount $MOUNT_OPTIONS $DEV $MNT

  dbench -s -t 600 -D $MNT $THREADS

  umount $MNT

The test was run on bare metal, no virtualization, on a box with 12 cores
(Intel i7-8700), 64Gb of RAM and using a NVMe device, with a kernel
configuration that is the default of typical distributions (debian in this
case), without debug options enabled (kasan, kmemleak, slub debug, debug
of page allocations, lock debugging, etc).

Results before this patch:

 Operation      Count    AvgLat    MaxLat
 ----------------------------------------
 NTCreateX    4410848     0.017   738.640
 Close        3240222     0.001     0.834
 Rename        186850     7.478  1272.476
 Unlink        890875     0.128   785.018
 Deltree          128     2.846    12.081
 Mkdir             64     0.002     0.003
 Qpathinfo    3997659     0.009    11.171
 Qfileinfo     701307     0.001     0.478
 Qfsinfo       733494     0.002     1.103
 Sfileinfo     359362     0.004     3.266
 Find         1546226     0.041     4.128
 WriteX       2202803     7.905  1376.989
 ReadX        6917775     0.003     3.887
 LockX          14392     0.002     0.043
 UnlockX        14392     0.001     0.085
 Flush         309225     0.128  1033.936

Throughput 231.555 MB/sec (sync open)  32 clients  32 procs  max_latency=1376.993 ms

Results after this patch:

Operation      Count    AvgLat    MaxLat
 ----------------------------------------
 NTCreateX    4603244     0.017   232.776
 Close        3381299     0.001     1.041
 Rename        194871     7.251  1073.165
 Unlink        929730     0.133   119.233
 Deltree          128     2.871    10.199
 Mkdir             64     0.002     0.004
 Qpathinfo    4171343     0.009    11.317
 Qfileinfo     731227     0.001     1.635
 Qfsinfo       765079     0.002     3.568
 Sfileinfo     374881     0.004     1.220
 Find         1612964     0.041     4.675
 WriteX       2296720     7.569  1178.204
 ReadX        7213633     0.003     3.075
 LockX          14976     0.002     0.076
 UnlockX        14976     0.001     0.061
 Flush         322635     0.102   579.505

Throughput 241.4 MB/sec (sync open)  32 clients  32 procs  max_latency=1178.207 ms
(+4.3% throughput, -14.4% max latency)

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: send: remove indirect callback parameter for changed_cb
David Sterba [Mon, 17 Aug 2020 10:16:57 +0000 (12:16 +0200)]
btrfs: send: remove indirect callback parameter for changed_cb

There's a custom callback passed to btrfs_compare_trees which happens to
be named exactly same as the existing function implementing it. This is
confusing and the indirection is not necessary for our needs. Compiler
is clever enough to call it directly so there's effectively no change.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: scrub: rename ratelimit state varaible to avoid shadowing
David Sterba [Mon, 17 Aug 2020 10:12:38 +0000 (12:12 +0200)]
btrfs: scrub: rename ratelimit state varaible to avoid shadowing

There's already defined _rs within ctree.h:btrfs_printk_ratelimited,
local variables should not use _ to avoid such name clashes with
macro-local variables.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: remove unnecessarily shadowed variables
David Sterba [Mon, 17 Aug 2020 10:08:37 +0000 (12:08 +0200)]
btrfs: remove unnecessarily shadowed variables

In btrfs_orphan_cleanup, there's another instance of fs_info, but it's
the same as the one we already have.

In btrfs_backref_finish_upper_links, rb_node is same type and used
as temporary cursor to the tree.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: compression: move declarations to header
David Sterba [Mon, 17 Aug 2020 08:58:38 +0000 (10:58 +0200)]
btrfs: compression: move declarations to header

The declarations of compression algorithm callbacks are defined in the
.c file as they're used from there. Compiler warns that there are no
declarations for public functions when compiling lzo.c/zlib.c/zstd.c.
Fix that by moving the declarations to the header as it's the common
place for all of them.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: remove const from btrfs_feature_set_name
David Sterba [Mon, 17 Aug 2020 08:56:00 +0000 (10:56 +0200)]
btrfs: remove const from btrfs_feature_set_name

The function btrfs_feature_set_name returns a const char pointer, the
second const is not necessary and reported as a warning:

