macan/pomegranate

An Efficient Distributed File/Blob/Key-Value Store for Billions of Small Files

pomegranate's Issues

Bug in xnet_simple st_update()

Sometimes, the accepted socket fd can not be set to the table. There is a conflict entry already table, maybe the last broken connection. Check it!

ITB active entry count is incorrect

Under heavy memory pressure, the active entry counter in ITB is not correct.
It is easy to reproduce this bug. Run the API unit test several times, you can got it!

xnet_free_msg() bug

In xnet-simple RESEND mode, there is a race between the reply message and the original xnet_msg, it is hard to trace the calling path.
For now, we just disable the RESEND mode to fix this bug:(

Source file missing from tree

[root@vs1 Pomegranate]# make
make: *** No rule to make target /usr/src/Pomegranate/test/xnet/cr.c', needed byunit_test'. Stop.

itb_dirty assertion failed!

itb_dirty in future access case, we got a assertion failure.
it is related to mds_read_itb() and cbht_itb_miss(), the itb->h.state can be CLEAN.

Crash on gossiping removed directory

[macan@gh01 hvfs]$ *** glibc detected *** /home/macan/hvfs/test/xnet/mds.ut: double free or corruption (!prev): 0x00002aaaac48fce0 ***
======= Backtrace: =========
/lib64/libc.so.6[0x3bab271ce2]
/lib64/libc.so.6[0x3bab2738e2]
/lib64/libc.so.6(realloc+0x1d0)[0x3bab275c30]
/home/macan/hvfs/test/xnet/mds.ut(txg_add_rdir+0x7d)[0x44373d]
/home/macan/hvfs/test/xnet/mds.ut(mds_gossip_rdir+0x1d7)[0x45ba37]
/home/macan/hvfs/test/xnet/mds.ut(mds_mds_dispatch+0x127)[0x464667]
/home/macan/hvfs/test/xnet/mds.ut[0x45d5b2]
/lib64/libpthread.so.0[0x3babe06367]
/lib64/libc.so.6(clone+0x6d)[0x3bab2d2f7d]

MDSL ITB toe lookup failure

The writebacked ITB can not be lookuped in the TOE list.
Maybe the itb counter is not correct or the MDSL code is not correct, should be rechecked!

ITB_JUST_SPLIT Assertion failed

Enable memory limit, run unit test.
After running a few TXGs, mds.ut crashed with ITB reference assertion failure.

It is a conflict between txg_wb() and aur_itb_split(). The usage of itb_put() should be revised.

Hash Imbalance in Postmark Test

Using postmark to test tiny file I/O, we found that the disk usage for each server is imbalance. The largest disk image can be 2 or 3 times of the smallest one. Maybe a new hash function is useful for postmark test.

xTable

How does xtable organize data in the Pomegranate file system? Could you please explain it with a diagram? Thank you very much

Large object read/write support

Add a reliable (replicated), striped OSD level to PFS.

OBJ replication for OSD layer

New Feature.

MDS scrub_thread stop at sem_wait with scrub_running set

scrub_running flag is set, but the scrub thread is not running.

I believe that there is a memory set conflict on scrub_running flag. So protect it!

PFS fuse client failed with mdtest

macan@MACANA ~/hvfs/working $ git bisect start
macan@MACANA ~/hvfs/working $ git bisect bad
macan@MACANA ~/hvfs/working $ git bisect good 1491353
Bisecting: 24 revisions left to test after this (roughly 5 steps)
[19f1fb1] Finish PFS xattr integrated test case. Seems that it works fine :)
macan@MACANA ~/hvfs/working $ git bisect good 19f1fb1
Bisecting: 12 revisions left to test after this (roughly 4 steps)
[15b2e14] Cleanup c2m.c and remove redundant code, and add fuse client latency distrbution. To use fuse client latency distribution, add 'CFLAGS='-DHVFS_FUSE_STAT'' and 'USE_FUSE=1' before make.
macan@MACANA ~/hvfs/working $ git bisect good 15b2e14
Bisecting: 6 revisions left to test after this (roughly 3 steps)
[ba24348] State checking and corrupt message checking in bitmap gossip
macan@MACANA ~/hvfs/working $ git bisect good
Bisecting: 2 revisions left to test after this (roughly 2 steps)
[2a11fef] Merge fuse microbenchmark from master branch
macan@MACANA ~/hvfs/working $ git bisect good
Bisecting: 0 revisions left to test after this (roughly 1 step)
[97d6bac] Add dbsearch bench, change xattr microbench for fast12 paper.
macan@MACANA ~/hvfs/working $ git bisect bad
Bisecting: 0 revisions left to test after this (roughly 0 steps)
[1d0595a] Internal API bug fix for hvfs_stat()