Skip to contentSkip to conversation

generic/064: allow 50 extents on F2FS after fcollapse

PATCH In review Jan Prusakowski 4 participants 11 messages Jun 22 – Jul 22

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 generic/064: allow 50 extents on F2FS after fcollapse

On F2FS, generic/064 fails with "extents mismatched before = 1 after = 50" following multiple fcollapse (collapse range) operations.

To ensure crash consistency and checkpoint integrity, F2FS forbids in-place SSR (Summary Standalone Replacement) overwrites on valid checkpointed blocks. When collapse range shifts blocks, F2FS allocates new data pages in LFS mode (out-of-place log writes). As a result, sequential collapse range calls rewrite shifted blocks at new log locations, intentionally leaving the file with 50 extents.

Adjust the extent verification in generic/064 to expect exactly 50 extents on F2FS, while preserving the strict 1-extent requirement for all other filesystems. Data integrity continues to be verified via byte comparison.

---
 tests/generic/064 | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
tests/generic/064 +7 -1
index aed13970..b77eccb4 100755
@@ -59,7 +59,13 @@ for (( j=0; j < $(($BLOCKS/2)); j++ )); do
 done
 
 extent_after=`_count_extents $dest`
-if [ $extent_before -ne $extent_after ]; then
+
+if [ "$FSTYP" == "f2fs" ];
+then
+	if [ $extent_before -ne 1 ] || [ $extent_after -ne 50 ]; then
+		echo "extents mismatched before = $extent_before after = $extent_after"
+	fi
+elif [ $extent_before -ne $extent_after ]; then
 	echo "extents mismatched before = $extent_before after = $extent_after"
 fi
 
Signature · 1 line
2.55.0.rc0.786.g65d90a0328-goog
Reviewed-by on v1 1/1
#1 lore raw
Quoted 10 lines · On 6/22/26 15:04, Jan Prusakowski via Linux-f2fs-devel wrote:

On F2FS, generic/064 fails with "extents mismatched before = 1 after = 50" following multiple fcollapse (collapse range) operations.

To ensure crash consistency and checkpoint integrity, F2FS forbids in-place SSR (Summary Standalone Replacement) overwrites on valid checkpointed blocks. When collapse range shifts blocks, F2FS allocates new data pages in LFS mode (out-of-place log writes). As a result, sequential collapse range calls rewrite shifted blocks at new log locations, intentionally leaving the file with 50 extents.

Quoted 7 lines

Adjust the extent verification in generic/064 to expect exactly 50 extents on F2FS, while preserving the strict 1-extent requirement for all other filesystems. Data integrity continues to be verified via byte comparison.

Thanks,

#2 lore raw
In reply to Jan Prusakowski
Quoted 9 lines · Jan Prusakowski

On F2FS, generic/064 fails with "extents mismatched before = 1 after = 50" following multiple fcollapse (collapse range) operations.

To ensure crash consistency and checkpoint integrity, F2FS forbids in-place SSR (Summary Standalone Replacement) overwrites on valid checkpointed blocks. When collapse range shifts blocks, F2FS allocates new data pages in LFS mode (out-of-place log writes). As a result, sequential collapse range calls rewrite shifted blocks at new log locations, intentionally leaving the file with 50 extents.

This sounds odd. The test allocates a contigous range and then just does insert/collapse on it, which should not lead to any new data block allocations. Given that the test works fine on zoned XFS and btrfs with strict out of place write policies we know it does not require overwriting blocks to work as well.

So I think something is fishy in f2fs if needs to allocate data blocks here.

#3 lore raw
Quoted 11 lines · Christoph Hellwig
Quoted 9 lines · Jan Prusakowski

On F2FS, generic/064 fails with "extents mismatched before = 1 after = 50" following multiple fcollapse (collapse range) operations.

To ensure crash consistency and checkpoint integrity, F2FS forbids in-place SSR (Summary Standalone Replacement) overwrites on valid checkpointed blocks. When collapse range shifts blocks, F2FS allocates new data pages in LFS mode (out-of-place log writes). As a result, sequential collapse range calls rewrite shifted blocks at new log locations, intentionally leaving the file with 50 extents.

