ext4: fix partial cluster initialization when splitting extent
RFC PATCH 1/1 ext4: fix partial cluster initialization when splitting extent
Hi Eric, would you mind explaining why the magic number '2' is used here when calculating the physical cluster number of the partial cluster in commit f4226d9ea400 ("ext4: fix partial cluster initialization") ?
```
+ /*
+ * If we're going to split the extent, note that
+ * the cluster containing the block after 'end' is
+ * in use to avoid freeing it when removing blocks.
+ */
+ if (sbi->s_cluster_ratio > 1) {
+ pblk = ext4_ext_pblock(ex) + end - ee_block + 2;
+ partial_cluster =
+ -(long long) EXT4_B2C(sbi, pblk);
+ }
```As far as I understand, there we are initializing the partial cluster describing the beginning of the split extent after 'end'. The corrsponding physical block number of the first block in the split extent should be 'ext4_ext_pblock(ex) + end - ee_block + 1'.
This bug will cause xfstests shared/298 failure on ext4 with bigalloc enabled sometimes. Ext4 error messages indicate that previously freed blocks are being freed again, and the following fsck will fail due to the inconsistency of block bitmap and bg descriptor.
The following is an example case:
1. First, Initialize a ext4 filesystem with cluster size '16K', block size '4K', in which case, one cluster contains four blocks.
2. Create one file (e.g., xxx.img) on this ext4 filesystem. Now the extent tree of this file is like:
...
36864:[0]4:220160
36868:[0]14332:145408
51200:[0]2:231424
...
3. Then execute PUNCH_HOLE fallocate on this file. The hole range is like:
..ext4_ext_remove_space: dev 254,16 ino 12 since 49506 end 49506 depth 1ext4_ext_remove_space: dev 254,16 ino 12 since 49544 end 49546 depth 1ext4_ext_remove_space: dev 254,16 ino 12 since 49605 end 49607 depth 1
...
4. Then the extent tree of this file after punching is like
...
49507:[0]37:158047
49547:[0]58:158087
...
5. Detailed procedure of punching hole [49544, 49546]
5.1. The block address space:
``` lblk ~49505 49506 49507~49543 49544~49546 49547~ ---------+------+-------------+----------------+-------- extent | hole | extent | hole | extent ---------+------+-------------+----------------+-------- pblk ~158045 158046 158047~158083 158084~158086 158087~ ```
5.2. The detailed layout of cluster 39521:
``` cluster 39521 <-------------------------------> hole extent <----------------------><-------- lblk 49544 49545 49546 49547 +-------+-------+-------+-------+ | | | | | +-------+-------+-------+-------+ pblk 158084 1580845 158086 158087 ```
5.3. The ftrace output when punching hole [49544, 49546]:
- ext4_ext_remove_space (start 49544, end 49546)
- ext4_ext_rm_leaf (start 49544, end 49546, last_extent [49507(158047), 40], partial [pclu 39522 lblk 0 state 2])
- ext4_remove_blocks (extent [49507(158047), 40], from 49544 to 49546, partial [pclu 39522 lblk 0 state 2]
- ext4_free_blocks: (block 158084 count 4)
- ext4_mballoc_free (extent 1/6753/1)
In this case, the whole cluster 39521 is freed mistakenly when freeing pblock 158084~158086 (i.e., the first three blocks of this cluster), although pblock 158087 (the last remaining block of this cluster) has not been freed yet.
The root cause of this isuue is that, the pclu of the partial cluster is calculated mistakenly in ext4_ext_remove_space(). The correct partial_cluster.pclu (i.e., the cluster number of the first block in the next extent, that is, lblock 49597 (pblock 158086)) should be 39521 rather than 39522.