 In file included from fs/btrfs/space-info.c:6:
 fs/btrfs/sysfs.h:16:1: warning: type qualifiers ignored on function return type [-Wignored-qualifiers]
    16 | const char * const btrfs_feature_set_name(enum btrfs_feature_set set);
       | ^~~~~

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: cleanup calculation of lockend in lock_and_cleanup_extent_if_need()
Qu Wenruo [Thu, 13 Aug 2020 06:33:52 +0000 (14:33 +0800)]
btrfs: cleanup calculation of lockend in lock_and_cleanup_extent_if_need()

We're just doing rounding up to sectorsize to calculate the lockend.
There is no need to do the unnecessary length calculation, just direct
round_up() is enough.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: fix possible infinite loop in data async reclaim
Josef Bacik [Tue, 25 Aug 2020 20:56:59 +0000 (16:56 -0400)]
btrfs: fix possible infinite loop in data async reclaim

Dave reported an issue where generic/102 would sometimes hang.  This
turned out to be because we'd get into this spot where we were no longer
making progress on data reservations because our exit condition was not
met.  The log is basically

while (!space_info->full && !list_empty(&space_info->tickets))
flush_space(space_info, flush_state);

where flush state is our various flush states, but doesn't include
ALLOC_CHUNK_FORCE.  This is because we actually lead with allocating
chunks, and so the assumption was that once you got to the actual
flushing states you could no longer allocate chunks.  This was a stupid
assumption, because you could have deleted block groups that would be
reclaimed by a transaction commit, thus unsetting space_info->full.
This is essentially what happens with generic/102, and so sometimes
you'd get stuck in the flushing loop because we weren't allocating
chunks, but flushing space wasn't giving us what we needed to make
progress.

Fix this by adding ALLOC_CHUNK_FORCE to the end of our flushing states,
that way we will eventually bail out because we did end up with
space_info->full if we free'd a chunk previously.  Otherwise, as is the
case for this test, we'll allocate our chunk and continue on our happy
merry way.

Reported-by: David Sterba <dsterba@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: add a comment explaining the data flush steps
Josef Bacik [Tue, 21 Jul 2020 14:22:34 +0000 (10:22 -0400)]
btrfs: add a comment explaining the data flush steps

The data flushing steps are not obvious to people other than myself and
Chris.  Write a giant comment explaining the reasoning behind each flush
step for data as well as why it is in that particular order.

Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: do async reclaim for data reservations
Josef Bacik [Tue, 21 Jul 2020 14:22:33 +0000 (10:22 -0400)]
btrfs: do async reclaim for data reservations

Now that we have the data ticketing stuff in place, move normal data
reservations to use an async reclaim helper to satisfy tickets.  Before
we could have multiple tasks race in and both allocate chunks, resulting
in more data chunks than we would necessarily need.  Serializing these
allocations and making a single thread responsible for flushing will
only allocate chunks as needed, as well as cut down on transaction
commits and other flush related activities.

Priority reservations will still work as they have before, simply
trying to allocate a chunk until they can make their reservation.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: flush delayed refs when trying to reserve data space
Josef Bacik [Tue, 21 Jul 2020 14:22:32 +0000 (10:22 -0400)]
btrfs: flush delayed refs when trying to reserve data space

We can end up with freed extents in the delayed refs, and thus
may_commit_transaction() may not think we have enough pinned space to
commit the transaction and we'll ENOSPC early.  Handle this by running
the delayed refs in order to make sure pinned is uptodate before we try
to commit the transaction.

Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: run delayed iputs before committing the transaction for data
Josef Bacik [Tue, 21 Jul 2020 14:22:31 +0000 (10:22 -0400)]
btrfs: run delayed iputs before committing the transaction for data

Before we were waiting on iputs after we committed the transaction, but
this doesn't really make much sense.  We want to reclaim any space we
may have in order to be more likely to commit the transaction, due to
pinned space being added by running the delayed iputs.  Fix this by
making delayed iputs run before committing the transaction.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: don't force commit if we are data
Josef Bacik [Tue, 21 Jul 2020 14:22:30 +0000 (10:22 -0400)]
btrfs: don't force commit if we are data

We used to unconditionally commit the transaction at least 2 times and
then on the 3rd try check against pinned space to make sure committing
the transaction was worth the effort.  This is overkill, we know nobody
is going to steal our reservation, and if we can't make our reservation
with the pinned amount simply bail out.

