diff mbox series

btrfs: relocation: Add an introduction for how relocation works.

Message ID 20200115062818.41268-1-wqu@suse.com (mailing list archive)
State New, archived
Headers show
Series btrfs: relocation: Add an introduction for how relocation works. | expand

Commit Message

Qu Wenruo Jan. 15, 2020, 6:28 a.m. UTC
Relocation is one of the most complex part of btrfs, while it's also the
foundation stone for online resizing, profile converting.

For such a complex facility, we should at least have some introduction
to it.

This patch will add an basic introduction at pretty a high level,
explaining:
- What relocation does
- How relocation is done
  Only mentioning how data reloc tree and reloc tree are involved in the
  operation.
  No details like the backref cache, or the data reloc tree contents.
- Which function to refer.

More detailed comments will be added for reloc tree creation, data reloc
tree creation and backref cache.

For now the introduction should save reader some time before digging
into the rabbit hole.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 fs/btrfs/relocation.c | 44 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 44 insertions(+)

Comments

Filipe Manana Jan. 15, 2020, 10:34 a.m. UTC | #1
On Wed, Jan 15, 2020 at 6:29 AM Qu Wenruo <wqu@suse.com> wrote:
>
> Relocation is one of the most complex part of btrfs, while it's also the
> foundation stone for online resizing, profile converting.
>
> For such a complex facility, we should at least have some introduction
> to it.
>
> This patch will add an basic introduction at pretty a high level,
> explaining:
> - What relocation does
> - How relocation is done
>   Only mentioning how data reloc tree and reloc tree are involved in the
>   operation.
>   No details like the backref cache, or the data reloc tree contents.
> - Which function to refer.
>
> More detailed comments will be added for reloc tree creation, data reloc
> tree creation and backref cache.
>
> For now the introduction should save reader some time before digging
> into the rabbit hole.
>
> Signed-off-by: Qu Wenruo <wqu@suse.com>
> ---
>  fs/btrfs/relocation.c | 44 +++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 44 insertions(+)
>
> diff --git a/fs/btrfs/relocation.c b/fs/btrfs/relocation.c
> index d897a8e5e430..cd3a15f1716d 100644
> --- a/fs/btrfs/relocation.c
> +++ b/fs/btrfs/relocation.c
> @@ -23,6 +23,50 @@
>  #include "delalloc-space.h"
>  #include "block-group.h"
>
> +/*
> + * Introduction for btrfs relocation.
> + *
> + * [What does relocation do]

For readability, a blank line here would help.

> + * The objective of relocation is to relocate all or some extents of one block
> + * group to other block groups.

Some? We always relocate all extents of a block group (except if
errors happen of course).

> + * This is utilized by resize (shrink only), profile converting, or just
> + * balance routine to free some block groups.
> + *
> + * In short, relocation wants to do:
> + *             Before          |               After
> + * ------------------------------------------------------------------
> + *  BG A: 10 data extents      | BG A: deleted
> + *  BG B:  2 data extents      | BG B: 10 data extents (2 old + 8 relocated)
> + *  BG C:  1 extents           | BG C:  3 data extents (1 old + 2 relocated)
> + *
> + * [How does relocation work]
> + * 1.   Mark the target bg RO
> + *      So that new extents won't be allocated from the target bg.
> + *
> + * 2.1  Record each extent in the target bg
> + *      To build a proper map of extents to be relocated.
> + *
> + * 2.2  Build data reloc tree and reloc trees
> + *      Data reloc tree will contain an inode, recording all newly relocated
> + *      data extents.
> + *      There will be only one data reloc tree for one data block group.
> + *
> + *      Reloc tree will be a special snapshot of its source tree, containing
> + *      relocated tree blocks.
> + *      Each tree referring to a tree block in target bg will get its reloc
> + *      tree built.
> + *
> + * 2.3  Swap source tree with its corresponding reloc tree
> + *      So that each involved tree only refers to new extents after swap.
> + *
> + * 3.   Cleanup reloc trees and data reloc tree.
> + *      As old extents in the target bg is still referred by reloc trees,
> + *      we need to clean them up before really freeing the target bg.
> + *
> + * The main complexity is in step 2.2 and 2.3.

step -> steps

Thanks.