--- fs/ext4/extents.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
fs/ext4/extents.c +1 -1
@@ -2828,7 +2828,7 @@ int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,* in use to avoid freeing it when removing blocks. */ if (sbi->s_cluster_ratio >1 ) { - pblk =ext4_ext_pblock (ex) + end - ee_block +2 ; + pblk =ext4_ext_pblock (ex) + end - ee_block +1 ; partial.pclu =EXT4_B2C (sbi, pblk); partial.state = nofree; }
Signature · 1 line
1.8.3.1
* Jeffle Xu <[email protected]>:
Quoted 123 lines
Hi Eric, would you mind explaining why the magic number '2' is used here when calculating the physical cluster number of the partial cluster in commit
f4226d9ea400("ext4: fix partial cluster initialization") ?```
(quoted diff) +10
-0+ /* + * If we're going to split the extent, note that + * the cluster containing the block after 'end' is + * in use to avoid freeing it when removing blocks. + */ + if (sbi->s_cluster_ratio > 1) { + pblk = ext4_ext_pblock(ex) + end - ee_block + 2; + partial_cluster = + -(long long) EXT4_B2C(sbi, pblk); + }``` As far as I understand, there we are initializing the partial cluster describing the beginning of the split extent after 'end'. The corrsponding physical block number of the first block in the split extent should be 'ext4_ext_pblock(ex) + end - ee_block + 1'. This bug will cause xfstests shared/298 failure on ext4 with bigalloc enabled sometimes. Ext4 error messages indicate that previously freed blocks are being freed again, and the following fsck will fail due to the inconsistency of block bitmap and bg descriptor. The following is an example case: 1. First, Initialize a ext4 filesystem with cluster size '16K', block size '4K', in which case, one cluster contains four blocks. 2. Create one file (e.g., xxx.img) on this ext4 filesystem. Now the extent tree of this file is like: ... 36864:[0]4:220160 36868:[0]14332:145408 51200:[0]2:231424 ... 3. Then execute PUNCH_HOLE fallocate on this file. The hole range is like: .. ext4_ext_remove_space: dev 254,16 ino 12 since 49506 end 49506 depth 1 ext4_ext_remove_space: dev 254,16 ino 12 since 49544 end 49546 depth 1 ext4_ext_remove_space: dev 254,16 ino 12 since 49605 end 49607 depth 1 ... 4. Then the extent tree of this file after punching is like ... 49507:[0]37:158047 49547:[0]58:158087 ... 5. Detailed procedure of punching hole [49544, 49546] 5.1. The block address space: ``` lblk ~49505 49506 49507~49543 49544~49546 49547~ ---------+------+-------------+----------------+-------- extent | hole | extent | hole | extent ---------+------+-------------+----------------+-------- pblk ~158045 158046 158047~158083 158084~158086 158087~ ``` 5.2. The detailed layout of cluster 39521: ``` cluster 39521 <-------------------------------> hole extent <----------------------><--------lblk 49544 49545 49546 49547
+-------+-------+-------+-------+ | | | | | +-------+-------+-------+-------+ pblk 158084 1580845 158086 158087```
5.3. The ftrace output when punching hole [49544, 49546]:
-ext4_ext_remove_space(start 49544, end 49546)
-ext4_ext_rm_leaf(start 49544, end 49546,last_extent[49507(158047), 40], partial [pclu 39522 lblk 0 state 2])
-ext4_remove_blocks(extent [49507(158047), 40], from 49544 to 49546, partial [pclu 39522 lblk 0 state 2]
-ext4_free_blocks: (block 158084 count 4)
-ext4_mballoc_free(extent 1/6753/1)In this case, the whole cluster 39521 is freed mistakenly when freeing pblock 158084~158086 (i.e., the first three blocks of this cluster), although pblock 158087 (the last remaining block of this cluster) has not been freed yet.