This also cleans up the passing of bytes_needed to
may_commit_transaction, as that was the thing we added into place in
order to accomplish this behavior.  We no longer need it so remove that
mess.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: drop the commit_cycles stuff for data reservations
Josef Bacik [Tue, 21 Jul 2020 14:22:29 +0000 (10:22 -0400)]
btrfs: drop the commit_cycles stuff for data reservations

This was an old wart left over from how we previously did data
reservations.  Before we could have people race in and take a
reservation while we were flushing space, so we needed to make sure we
looped a few times before giving up.  Now that we're using the ticketing
infrastructure we don't have to worry about this and can drop the logic
altogether.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: use the same helper for data and metadata reservations
Josef Bacik [Tue, 21 Jul 2020 14:22:28 +0000 (10:22 -0400)]
btrfs: use the same helper for data and metadata reservations

Now that data reservations follow the same pattern as metadata
reservations we can simply rename __reserve_metadata_bytes to
__reserve_bytes and use that helper for data reservations.

Things to keep in mind, btrfs_can_overcommit() returns 0 for data,
because we can never overcommit.  We also will never pass in FLUSH_ALL
for data, so we'll simply be added to the priority list and go straight
into handle_reserve_ticket.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: serialize data reservations if we are flushing
Josef Bacik [Tue, 21 Jul 2020 14:22:27 +0000 (10:22 -0400)]
btrfs: serialize data reservations if we are flushing

Nikolay reported a problem where generic/371 would fail sometimes with a
slow drive.  The gist of the test is that we fallocate a file in
parallel with a pwrite of a different file.  These two files combined
are smaller than the file system, but sometimes the pwrite would ENOSPC.

A fair bit of investigation uncovered the fact that the fallocate
workload was racing in and grabbing the free space that the pwrite
workload was trying to free up so it could make its own reservation.
After a few loops of this eventually the pwrite workload would error out
with an ENOSPC.

We've had the same problem with metadata as well, and we serialized all
metadata allocations to satisfy this problem.  This wasn't usually a
problem with data because data reservations are more straightforward,
but obviously could still happen.

Fix this by not allowing reservations to occur if there are any pending
tickets waiting to be satisfied on the space info.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: use ticketing for data space reservations
Josef Bacik [Tue, 21 Jul 2020 14:22:26 +0000 (10:22 -0400)]
btrfs: use ticketing for data space reservations

Now that we have all the infrastructure in place, use the ticketing
infrastructure to make data allocations.  This still maintains the exact
same flushing behavior, but now we're using tickets to get our
reservations satisfied.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: add btrfs_reserve_data_bytes and use it
Josef Bacik [Tue, 21 Jul 2020 14:22:25 +0000 (10:22 -0400)]
btrfs: add btrfs_reserve_data_bytes and use it

Create a new function btrfs_reserve_data_bytes() in order to handle data
reservations.  This uses the new flush types and flush states to handle
making data reservations.

This patch specifically does not change any functionality, and is
purposefully not cleaned up in order to make bisection easier for the
future patches.  The new helper is identical to the old helper in how it
handles data reservations.  We first try to force a chunk allocation,
and then we run through the flush states all at once and in the same
order that they were done with the old helper.

Subsequent patches will clean this up and change the behavior of the
flushing, and it is important to keep those changes separate so we can
easily bisect down to the patch that caused the regression, rather than
the patch that made us start using the new infrastructure.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: add the data transaction commit logic into may_commit_transaction
Josef Bacik [Tue, 21 Jul 2020 14:22:24 +0000 (10:22 -0400)]
btrfs: add the data transaction commit logic into may_commit_transaction

Data space flushing currently unconditionally commits the transaction
twice in a row, and the last time it checks if there's enough pinned
extents to satisfy its reservation before deciding to commit the
transaction for the 3rd and final time.

Encode this logic into may_commit_transaction().  In the next patch we
will pass in U64_MAX for bytes_needed the first two times, and the final
time we will pass in the actual bytes we need so the normal logic will
apply.

This patch exists solely to make the logical changes I will make to the
flushing state machine separate to make it easier to bisect any
performance related regressions.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: add flushing states for handling data reservations
Josef Bacik [Tue, 21 Jul 2020 14:22:23 +0000 (10:22 -0400)]
btrfs: add flushing states for handling data reservations

Currently the way we do data reservations is by seeing if we have enough
space in our space_info.  If we do not and we're a normal inode we'll

1) Attempt to force a chunk allocation until we can't anymore.
2) If that fails we'll flush delalloc, then commit the transaction, then
   run the delayed iputs.

