Skip to contentSkip to conversation

btrfs: scrub: make scrub uses less memory for metadata scrub

PATCH Superseded Qu Wenruo 3 participants 5 messages Mar 2, 2022 – May 27, 2022 Btrfs file system

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Qu Wenruo

author
#0 lore raw

cover btrfs: scrub: make scrub uses less memory for metadata scrub

Although btrfs scrub works for subpage from day one, it has a small pitfall:

Scrub will always allocate a full page for each sector.

This causes increased memory usage, although not a big deal, it's still not ideal.

The patchset will change the behavior by integrating all pages into scrub_block::pages[], instead of using scrub_sector::page.

Now scrub_sector will no longer hold a page pointer, but uses its logical bytenr to caculate which page and page range it should use.

This behavior unfortunately still only affects memory usage on metadata scrub, which uses nodesize for scrub.

For the best case, 64K node size with 64K page size, we waste no memory to scrub one tree block.

For the worst case, 4K node size with 64K page size, we are no worse than the existing behavior (still one 64K page for the tree block)

For the default case (16K nodesize), we use one 64K page, compared to 4x64K pages previously.

For data scrubing, we uses sector size, thus it causes no difference. We need to do more work on data scrubing size to properly handle mutilpe sectors for non-RAID56 profiles.

The patchset requires the rename patchset.
(https://lore.kernel.org/linux-btrfs/[email protected]/)

If David is not happy with the big change again, at least first 3 patches can be considered as some cleanup.

Qu Wenruo (5):
btrfs: scrub: use pointer array to replace @sblocks_for_recheck
btrfs: extract the initialization of scrub_block into a helper
function
btrfs: extract the allocation and initialization of scrub_sector into
a helper
btrfs: scrub: introduce scrub_block::pages for more efficient memory
usage for subpage
btrfs: scrub: remove scrub_sector::page and use scrub_block::pages
instead

 fs/btrfs/scrub.c | 399 ++++++++++++++++++++++++++++++++---------------
 1 file changed, 270 insertions(+), 129 deletions(-)
Signature · 1 line
2.35.1
#1 lore raw
Quoted 36 lines · Qu Wenruo

Although btrfs scrub works for subpage from day one, it has a small pitfall:

Scrub will always allocate a full page for each sector.

This causes increased memory usage, although not a big deal, it's still not ideal.

The patchset will change the behavior by integrating all pages into scrub_block::pages[], instead of using scrub_sector::page.

Now scrub_sector will no longer hold a page pointer, but uses its logical bytenr to caculate which page and page range it should use.

This behavior unfortunately still only affects memory usage on metadata scrub, which uses nodesize for scrub.

For the best case, 64K node size with 64K page size, we waste no memory to scrub one tree block.

For the worst case, 4K node size with 64K page size, we are no worse than the existing behavior (still one 64K page for the tree block)

For the default case (16K nodesize), we use one 64K page, compared to 4x64K pages previously.

For data scrubing, we uses sector size, thus it causes no difference. We need to do more work on data scrubing size to properly handle mutilpe sectors for non-RAID56 profiles.

The patchset requires the rename patchset.
(https://lore.kernel.org/linux-btrfs/[email protected]/)

If David is not happy with the big change again, at least first 3 patches can be considered as some cleanup.

Quoted 10 lines

Qu Wenruo (5):
btrfs: scrub: use pointer array to replace @sblocks_for_recheck
btrfs: extract the initialization of scrub_block into a helper
function
btrfs: extract the allocation and initialization of scrub_sector into
a helper
btrfs: scrub: introduce scrub_block::pages for more efficient memory
usage for subpage
btrfs: scrub: remove scrub_sector::page and use scrub_block::pages
instead

Added to for-next as topic branch for now.

#2 lore raw
Quoted 48 lines · David Sterba
Quoted 36 lines · Qu Wenruo

Although btrfs scrub works for subpage from day one, it has a small pitfall:

Scrub will always allocate a full page for each sector.

This causes increased memory usage, although not a big deal, it's still not ideal.

The patchset will change the behavior by integrating all pages into scrub_block::pages[], instead of using scrub_sector::page.

Now scrub_sector will no longer hold a page pointer, but uses its logical bytenr to caculate which page and page range it should use.