The root cause of this isuue is that, the pclu of the partial cluster is calculated mistakenly in
ext4_ext_remove_space(). The correct partial_cluster.pclu (i.e., the cluster number of the first block in the next extent, that is, lblock 49597 (pblock 158086)) should be 39521 rather than 39522.Fixes f4226d9ea400 ("ext4: fix partial cluster initialization")Signed-off-by Jeffle Xu <[email protected]>--- fs/ext4/extents.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)fs/ext4/extents.c +1
-1@@ -2828,7 +2828,7 @@ int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,* in use to avoid freeing it when removing blocks. */ if (sbi->s_cluster_ratio >1 ) { - pblk =ext4_ext_pblock (ex) + end - ee_block +2 ; + pblk =ext4_ext_pblock (ex) + end - ee_block +1 ; partial.pclu =EXT4_B2C (sbi, pblk); partial.state = nofree; }Signature · 1 line
1.8.3.1
Hi, Jeffle:
Thanks for the report and the patch. At first glance I suspect the "2" is simply a bug; logically we're just looking for the first block after the extent split to set the partial cluster, as you suggest. I'll post the results of my review once I've had a chance to refresh my memory of the code and run some more tests.
Thanks,
Eric
* Eric Whitney <[email protected]>:
Quoted 127 lines
* Jeffle Xu <[email protected]>:
Quoted 123 lines
Hi Eric, would you mind explaining why the magic number '2' is used here when calculating the physical cluster number of the partial cluster in commit
f4226d9ea400("ext4: fix partial cluster initialization") ?```
(quoted diff) +10
-0+ /* + * If we're going to split the extent, note that + * the cluster containing the block after 'end' is + * in use to avoid freeing it when removing blocks. + */ + if (sbi->s_cluster_ratio > 1) { + pblk = ext4_ext_pblock(ex) + end - ee_block + 2; + partial_cluster = + -(long long) EXT4_B2C(sbi, pblk); + }``` As far as I understand, there we are initializing the partial cluster describing the beginning of the split extent after 'end'. The corrsponding physical block number of the first block in the split extent should be 'ext4_ext_pblock(ex) + end - ee_block + 1'. This bug will cause xfstests shared/298 failure on ext4 with bigalloc enabled sometimes. Ext4 error messages indicate that previously freed blocks are being freed again, and the following fsck will fail due to the inconsistency of block bitmap and bg descriptor. The following is an example case: 1. First, Initialize a ext4 filesystem with cluster size '16K', block size '4K', in which case, one cluster contains four blocks. 2. Create one file (e.g., xxx.img) on this ext4 filesystem. Now the extent tree of this file is like: ... 36864:[0]4:220160 36868:[0]14332:145408 51200:[0]2:231424 ... 3. Then execute PUNCH_HOLE fallocate on this file. The hole range is like: .. ext4_ext_remove_space: dev 254,16 ino 12 since 49506 end 49506 depth 1 ext4_ext_remove_space: dev 254,16 ino 12 since 49544 end 49546 depth 1 ext4_ext_remove_space: dev 254,16 ino 12 since 49605 end 49607 depth 1 ... 4. Then the extent tree of this file after punching is like ... 49507:[0]37:158047 49547:[0]58:158087 ... 5. Detailed procedure of punching hole [49544, 49546] 5.1. The block address space: ``` lblk ~49505 49506 49507~49543 49544~49546 49547~ ---------+------+-------------+----------------+-------- extent | hole | extent | hole | extent ---------+------+-------------+----------------+-------- pblk ~158045 158046 158047~158083 158084~158086 158087~ ``` 5.2. The detailed layout of cluster 39521: ``` cluster 39521 <-------------------------------> hole extent <----------------------><--------lblk 49544 49545 49546 49547
+-------+-------+-------+-------+ | | | | | +-------+-------+-------+-------+ pblk 158084 1580845 158086 158087```
5.3. The ftrace output when punching hole [49544, 49546]:
-ext4_ext_remove_space(start 49544, end 49546)
-ext4_ext_rm_leaf(start 49544, end 49546,last_extent[49507(158047), 40], partial [pclu 39522 lblk 0 state 2])
-ext4_remove_blocks(extent [49507(158047), 40], from 49544 to 49546, partial [pclu 39522 lblk 0 state 2]
-ext4_free_blocks: (block 158084 count 4)
-ext4_mballoc_free(extent 1/6753/1)In this case, the whole cluster 39521 is freed mistakenly when freeing pblock 158084~158086 (i.e., the first three blocks of this cluster), although pblock 158087 (the last remaining block of this cluster) has not been freed yet.