Quoted 8 lines

This sounds odd. The test allocates a contigous range and then just does insert/collapse on it, which should not lead to any new data block allocations. Given that the test works fine on zoned XFS and btrfs with strict out of place write policies we know it does not require overwriting blocks to work as well.

So I think something is fishy in f2fs if needs to allocate data blocks here.

I have the same question with Christoph. Shouldn't fcollapse in F2FS just remap the metadata instead of allocating new data blocks?

Quoted 1 line
#4 lore raw
Quoted 7 lines · On 6/30/26 00:37, Zorro Lang via Linux-f2fs-devel wrote:
Quoted 6 lines · Christoph Hellwig

This sounds odd. The test allocates a contigous range and then just does insert/collapse on it, which should not lead to any new data block allocations. Given that the test works fine on zoned XFS and btrfs with strict out of place write policies we know it does not require overwriting blocks to work as well.

Quoted 5 lines
Quoted 2 lines

So I think something is fishy in f2fs if needs to allocate data blocks here.

I have the same question with Christoph. Shouldn't fcollapse in F2FS just remap the metadata instead of allocating new data blocks?

Hi Christoph, Zorro,

Let me explain it briefly, this is because f2fs has only one copy of reverse mapping table (aka SSA), for any checkpointed data, we can not modify SSA info directly during remap, otherwise, after sudden power-cut, we will suffer inconsistent issue in between reverse mapping table and mapping table. So, f2fs choose to rewrite for any checkpointed data.

Thanks,

#5 lore raw
Quoted 14 lines · Chao Yu
Quoted 7 lines · On 6/30/26 00:37, Zorro Lang via Linux-f2fs-devel wrote:
Quoted 6 lines · Christoph Hellwig

This sounds odd. The test allocates a contigous range and then just does insert/collapse on it, which should not lead to any new data block allocations. Given that the test works fine on zoned XFS and btrfs with strict out of place write policies we know it does not require overwriting blocks to work as well.

Quoted 5 lines
Quoted 2 lines

So I think something is fishy in f2fs if needs to allocate data blocks here.

I have the same question with Christoph. Shouldn't fcollapse in F2FS just remap the metadata instead of allocating new data blocks?

Quoted 7 lines

Hi Christoph, Zorro,

Let me explain it briefly, this is because f2fs has only one copy of reverse mapping table (aka SSA), for any checkpointed data, we can not modify SSA info directly during remap, otherwise, after sudden power-cut, we will suffer inconsistent issue in between reverse mapping table and mapping table. So, f2fs choose to rewrite for any checkpointed data.

Hi Chao,

Thank you for the detailed explanation. I now see why F2FS handles this differently.

Since this is by design for F2FS, I'd like to merge this fix to help f2fs avoid this test failure at first. As for the potential write amplification and performance overhead introduced by this rewrite strategy, that can be discussed further on the f2fs list if necessary.

I will merge the patch with some clarifying comments:

/*
* F2FS uses a rewrite strategy for fcollapse/finsert on checkpointed data.
* Since F2FS has only a single copy of the reverse mapping table (SSA),
* directly remapping checkpointed metadata could cause inconsistency
* issues after a sudden power-cut.
*/

Is this ^^ good to you, or feel free to give me a proper comment to add into this patch.

Thanks,
Zorro

Quoted 2 lines

Thanks,

#6 lore raw

Hi Zorro,

Quoted 25 lines · Zorro Lang
Quoted 14 lines · Chao Yu
Quoted 7 lines · On 6/30/26 00:37, Zorro Lang via Linux-f2fs-devel wrote:
Quoted 6 lines · Christoph Hellwig

This sounds odd. The test allocates a contigous range and then just does insert/collapse on it, which should not lead to any new data block allocations. Given that the test works fine on zoned XFS and btrfs with strict out of place write policies we know it does not require overwriting blocks to work as well.

