Infinite assertion failures and/or hang with metadata_csum=1 and replica_metadata=0

[stevenjswanson]

With these configuration options, I get a whole system hang. Sometimes w/ and sometimes w/o a bunch of assertion failures on dmesg.

on 84d3e6a, but it it's been present since at least 10142f3.

This happens with

measure_timing=0
inplace_data_updates=0
wprotect=0 
mmap_cow=1      
unsafe_metadata=0     
replica_metadata=0 
metadata_csum=1 
dram_struct_csum=1      
data_csum=1 
data_parity=1

but not

measure_timing=0
inplace_data_updates=0
wprotect=0
mmap_cow=1
unsafe_metadata=0
replica_metadata=0 
metadata_csum=0 ***
dram_struct_csum=1 
data_csum=1 
data_parity=1

# sudo umount /mnt/ramdisk; sudo rmmod nova; sudo modprobe nova measure_timing=0      inplace_data_updates=0      wprotect=0 mmap_cow=1      unsafe_metadata=0     replica_metadata=0 metadata_csum=1 dram_struct_csum=1      data_csum=1 data_parity=1; sudo mount -t NOVA -o init /dev/pmem0 /mnt/ramdisk; echo 1 | sudo tee  /proc/fs/NOVA/pmem0/create_snapshot ;cat /proc/fs/NOVA/pmem0/snapshots
<hangs->

# dmesg
[ 4064.180232] nova: nova_rebuild_dir_inode_tree: unknown type 195, 0x3eee000
[ 4064.181171] nova: nova_rebuild_dir_inode_tree: unknown type 195, 0x3eee000
[ 4064.181208] nova: nova_rebuild_dir_inode_tree: unknown type 195, 0x3eee000
[ 4064.181229] assertion failed fs/nova/rebuild.c:673: 0
[ 4064.181348] nova: nova_rebuild_dir_inode_tree: unknown type 195, 0x3eee000
[ 4064.181386] nova: nova_rebuild_dir_inode_tree: unknown type 195, 0x3eee000
[ 4064.181442] nova: nova_rebuild_dir_inode_tree: unknown type 195, 0x3eee000
[ 4064.181457] nova: nova_rebuild_dir_inode_tree: unknown type 195, 0x3eee000
[ 4064.181508] assertion failed fs/nova/rebuild.c:673: 0
[ 4064.181541] assertion failed fs/nova/rebuild.c:673: 0
[ 4064.181556] nova: nova_rebuild_dir_inode_tree: unknown type 195, 0x3eee000
<...forever...>

[luzh]

metadata_csum shouldn't be used without replica_metadata, as we rely on the tick-tock scheme to maintain consistency, since at least two fields are updated in some metadata structure: something + checksum. So some code may not handle this configuration well. I think we can fix it at mount time by setting replica_metadata = 1 when metadata_csum = 1.

[Andiry]

That's true. I think we should just combine this two flags.

…

On Fri, Jul 7, 2017 at 5:03 PM, Lu Zhang ***@***.***> wrote: metadata_csum shouldn't be used without replica_metadata, as we rely on the tick-tock scheme to maintain consistency, since at least two fields are updated in some metadata structure: something + checksum. So some code may not handle this configuration well. I think we can fix it at mount time by setting replica_metadata = 1 when metadata_csum = 1. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#30 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABQcEM4YJ1MtQ1al9SmS-hczdG-vfHgGks5sLsczgaJpZM4ON8Tz> .

[stevenjswanson]

I agree.

-steve

-- Composed on (and maybe dictated to) my phone. On Jul 7, 2017, at 17:29, Andiry Xu <notifications@github.com> wrote: That's true. I think we should just combine this two flags.

On Fri, Jul 7, 2017 at 5:03 PM, Lu Zhang ***@***.***> wrote: metadata_csum shouldn't be used without replica_metadata, as we rely on the tick-tock scheme to maintain consistency, since at least two fields

are

updated in some metadata structure: something + checksum. So some code may not handle this configuration well. I think we can fix it at mount time by setting replica_metadata = 1 when metadata_csum = 1. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#30 (comment)>, or

mute

the thread <

https://github.com/notifications/unsubscribe-auth/ABQcEM4YJ1MtQ1al9SmS-hczdG-vfHgGks5sLsczgaJpZM4ON8Tz

.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#30 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIpg3cS3K8ezrcCprA7DPvNd6fuik-blks5sLs1QgaJpZM4ON8Tz> .

[luzh]

A quick fix just to make them equal: a5f89a2

[stevenjswanson]

If we are going to fix, let’s fix it right and just have on flag. This just makes the code messier and its behavior harder to understand. -steve

…

On Jul 7, 2017, at 10:45 PM, Lu Zhang ***@***.***> wrote: A quick fix just to make them equal: a5f89a2 <a5f89a2> — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#30 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIpg3XJt2YVoiPDps3OsIkAY9y3pQ72wks5sLxdqgaJpZM4ON8Tz>.

[luzh]

Moved to #35