The root cause of this isuue is that, the pclu of the partial cluster is calculated mistakenly in
ext4_ext_remove_space(). The correct partial_cluster.pclu (i.e., the cluster number of the first block in the next extent, that is, lblock 49597 (pblock 158086)) should be 39521 rather than 39522.Fixes f4226d9ea400 ("ext4: fix partial cluster initialization")Signed-off-by Jeffle Xu <[email protected]>--- fs/ext4/extents.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)fs/ext4/extents.c +1
-1@@ -2828,7 +2828,7 @@ int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,* in use to avoid freeing it when removing blocks. */ if (sbi->s_cluster_ratio >1 ) { - pblk =ext4_ext_pblock (ex) + end - ee_block +2 ; + pblk =ext4_ext_pblock (ex) + end - ee_block +1 ; partial.pclu =EXT4_B2C (sbi, pblk); partial.state = nofree; }Signature · 1 line
1.8.3.1Hi, Jeffle:
Quoted 8 lines
Thanks for the report and the patch. At first glance I suspect the "2" is simply a bug; logically we're just looking for the first block after the extent split to set the partial cluster, as you suggest. I'll post the results of my review once I've had a chance to refresh my memory of the code and run some more tests.
Thanks,
Eric
Hi, Jeffle:
What kernel were you running when you observed your failures? Does your patch resolve all observed failures, or do any remain? Do you have a simple test script that reproduces the bug?
I've made almost 1000 runs of shared/298 on various bigalloc configurations using Ted's test appliance on 5.7-rc5 and have not observed a failure. Several auto group runs have also passed without failures. Ideally, I'd like to be able to reproduce your failure to be sure we fully understand what's going on. It's still the case that the "2" is wrong, but I think that code in rm_leaf may be involved in an unexpected way.
Thanks,
Eric
Quoted 13 lines · Eric Whitney
Hi, Jeffle:
What kernel were you running when you observed your failures? Does your patch resolve all observed failures, or do any remain? Do you have a simple test script that reproduces the bug?
I've made almost 1000 runs of shared/298 on various bigalloc configurations using Ted's test appliance on 5.7-rc5 and have not observed a failure. Several auto group runs have also passed without failures. Ideally, I'd like to be able to reproduce your failure to be sure we fully understand what's going on. It's still the case that the "2" is wrong, but I think that code in
rm_leafmay be involved in an unexpected way.
Quoted 2 lines
Thanks,
Eric
Hi Eric,
Following on is my test environment.
kernel: 5.7-rc4-git-eb24fdd8e6f5c6bb95129748a1801c6476492aba
e2fsprog: latest release version 1.45.6 (20-Mar-2020)
xfstests: git://git.kernel.org/pub/scm/fs/xfs/xfstests-dev.git, master branch, latest commit
1. Test device
I run the test in a VM and the VM is setup by qemu. The size of vdb is 1G,
``` #lsblk NAME MAJ:MIN RM SIZE RO TYPE MOUNTPOINT vdb 254:16 0 1G 0 disk ```
and is initialized by:
``` qemu-img create -f qcow2 /XX/disk1.qcow2 1G qemu-kvm -drive file=/XX/disk1.qcow2,if=virtio,format=qcow2 ... ```
2. Test script
local.config of xfstests is like:
export TEST_DEV=/dev/vdb
export TEST_DIR=/mnt/test
export SCRATCH_DEV=/dev/vdc
export SCRATCH_MNT=/mnt/scratch
Following on is an example script to reproduce the failure:
```sh #!/bin/bash for i in `seq 100`; do echo y | mkfs.ext4 -O bigalloc -C 16K /dev/vdb ./check shared/298 status=$? if [[ $status == 1 ]]; then echo "$i exit" exit fi done ```
Indeed the failure occurs occasionally. Sometimes the script stops at iteration 4, or sometimes
at iteration 2, 7, 24.
