W dniu 16.09.2016 o 05:33, Kent Overstreet pisze:

On Thu, Sep 15, 2016 at 11:36:14AM +0200, Marcin Mirosław wrote:

quoted

Hi!
I was playing with fs without tiering. I was using it for tmp dir for
compilation. Next I changed in sys:
echo crc64 > options/data_checksum
echo crc64 > options/metadata_checksum
echo crc64 > options/str_hash

After a couple of minutes I got:
[ 8372.574346] bcache (dm-10): IO error on dm-10 for checksum error
[ 8372.680196] bcache (dm-10): IO error on dm-10 for checksum error
[ 8464.361860] bcache (dm-10): IO error on dm-10 for checksum error
[ 8466.146966] bcache (dm-10): IO error on dm-10 for checksum error
[ 8466.995095] bcache (dm-10): IO error on dm-10 for checksum error
[ 8469.199749] bcache (dm-10): IO error on dm-10 for checksum error
[ 8469.441408] bcache (dm-10): IO error on dm-10 for checksum error
[ 8469.722676] bcache (dm-10): IO error on dm-10 for checksum error
[ 8469.827055] bcache (dm-10): IO error on dm-10 for checksum error
[ 8470.038869] bcache (dm-10): IO error on dm-10 for checksum error
[ 8470.236663] bcache (dm-10): IO error on dm-10 for checksum error
[ 8470.427094] bcache (dm-10): IO error on dm-10 for checksum error
[ 8472.030519] bcache (dm-10): IO error on dm-10 for checksum error
[ 8473.098820] bcache (dm-10): IO error on dm-10 for checksum error
[ 8916.491297] bcache (dm-10): IO error on dm-10 for checksum error
[ 8916.715057] bcache (dm-10): IO error on dm-10 for checksum error
[ 8916.715111] bcache (dm-10): too many IO errors on dm-10, setting
filesystem RO
[ 8916.733056] bcache (dm-10): IO error on dm-10 for checksum error
[ 8916.733125] bcache (dm-10): dm-10 read only
[ 8916.733161] bcache (dm-10): too many IO errors on dm-10, setting
device RO
[ 8916.988286] bcache (dm-10): IO error: read only
[ 8916.988545] bcache (dm-10): IO error: read only

Ok, it turns out the crc64 for data checksums code was just fubar. Fix is up
(the fix does change how crc64 is computed for bios though, so it'll be
incompatible with your existing filesystem).

Also pushed a patch that adds some more error messages to fs-gc, we should
figure out why it wouldn't mount. I can't think of any reason why data checksum
errors would've caused that.

Hi Kent, hi all,
when I tried to mount fs that has troubles yesterday I've got:
[  494.296818] bcache (dm-10): dm-10: journal checksum bad (got
18446744072224191025 expect 2809606705), sector 2048u
[  494.309973] bcache (dm-10): dm-10: journal checksum bad (got
18446744073320597786 expect 3906013466), sector 2304u
[  494.311597] bcache (dm-10): dm-10: journal checksum bad (got
18446744070980686285 expect 1566101965), sector 2560u
[  494.313038] bcache (dm-10): dm-10: journal checksum bad (got
18446744073177643543 expect 3763059223), sector 2816u
[  494.324082] bcache (dm-10): dm-10: journal checksum bad (got
18446744070081456445 expect 666872125), sector 3072u
[... many similar lines...]
[  495.000229] bcache (dm-10): dm-10: journal checksum bad (got
18446744071270315299 expect 1855730979), sector 90368u
[  495.001373] bcache (dm-10): dm-10: journal checksum bad (got
18446744070901133954 expect 1486549634), sector 90624u
[  495.002696] bcache (dm-10): dm-10: journal checksum bad (got
18446744071373615633 expect 1959031313), sector 90880u
[  496.618084] bcache (dm-10): journal replay error: -28
[  496.618124] bcache: bch_open_as_blockdevs() register_cache_set err
journal replay failed
[  496.796085] bcache (dm-10): stopped


What str_hash does?

Today I formated block device and again I play with changing
"compression, data_checksum, metadata_checksum, str_hash". I was
changing options while intensive writing to fs. Two times I had hard
lockup of kernel. No chance for getting dmesg. After first lockup I
caouldn't mount fs again due to:
kernel: [  260.141942] bcache: bch_open_as_blockdevs()
register_cache_set err bad btree root

So -> format -> testing - hard lockup. On the second time I could mount
again fs:
kernel: [  234.920846] bcache (dm-11): journal replay done, 29 keys in 1
entries, seq 3447

I'm thinking about using netconsole but I'm not sure I would have a time
for this before tuesday.

Thanks,
Marcin