On 2019/6/28 上午11:28, Anand Jain wrote:
> On 28/6/19 10:44 AM, Qu Wenruo wrote:
>>
>>
>> On 2019/6/28 上午10:26, Anand Jain wrote:
>>> At the time mkfs.btrfs the device id and stripe index gets reversed as
>>> shown in [1]. This patch helps to keep them in order at the time of
>>> mkfs.btrfs. And makes it easier to debug.
>>>
>>> Before:
>>> Stripe 0 is on devid 2; Stipe 1 is on devid 1;
>>>
>>> ./mkfs.btrfs -fq -draid1 -mraid1 /dev/sdb /dev/sdc && btrfs in
>>> dump-tree -d /dev/sdb | grep -A 10000 "chunk tree" | grep -B 10000
>>> "device tree" | grep -A 13  "FIRST_CHUNK_TREE CHUNK_ITEM"
>>>     item 2 key (FIRST_CHUNK_TREE CHUNK_ITEM 22020096) itemoff 15975
>>> itemsize 112
>>>         length 8388608 owner 2 stripe_len 65536 type SYSTEM|RAID1
>>>         io_align 65536 io_width 65536 sector_size 4096
>>>         num_stripes 2 sub_stripes 0
>>>             stripe 0 devid 2 offset 1048576
>>>             dev_uuid d9fe51c4-6e79-446d-87ee-5be3184798cd
>>>             stripe 1 devid 1 offset 22020096
>>>             dev_uuid 16f626ca-1a54-469b-ac7e-25623af884ab
>>>     item 3 key (FIRST_CHUNK_TREE CHUNK_ITEM 30408704) itemoff 15863
>>> itemsize 112
>>>         length 268435456 owner 2 stripe_len 65536 type METADATA|RAID1
>>>         io_align 65536 io_width 65536 sector_size 4096
>>>         num_stripes 2 sub_stripes 0
>>>             stripe 0 devid 2 offset 9437184
>>>             dev_uuid d9fe51c4-6e79-446d-87ee-5be3184798cd
>>>             stripe 1 devid 1 offset 30408704
>>>             dev_uuid 16f626ca-1a54-469b-ac7e-25623af884ab
>>>     item 4 key (FIRST_CHUNK_TREE CHUNK_ITEM 298844160) itemoff 15751
>>> itemsize 112
>>>         length 314572800 owner 2 stripe_len 65536 type DATA|RAID1
>>>         io_align 65536 io_width 65536 sector_size 4096
>>>         num_stripes 2 sub_stripes 0
>>>             stripe 0 devid 2 offset 277872640
>>>             dev_uuid d9fe51c4-6e79-446d-87ee-5be3184798cd
>>>             stripe 1 devid 1 offset 298844160
>>>             dev_uuid 16f626ca-1a54-469b-ac7e-25623af884ab
>>>
>>> After:
>>> Stripe 0 is on devid 1; Stripe 1 is on devid 2
>>>
>>> ./mkfs.btrfs -fq -draid1 -mraid1 /dev/sdb /dev/sdc && btrfs in
>>> dump-tree -d /dev/sdb | grep -A 10000 "chunk tree" | grep -B 10000
>>> "device tree" | grep -A 13  "FIRST_CHUNK_TREE CHUNK_ITEM"
>>> /dev/sdb: 8 bytes were erased at offset 0x00010040 (btrfs): 5f 42 48
>>> 52 66 53 5f 4d
>>> /dev/sdc: 8 bytes were erased at offset 0x00010040 (btrfs): 5f 42 48
>>> 52 66 53 5f 4d
>>>     item 2 key (FIRST_CHUNK_TREE CHUNK_ITEM 22020096) itemoff 15975
>>> itemsize 112
>>>         length 8388608 owner 2 stripe_len 65536 type SYSTEM|RAID1
>>>         io_align 65536 io_width 65536 sector_size 4096
>>>         num_stripes 2 sub_stripes 0
>>>             stripe 0 devid 1 offset 22020096
>>>             dev_uuid 6abc88fa-f42e-4f0c-9bc3-2225735e51d1
>>>             stripe 1 devid 2 offset 1048576
>>>             dev_uuid 73746d27-13a6-4d58-ac6b-48c90c31d94d
>>>     item 3 key (FIRST_CHUNK_TREE CHUNK_ITEM 30408704) itemoff 15863
>>> itemsize 112
>>>         length 268435456 owner 2 stripe_len 65536 type METADATA|RAID1