This behavior unfortunately still only affects memory usage on metadata scrub, which uses nodesize for scrub.

For the best case, 64K node size with 64K page size, we waste no memory to scrub one tree block.

For the worst case, 4K node size with 64K page size, we are no worse than the existing behavior (still one 64K page for the tree block)

For the default case (16K nodesize), we use one 64K page, compared to 4x64K pages previously.

For data scrubing, we uses sector size, thus it causes no difference. We need to do more work on data scrubing size to properly handle mutilpe sectors for non-RAID56 profiles.

The patchset requires the rename patchset.
(https://lore.kernel.org/linux-btrfs/[email protected]/)

If David is not happy with the big change again, at least first 3 patches can be considered as some cleanup.

Quoted 10 lines

Qu Wenruo (5):
btrfs: scrub: use pointer array to replace @sblocks_for_recheck
btrfs: extract the initialization of scrub_block into a helper
function
btrfs: extract the allocation and initialization of scrub_sector into
a helper
btrfs: scrub: introduce scrub_block::pages for more efficient memory
usage for subpage
btrfs: scrub: remove scrub_sector::page and use scrub_block::pages
instead

Quoted 1 line

Added to for-next as topic branch for now.

I guess you replied to the wrong patch?

The for-next branch only contains the scrub entrance refactor v3.

No the renaming nor the subpage optimization.

Thanks,
Qu

#3 lore raw
Quoted 17 lines · Qu Wenruo
Quoted 17 lines
Quoted 16 lines

The patchset requires the rename patchset.
(https://lore.kernel.org/linux-btrfs/[email protected]/)

If David is not happy with the big change again, at least first 3 patches can be considered as some cleanup.

Qu Wenruo (5):
btrfs: scrub: use pointer array to replace @sblocks_for_recheck
btrfs: extract the initialization of scrub_block into a helper
function
btrfs: extract the allocation and initialization of scrub_sector into
a helper
btrfs: scrub: introduce scrub_block::pages for more efficient memory
usage for subpage
btrfs: scrub: remove scrub_sector::page and use scrub_block::pages
instead

Quoted 3 lines
Quoted 1 line

Added to for-next as topic branch for now.

I guess you replied to the wrong patch?

Yeah.

Quoted 3 lines

The for-next branch only contains the scrub entrance refactor v3.

No the renaming nor the subpage optimization.

The scrub renaming is in misc-next, please refresh this patchset, thanks.

#4 lore raw
In reply to Qu Wenruo
Quoted 36 lines · Qu Wenruo

Although btrfs scrub works for subpage from day one, it has a small pitfall:

Scrub will always allocate a full page for each sector.

This causes increased memory usage, although not a big deal, it's still not ideal.

The patchset will change the behavior by integrating all pages into scrub_block::pages[], instead of using scrub_sector::page.

Now scrub_sector will no longer hold a page pointer, but uses its logical bytenr to caculate which page and page range it should use.

This behavior unfortunately still only affects memory usage on metadata scrub, which uses nodesize for scrub.

For the best case, 64K node size with 64K page size, we waste no memory to scrub one tree block.

For the worst case, 4K node size with 64K page size, we are no worse than the existing behavior (still one 64K page for the tree block)

For the default case (16K nodesize), we use one 64K page, compared to 4x64K pages previously.

For data scrubing, we uses sector size, thus it causes no difference. We need to do more work on data scrubing size to properly handle mutilpe sectors for non-RAID56 profiles.

The patchset requires the rename patchset.
(https://lore.kernel.org/linux-btrfs/[email protected]/)

If David is not happy with the big change again, at least first 3 patches can be considered as some cleanup.

Quoted 10 lines

Qu Wenruo (5):
btrfs: scrub: use pointer array to replace @sblocks_for_recheck
btrfs: extract the initialization of scrub_block into a helper
function
btrfs: extract the allocation and initialization of scrub_sector into
a helper
btrfs: scrub: introduce scrub_block::pages for more efficient memory
usage for subpage
btrfs: scrub: remove scrub_sector::page and use scrub_block::pages
instead

This patchset still seems relevant, but does not apply after recent changes. Please refresh and resend, thanks.