> + *
> + * The core entrance for relocation is relocate_block_group() function.
> + */
>  /*
>   * backref_node, mapping_node and tree_block start with this
>   */
> --
> 2.24.1
>
Qu Wenruo Jan. 15, 2020, 11:11 a.m. UTC | #2
On 2020/1/15 下午6:34, Filipe Manana wrote:
> On Wed, Jan 15, 2020 at 6:29 AM Qu Wenruo <wqu@suse.com> wrote:
>>
>> Relocation is one of the most complex part of btrfs, while it's also the
>> foundation stone for online resizing, profile converting.
>>
>> For such a complex facility, we should at least have some introduction
>> to it.
>>
>> This patch will add an basic introduction at pretty a high level,
>> explaining:
>> - What relocation does
>> - How relocation is done
>>   Only mentioning how data reloc tree and reloc tree are involved in the
>>   operation.
>>   No details like the backref cache, or the data reloc tree contents.
>> - Which function to refer.
>>
>> More detailed comments will be added for reloc tree creation, data reloc
>> tree creation and backref cache.
>>
>> For now the introduction should save reader some time before digging
>> into the rabbit hole.
>>
>> Signed-off-by: Qu Wenruo <wqu@suse.com>
>> ---
>>  fs/btrfs/relocation.c | 44 +++++++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 44 insertions(+)
>>
>> diff --git a/fs/btrfs/relocation.c b/fs/btrfs/relocation.c
>> index d897a8e5e430..cd3a15f1716d 100644
>> --- a/fs/btrfs/relocation.c
>> +++ b/fs/btrfs/relocation.c
>> @@ -23,6 +23,50 @@
>>  #include "delalloc-space.h"
>>  #include "block-group.h"
>>
>> +/*
>> + * Introduction for btrfs relocation.
>> + *
>> + * [What does relocation do]
> 
> For readability, a blank line here would help.
> 
>> + * The objective of relocation is to relocate all or some extents of one block
>> + * group to other block groups.
> 
> Some? We always relocate all extents of a block group (except if
> errors happen of course).

Error is the exact reason I mention "some".

Thanks,
Qu

> 
>> + * This is utilized by resize (shrink only), profile converting, or just
>> + * balance routine to free some block groups.
>> + *
>> + * In short, relocation wants to do:
>> + *             Before          |               After
>> + * ------------------------------------------------------------------
>> + *  BG A: 10 data extents      | BG A: deleted
>> + *  BG B:  2 data extents      | BG B: 10 data extents (2 old + 8 relocated)
>> + *  BG C:  1 extents           | BG C:  3 data extents (1 old + 2 relocated)
>> + *
>> + * [How does relocation work]
>> + * 1.   Mark the target bg RO
>> + *      So that new extents won't be allocated from the target bg.
>> + *
>> + * 2.1  Record each extent in the target bg
>> + *      To build a proper map of extents to be relocated.
>> + *
>> + * 2.2  Build data reloc tree and reloc trees
>> + *      Data reloc tree will contain an inode, recording all newly relocated
>> + *      data extents.
>> + *      There will be only one data reloc tree for one data block group.
>> + *
>> + *      Reloc tree will be a special snapshot of its source tree, containing
>> + *      relocated tree blocks.
>> + *      Each tree referring to a tree block in target bg will get its reloc
>> + *      tree built.
>> + *
>> + * 2.3  Swap source tree with its corresponding reloc tree
>> + *      So that each involved tree only refers to new extents after swap.
>> + *
>> + * 3.   Cleanup reloc trees and data reloc tree.
>> + *      As old extents in the target bg is still referred by reloc trees,
>> + *      we need to clean them up before really freeing the target bg.
>> + *
>> + * The main complexity is in step 2.2 and 2.3.
> 
> step -> steps
> 
> Thanks.
> 
>> + *
>> + * The core entrance for relocation is relocate_block_group() function.
>> + */
>>  /*
>>   * backref_node, mapping_node and tree_block start with this
>>   */
>> --
>> 2.24.1
>>
> 
>
Filipe Manana Jan. 15, 2020, 11:25 a.m. UTC | #3
On Wed, Jan 15, 2020 at 11:11 AM Qu Wenruo <quwenruo.btrfs@gmx.com> wrote:
>
>
>
> On 2020/1/15 下午6:34, Filipe Manana wrote:
> > On Wed, Jan 15, 2020 at 6:29 AM Qu Wenruo <wqu@suse.com> wrote:
> >>
> >> Relocation is one of the most complex part of btrfs, while it's also the
> >> foundation stone for online resizing, profile converting.
> >>
> >> For such a complex facility, we should at least have some introduction
> >> to it.
> >>
> >> This patch will add an basic introduction at pretty a high level,
> >> explaining:
> >> - What relocation does
> >> - How relocation is done
> >>   Only mentioning how data reloc tree and reloc tree are involved in the
> >>   operation.
> >>   No details like the backref cache, or the data reloc tree contents.
> >> - Which function to refer.
> >>
> >> More detailed comments will be added for reloc tree creation, data reloc
> >> tree creation and backref cache.
> >>
> >> For now the introduction should save reader some time before digging
> >> into the rabbit hole.
> >>
> >> Signed-off-by: Qu Wenruo <wqu@suse.com>
> >> ---
> >>  fs/btrfs/relocation.c | 44 +++++++++++++++++++++++++++++++++++++++++++
> >>  1 file changed, 44 insertions(+)
> >>
> >> diff --git a/fs/btrfs/relocation.c b/fs/btrfs/relocation.c
> >> index d897a8e5e430..cd3a15f1716d 100644
> >> --- a/fs/btrfs/relocation.c
> >> +++ b/fs/btrfs/relocation.c
> >> @@ -23,6 +23,50 @@
> >>  #include "delalloc-space.h"
> >>  #include "block-group.h"
> >>
> >> +/*
> >> + * Introduction for btrfs relocation.
> >> + *
> >> + * [What does relocation do]
> >
> > For readability, a blank line here would help.
> >
> >> + * The objective of relocation is to relocate all or some extents of one block
> >> + * group to other block groups.
> >
> > Some? We always relocate all extents of a block group (except if
> > errors happen of course).
>
> Error is the exact reason I mention "some".