>>>         io_align 65536 io_width 65536 sector_size 4096
>>>         num_stripes 2 sub_stripes 0
>>>             stripe 0 devid 1 offset 30408704
>>>             dev_uuid 6abc88fa-f42e-4f0c-9bc3-2225735e51d1
>>>             stripe 1 devid 2 offset 9437184
>>>             dev_uuid 73746d27-13a6-4d58-ac6b-48c90c31d94d
>>>     item 4 key (FIRST_CHUNK_TREE CHUNK_ITEM 298844160) itemoff 15751
>>> itemsize 112
>>>         length 314572800 owner 2 stripe_len 65536 type DATA|RAID1
>>>         io_align 65536 io_width 65536 sector_size 4096
>>>         num_stripes 2 sub_stripes 0
>>>             stripe 0 devid 1 offset 298844160
>>>             dev_uuid 6abc88fa-f42e-4f0c-9bc3-2225735e51d1
>>>             stripe 1 devid 2 offset 277872640
>>>             dev_uuid 73746d27-13a6-4d58-ac6b-48c90c31d94d
>>>
>>> Signed-off-by: Anand Jain <anand.jain@oracle.com>
>>
>> Reviewed-by: Qu Wenruo <wqu@suse.com>
>>
>> But please also check the comment inlined below.
>>> ---
>>>   volumes.c | 4 ++--
>>>   1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/volumes.c b/volumes.c
>>> index 79d1d6a07fb7..8c8b17e814b8 100644
>>> --- a/volumes.c
>>> +++ b/volumes.c
>>> @@ -1109,7 +1109,7 @@ again:
>>>               return ret;
>>>           cur = cur->next;
>>>           if (avail >= min_free) {
>>> -            list_move_tail(&device->dev_list, &private_devs);
>>> +            list_move(&device->dev_list, &private_devs);
>>
>> This is OK since current btrfs-progs chunk allocator doesn't follow the
>> kernel behavior by sorting devices with its unallocated space.
>> So it's completely devid based.
>>
>> But please keep in mind that, if we're going to unify the chunk
>> allocator behavior of kernel and btrfs-progs, the behavior will change.
>>
>> As the initial temporary chunk is always allocated on devid 1, reducing
>> its unallocated space thus reducing its priority in chunk allocator, and
>> making the devid sequence more unreliable.
> 
>  Right. For the debug here, I have an experimental code which disables
>  the unallocated space sort in the kernel. I don't have a strong reason
>  to disable the sort in the kernel so didn't send the patch.

I'd say that unallocated sort is a hidden way to prevent starvation.

The mostly common case is 3 disk RAID1. (1024M X 3)
With the unallocated space sort, we can take full use of 1.5T.

While without that, we can only use 1T, as all allocation will happen on
the first (or last) 2 devices, not utilize the remaining disk at all.

So that kernel part is very helpful to prevent starvation.

Thanks,
Qu

> 
> Thanks, Anand
> 
>> Thanks,
>> Qu
>>
>>>               index++;
>>>               if (type & BTRFS_BLOCK_GROUP_DUP)
>>>                   index++;
>>> @@ -1166,7 +1166,7 @@ again:
>>>           /* loop over this device again if we're doing a dup group */
>>>           if (!(type & BTRFS_BLOCK_GROUP_DUP) ||
>>>               (index == num_stripes - 1))
>>> -            list_move_tail(&device->dev_list, dev_list);
>>> +            list_move(&device->dev_list, dev_list);
>>>             ret = btrfs_alloc_dev_extent(trans, device, key.offset,
>>>                    calc_size, &dev_offset);
>>>
>>
>