Quoted 5 lines
Quoted 2 lines

So I think something is fishy in f2fs if needs to allocate data blocks here.

I have the same question with Christoph. Shouldn't fcollapse in F2FS just remap the metadata instead of allocating new data blocks?

Quoted 7 lines

Hi Christoph, Zorro,

Let me explain it briefly, this is because f2fs has only one copy of reverse mapping table (aka SSA), for any checkpointed data, we can not modify SSA info directly during remap, otherwise, after sudden power-cut, we will suffer inconsistent issue in between reverse mapping table and mapping table. So, f2fs choose to rewrite for any checkpointed data.

Hi Chao,

Quoted 7 lines

Thank you for the detailed explanation. I now see why F2FS handles this differently.

Since this is by design for F2FS, I'd like to merge this fix to help f2fs avoid this test failure at first. As for the potential write amplification and performance overhead introduced by this rewrite strategy, that can be discussed further on the f2fs list if necessary.

It makes sense to me, we can discuss about the detailed implementation in f2fs list if there is any concern.

Quoted 10 lines

I will merge the patch with some clarifying comments:

/*
* F2FS uses a rewrite strategy for fcollapse/finsert on checkpointed data.
* Since F2FS has only a single copy of the reverse mapping table (SSA),
* directly remapping checkpointed metadata could cause inconsistency
* issues after a sudden power-cut.
*/

Quoted 2 lines

Is this ^^ good to you, or feel free to give me a proper comment to add into this patch.

Looks good, thanks for the help!

Thanks

Quoted 6 lines

Thanks,
Zorro

Quoted 2 lines

Thanks,

#7 lore raw
In reply to Chao Yu
Quoted 6 lines · Chao Yu

Let me explain it briefly, this is because f2fs has only one copy of reverse mapping table (aka SSA), for any checkpointed data, we can not modify SSA info directly during remap, otherwise, after sudden power-cut, we will suffer inconsistent issue in between reverse mapping table and mapping table. So, f2fs choose to rewrite for any checkpointed data.

So looking at this explanation and the code, f2fs does not actually support COLLAPSE_RANGE at all, but emulates by doing a full data copy. What is the point of this and lying to the user that it supports and operation that isn't actually supported as the intended metadata only swizzle?

#8 lore raw

Hi Christoph,

Quoted 2 lines · Christoph Hellwig

So looking at this explanation and the code, f2fs does not actually support COLLAPSE_RANGE at all, but emulates by doing a full data

For non-checkpoint data, we can just update metadata w/o any data copying to support COLLAPSE_RANGE.

Quoted 3 lines

copy. What is the point of this and lying to the user that it supports and operation that isn't actually supported as the intended metadata only swizzle?

I added the fcollapse support long time ago in commit b4ace3370324 ("f2fs: support FALLOC_FL_COLLAPSE_RANGE"), and then, a following patch was committed to fix checkpoint case by copying data.

IIRC, the intention at that time is trying to support modern filesystem features as many as possible, so while commercializing f2fs filesystem, it won't break userspace application use if previous filesystem support the feature.

I just followed manual of fallocate 2 to implement the fcollapse, I didn't notice fcollapse is a metadata operation only interface, but I may missed something in manual or discussion conclusion about fcollapse in community, please let me know if I missed something. :)

Thanks,

#9 lore raw
Quoted 8 lines · Chao Yu

Hi Christoph,

Quoted 2 lines · Christoph Hellwig

So looking at this explanation and the code, f2fs does not actually support COLLAPSE_RANGE at all, but emulates by doing a full data

For non-checkpoint data, we can just update metadata w/o any data copying to support COLLAPSE_RANGE.

Can you clarify what "non-checkpoint data" data means? Older data?

#10 lore raw
Quoted 1 line · Christoph Hellwig

Can you clarify what "non-checkpoint data" data means? Older data?

It is newer data.

1. write data first, and then trigger checkpoint, after checkpoint, data will become checkpointed one, for such case we can only collapse w/ data copying.
2. write data (data is non-checkpointed if there is no following checkpoint), then we can collapse w/ metadata update.