Then it should be rephrased imo. It's confusing. The objective is to
always relocate all extents.
You can mention instead in parentheses or in a separate statement that
in case of an error only some may be relocated:

The objective of relocation is to relocate all the extents of one
block group to other block groups.
If an error happens during relocation, it's possible that only some
extents were relocated.

Thanks.

>
> Thanks,
> Qu
>
> >
> >> + * This is utilized by resize (shrink only), profile converting, or just
> >> + * balance routine to free some block groups.
> >> + *
> >> + * In short, relocation wants to do:
> >> + *             Before          |               After
> >> + * ------------------------------------------------------------------
> >> + *  BG A: 10 data extents      | BG A: deleted
> >> + *  BG B:  2 data extents      | BG B: 10 data extents (2 old + 8 relocated)
> >> + *  BG C:  1 extents           | BG C:  3 data extents (1 old + 2 relocated)
> >> + *
> >> + * [How does relocation work]
> >> + * 1.   Mark the target bg RO
> >> + *      So that new extents won't be allocated from the target bg.
> >> + *
> >> + * 2.1  Record each extent in the target bg
> >> + *      To build a proper map of extents to be relocated.
> >> + *
> >> + * 2.2  Build data reloc tree and reloc trees
> >> + *      Data reloc tree will contain an inode, recording all newly relocated
> >> + *      data extents.
> >> + *      There will be only one data reloc tree for one data block group.
> >> + *
> >> + *      Reloc tree will be a special snapshot of its source tree, containing
> >> + *      relocated tree blocks.
> >> + *      Each tree referring to a tree block in target bg will get its reloc
> >> + *      tree built.
> >> + *
> >> + * 2.3  Swap source tree with its corresponding reloc tree
> >> + *      So that each involved tree only refers to new extents after swap.
> >> + *
> >> + * 3.   Cleanup reloc trees and data reloc tree.
> >> + *      As old extents in the target bg is still referred by reloc trees,
> >> + *      we need to clean them up before really freeing the target bg.
> >> + *
> >> + * The main complexity is in step 2.2 and 2.3.
> >
> > step -> steps
> >
> > Thanks.
> >
> >> + *
> >> + * The core entrance for relocation is relocate_block_group() function.
> >> + */
> >>  /*
> >>   * backref_node, mapping_node and tree_block start with this
> >>   */
> >> --
> >> 2.24.1
> >>
> >
> >
>
diff mbox series

Patch

diff --git a/fs/btrfs/relocation.c b/fs/btrfs/relocation.c
index d897a8e5e430..cd3a15f1716d 100644
--- a/fs/btrfs/relocation.c
+++ b/fs/btrfs/relocation.c
@@ -23,6 +23,50 @@ 
 #include "delalloc-space.h"
 #include "block-group.h"
 
+/*
+ * Introduction for btrfs relocation.
+ *
+ * [What does relocation do]
+ * The objective of relocation is to relocate all or some extents of one block
+ * group to other block groups.
+ * This is utilized by resize (shrink only), profile converting, or just
+ * balance routine to free some block groups.
+ *
+ * In short, relocation wants to do:
+ * 		Before		|		After
+ * ------------------------------------------------------------------
+ *  BG A: 10 data extents	| BG A: deleted
+ *  BG B:  2 data extents	| BG B: 10 data extents (2 old + 8 relocated)
+ *  BG C:  1 extents		| BG C:  3 data extents (1 old + 2 relocated)
+ *
+ * [How does relocation work]
+ * 1.   Mark the target bg RO
+ *      So that new extents won't be allocated from the target bg.
+ *
+ * 2.1  Record each extent in the target bg
+ *      To build a proper map of extents to be relocated.
+ *
+ * 2.2  Build data reloc tree and reloc trees
+ *      Data reloc tree will contain an inode, recording all newly relocated
+ *      data extents.
+ *      There will be only one data reloc tree for one data block group.
+ *
+ *      Reloc tree will be a special snapshot of its source tree, containing
+ *      relocated tree blocks.
+ *      Each tree referring to a tree block in target bg will get its reloc
+ *      tree built.
+ *
+ * 2.3  Swap source tree with its corresponding reloc tree
+ *      So that each involved tree only refers to new extents after swap.
+ *
+ * 3.   Cleanup reloc trees and data reloc tree.
+ *      As old extents in the target bg is still referred by reloc trees,
+ *      we need to clean them up before really freeing the target bg.
+ *
+ * The main complexity is in step 2.2 and 2.3.
+ *
+ * The core entrance for relocation is relocate_block_group() function.
+ */
 /*
  * backref_node, mapping_node and tree_block start with this
  */