The failure occurs with the following dmesg report:
``` [ 387.471876] EXT4-fs error (device vdb): mb_free_blocks:1457: group 1, block 158084:freeing already freed block (bit 6753); block bitmap corrupt. [ 387.473729] EXT4-fs error (device vdb): ext4_mb_generate_buddy:747: group 1, block bitmap and bg descriptor inconsistent: 19550 vs 19551 free clusters ```
3. About the applied patch
The applied patch does fix the failure in my test environment. At least the failure doesn't occur after running the full 100 iterations.
Thanks
Jeffle
Eric Whitney
Reviewed-by* JeffleXu <[email protected]>:
Quoted 108 lines · Eric Whitney
Quoted 13 lines
Hi, Jeffle:
What kernel were you running when you observed your failures? Does your patch resolve all observed failures, or do any remain? Do you have a simple test script that reproduces the bug?
I've made almost 1000 runs of shared/298 on various bigalloc configurations using Ted's test appliance on 5.7-rc5 and have not observed a failure. Several auto group runs have also passed without failures. Ideally, I'd like to be able to reproduce your failure to be sure we fully understand what's going on. It's still the case that the "2" is wrong, but I think that code in
rm_leafmay be involved in an unexpected way.Quoted 2 lines
Thanks,
EricHi Eric,
Following on is my test environment.
kernel: 5.7-rc4-git-eb24fdd8e6f5c6bb95129748a1801c6476492aba
e2fsprog: latest release version 1.45.6 (20-Mar-2020)
xfstests:
git://git.kernel.org/pub/scm/fs/xfs/xfstests-dev.git, master branch, latest commit1. Test device
I run the test in a VM and the VM is setup by qemu. The size of vdb is 1G,
``` #lsblk NAME MAJ:MIN RM SIZE RO TYPE MOUNTPOINT vdb 254:16 0 1G 0 disk ```and is initialized by:
``` qemu-img create -f qcow2 /XX/disk1.qcow2 1G qemu-kvm -drive file=/XX/disk1.qcow2,if=virtio,format=qcow2 ... ```2. Test script
local.config of xfstests is like:
export TEST_DEV=/dev/vdb
export TEST_DIR=/mnt/test
export SCRATCH_DEV=/dev/vdc
export SCRATCH_MNT=/mnt/scratchFollowing on is an example script to reproduce the failure:
```sh #!/bin/bash for i in `seq 100`; do echo y | mkfs.ext4 -O bigalloc -C 16K /dev/vdb ./check shared/298 status=$? if [[ $status == 1 ]]; then echo "$i exit" exit fi done ```Indeed the failure occurs occasionally. Sometimes the script stops at iteration 4, or sometimes
at iteration 2, 7, 24.
The failure occurs with the following dmesg report:
``` [ 387.471876] EXT4-fs error (device vdb): mb_free_blocks:1457: group 1, block 158084:freeing already freed block (bit 6753); block bitmap corrupt. [ 387.473729] EXT4-fs error (device vdb): ext4_mb_generate_buddy:747: group 1, block bitmap and bg descriptor inconsistent: 19550 vs 19551 free clusters ```3. About the applied patch
Quoted 10 lines
The applied patch does fix the failure in my test environment. At least the failure doesn't occur after running the full 100 iterations.
Thanks
Jeffle
Hi, Jeffle:
Thanks for that information. I'm still unable to reproduce your failure, but by inspection your patch clearly fixes a bug, and of course, you're seeing that. I suspect the code in rm_leaf that also sets the partial cluster nofree state is masking the bug in my testing. In your case, my best guess is that your testing is occasionally getting into the retry loop for EAGAIN in remove_space. This would effectively expose the bug again and could lead to the failure you've described.