If we are a free space inode we're only allowed to force a chunk
allocation.  In order to use the normal flushing mechanism we need to
encode this into a flush state array for normal inodes.  Since both will
start with allocating chunks until the space info is full there is no
need to add this as a flush state, this will be handled specially.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: check tickets after waiting on ordered extents
Josef Bacik [Tue, 21 Jul 2020 14:22:22 +0000 (10:22 -0400)]
btrfs: check tickets after waiting on ordered extents

Right now if the space is freed up after the ordered extents complete
(which is likely since the reservations are held until they complete),
we would do extra delalloc flushing before we'd notice that we didn't
have any more tickets.  Fix this by moving the tickets check after our
wait_ordered_extents check.

Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: use btrfs_start_delalloc_roots in shrink_delalloc
Josef Bacik [Tue, 21 Jul 2020 14:22:21 +0000 (10:22 -0400)]
btrfs: use btrfs_start_delalloc_roots in shrink_delalloc

The original iteration of flushing had us flushing delalloc and then
checking to see if we could make our reservation, thus we were very
careful about how many pages we would flush at once.

But now that everything is async and we satisfy tickets as the space
becomes available we don't have to keep track of any of this, simply
try and flush the number of dirty inodes we may have in order to
reclaim space to make our reservation.  This cleans up our delalloc
flushing significantly.

The async_pages stuff is dropped because btrfs_start_delalloc_roots()
handles the case that we generate async extents for us, so we no longer
require this extra logic.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: use the btrfs_space_info_free_bytes_may_use helper for delalloc
Josef Bacik [Tue, 21 Jul 2020 14:22:20 +0000 (10:22 -0400)]
btrfs: use the btrfs_space_info_free_bytes_may_use helper for delalloc

We are going to use the ticket infrastructure for data, so use the
btrfs_space_info_free_bytes_may_use() helper in
btrfs_free_reserved_data_space_noquota() so we get the
btrfs_try_granting_tickets call when we free our reservation.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: call btrfs_try_granting_tickets when reserving space
Josef Bacik [Tue, 21 Jul 2020 14:22:19 +0000 (10:22 -0400)]
btrfs: call btrfs_try_granting_tickets when reserving space

If we have compression on we could free up more space than we reserved,
and thus be able to make a space reservation.  Add the call for this
scenario.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: call btrfs_try_granting_tickets when unpinning anything
Josef Bacik [Tue, 21 Jul 2020 14:22:18 +0000 (10:22 -0400)]
btrfs: call btrfs_try_granting_tickets when unpinning anything

When unpinning we were only calling btrfs_try_granting_tickets() if
global_rsv->space_info == space_info, which is problematic because we
use ticketing for SYSTEM chunks, and want to use it for DATA as well.
Fix this by moving this call outside of that if statement.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: call btrfs_try_granting_tickets when freeing reserved bytes
Josef Bacik [Tue, 21 Jul 2020 14:22:17 +0000 (10:22 -0400)]
btrfs: call btrfs_try_granting_tickets when freeing reserved bytes

We were missing a call to btrfs_try_granting_tickets in
btrfs_free_reserved_bytes, so add it to handle the case where we're able
to satisfy an allocation because we've freed a pending reservation.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make ALLOC_CHUNK use the space info flags
Josef Bacik [Tue, 21 Jul 2020 14:22:16 +0000 (10:22 -0400)]
btrfs: make ALLOC_CHUNK use the space info flags

We have traditionally used flush_space() to flush metadata space, so
we've been unconditionally using btrfs_metadata_alloc_profile() for our
profile to allocate a chunk. However if we're going to use this for
data we need to use btrfs_get_alloc_profile() on the space_info we pass
in.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
4 years agobtrfs: make shrink_delalloc take space_info as an arg
Josef Bacik [Tue, 21 Jul 2020 14:22:15 +0000 (10:22 -0400)]
btrfs: make shrink_delalloc take space_info as an arg

Currently shrink_delalloc just looks up the metadata space info, but
this won't work if we're trying to reclaim space for data chunks.  We
get the right space_info we want passed into flush_space, so simply pass
that along to shrink_delalloc.

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Tested-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>