Your patch has survived all the heavy testing I've thrown at it. So, please repost your RFC patch as a fix, and feel free to add:
This points out that the cluster freeing code really needs to be cleaned up, so I'm working on a patch series that does that.
Thanks for your patience,
Eric
Thanks for reviewing. I will send a formal patch later ;)
Thanks,
Jeffle
Quoted 136 lines · Eric Whitney
* JeffleXu <[email protected]>:
Quoted 106 lines · Eric Whitney
Quoted 13 lines
Hi, Jeffle:
What kernel were you running when you observed your failures? Does your patch resolve all observed failures, or do any remain? Do you have a simple test script that reproduces the bug?
I've made almost 1000 runs of shared/298 on various bigalloc configurations using Ted's test appliance on 5.7-rc5 and have not observed a failure. Several auto group runs have also passed without failures. Ideally, I'd like to be able to reproduce your failure to be sure we fully understand what's going on. It's still the case that the "2" is wrong, but I think that code in
rm_leafmay be involved in an unexpected way.Quoted 2 lines
Thanks,
EricHi Eric,
Following on is my test environment.
kernel: 5.7-rc4-git-eb24fdd8e6f5c6bb95129748a1801c6476492aba
e2fsprog: latest release version 1.45.6 (20-Mar-2020)
xfstests:
git://git.kernel.org/pub/scm/fs/xfs/xfstests-dev.git, master branch, latest commit1. Test device
I run the test in a VM and the VM is setup by qemu. The size of vdb is 1G,
``` #lsblk NAME MAJ:MIN RM SIZE RO TYPE MOUNTPOINT vdb 254:16 0 1G 0 disk ```and is initialized by:
``` qemu-img create -f qcow2 /XX/disk1.qcow2 1G qemu-kvm -drive file=/XX/disk1.qcow2,if=virtio,format=qcow2 ... ```2. Test script
local.config of xfstests is like:
export TEST_DEV=/dev/vdb
export TEST_DIR=/mnt/test
export SCRATCH_DEV=/dev/vdc
export SCRATCH_MNT=/mnt/scratchFollowing on is an example script to reproduce the failure:
```sh #!/bin/bash for i in `seq 100`; do echo y | mkfs.ext4 -O bigalloc -C 16K /dev/vdb ./check shared/298 status=$? if [[ $status == 1 ]]; then echo "$i exit" exit fi done ```Indeed the failure occurs occasionally. Sometimes the script stops at iteration 4, or sometimes
at iteration 2, 7, 24.
The failure occurs with the following dmesg report:
``` [ 387.471876] EXT4-fs error (device vdb): mb_free_blocks:1457: group 1, block 158084:freeing already freed block (bit 6753); block bitmap corrupt. [ 387.473729] EXT4-fs error (device vdb): ext4_mb_generate_buddy:747: group 1, block bitmap and bg descriptor inconsistent: 19550 vs 19551 free clusters ```3. About the applied patch
Quoted 10 lines
The applied patch does fix the failure in my test environment. At least the failure doesn't occur after running the full 100 iterations.
Thanks
Jeffle
Hi, Jeffle:
Thanks for that information. I'm still unable to reproduce your failure, but by inspection your patch clearly fixes a bug, and of course, you're seeing that. I suspect the code in
rm_leafthat also sets the partial cluster nofree state is masking the bug in my testing. In your case, my best guess is that your testing is occasionally getting into the retry loop for EAGAIN inremove_space. This would effectively expose the bug again and could lead to the failure you've described.Your patch has survived all the heavy testing I've thrown at it. So, please repost your RFC patch as a fix, and feel free to add:
Reviewed-by Eric Whitney <[email protected]>This points out that the cluster freeing code really needs to be cleaned up, so I'm working on a patch series that does that.
Thanks for your patience,
Eric