Whamcloud - gitweb
wangdi [Fri, 1 Dec 2006 07:40:45 +0000 (07:40 +0000)]
Branch:b_new_cmd
Init cat log for mds when add ost in recoverying, to make sure all unlink log can
be initialized after config log is processed. This could avoid race beween unlink
and llog init postrecovery, which should be the reason of b11218.
b:11218
nikita [Thu, 30 Nov 2006 21:11:19 +0000 (21:11 +0000)]
try smaller bh lru size.
tappro [Thu, 30 Nov 2006 14:15:22 +0000 (14:15 +0000)]
- call md_clear_open_replay_data() right before freeing och instead of doing
that before ll_done_writing(). This fix prevents possible handle loss when
replay happens between open/close and done_writting.
nikita [Thu, 30 Nov 2006 10:27:46 +0000 (10:27 +0000)]
sabdwich hash.
nikita [Thu, 30 Nov 2006 10:18:53 +0000 (10:18 +0000)]
another return value fix.
nikita [Wed, 29 Nov 2006 21:22:39 +0000 (21:22 +0000)]
2.6-rhel series: increase BH_LRU_SIZE.
nikita [Wed, 29 Nov 2006 19:09:54 +0000 (19:09 +0000)]
log_do_checkpoint(): fix return value.
nikita [Wed, 29 Nov 2006 19:09:12 +0000 (19:09 +0000)]
remove obsolete debugging.
fanyong [Wed, 29 Nov 2006 13:57:54 +0000 (13:57 +0000)]
Fix BUG for mount permission:
(1) if client failed to identification, it should not reconnect endlessly.
(2) client should check the connection error and return the value correctly.
huanghua [Wed, 29 Nov 2006 09:59:34 +0000 (09:59 +0000)]
fix the create mode bug
wangdi [Wed, 29 Nov 2006 09:44:34 +0000 (09:44 +0000)]
Branch:b_new_cmd
add evict_client proc for mdt
tappro [Tue, 28 Nov 2006 22:03:22 +0000 (22:03 +0000)]
- bug in lustre_msg_clear_flags(). For version 2 it clears ALL flags instead of
passed as parameter. b1_5 is affected too.
ericm [Tue, 28 Nov 2006 21:23:23 +0000 (21:23 +0000)]
branch: b_new_cmd
port from b1_8_gss: add a gss sanity test for enlarge_reqbuf().
nikita [Tue, 28 Nov 2006 21:07:29 +0000 (21:07 +0000)]
fix for LASSERT(*start <= hash) bug.
huanghua [Tue, 28 Nov 2006 15:24:20 +0000 (15:24 +0000)]
fix a memleak @lustre_md->mea in ll_prep_inode().
fanyong [Tue, 28 Nov 2006 13:13:48 +0000 (13:13 +0000)]
(1) Fix setxid permission check order according to the upcall result packed order.
(2) More detail comment.
fanyong [Tue, 28 Nov 2006 09:39:47 +0000 (09:39 +0000)]
(1)Add more rules for setxid.conf, permit to enable/disable setxid permission explicitly.
(2)Fix bug in cmd3-53 for local client setgrp permission out-of-control.
(3)More detail comment.
yury [Tue, 28 Nov 2006 07:58:05 +0000 (07:58 +0000)]
- make split size 128K to ensure that no split happens in recovery but still happens in perf. tests.
tappro [Mon, 27 Nov 2006 19:46:51 +0000 (19:46 +0000)]
- fix recovery issue with SOM. The setattr/done_writting use mfd on server but
can't find the correct one during replay because mfd handle is changed.
Current fix add old_handle field in mfd so during replay it is possible to find
it comparing old handle.
vitaly [Mon, 27 Nov 2006 17:44:01 +0000 (17:44 +0000)]
Branch b_new_cmd
b=22564
cancel all the locks on umount before shutting down the close thread
to flush all the dirty cache and send needed Size-on-MDS updates on MDS.
vitaly [Mon, 27 Nov 2006 14:48:53 +0000 (14:48 +0000)]
Branch b_new_cmd
b=22564,11103
check if unhashed and rehash under the dcache lock.
print epoch to be closed properly.
wangdi [Mon, 27 Nov 2006 08:58:05 +0000 (08:58 +0000)]
Branch:b_new_cmd
remove useless debug info after 11123 is fixed
fanyong [Mon, 27 Nov 2006 05:40:09 +0000 (05:40 +0000)]
Fix test script for local acl operation.
wangdi [Sun, 26 Nov 2006 04:41:20 +0000 (04:41 +0000)]
Branch:b_new_cmd
if ost failed is detected in lfs find, only return 0 to make tranverse continue
for cmd3-test18
yury [Sat, 25 Nov 2006 13:08:44 +0000 (13:08 +0000)]
- some cleanups in lmv.
ericm [Fri, 24 Nov 2006 18:15:11 +0000 (18:15 +0000)]
branch: b_new_cmd
the part missing from last commit: don't taint passed-in obj until we
know everything is fine.
nikita [Fri, 24 Nov 2006 17:27:15 +0000 (17:27 +0000)]
lmv: 0. more fid correctness checks, 1. fix wrong mea_hash_segment calculation.
huanghua [Fri, 24 Nov 2006 13:48:29 +0000 (13:48 +0000)]
this fix a bug in Lustre over NFS. But this flag is temparory as the patch said.
yury [Fri, 24 Nov 2006 13:29:37 +0000 (13:29 +0000)]
- cleanups in lmv:
- removed not needed @body initializatiion;
- added assert that body field in reply should be swabbed in mdc before using it in lmv.
wangdi [Fri, 24 Nov 2006 12:39:18 +0000 (12:39 +0000)]
Branch: b_new_cmd
We should try to retrieve the MEA from MDS anyway in getattr
fanyong [Fri, 24 Nov 2006 09:49:29 +0000 (09:49 +0000)]
Add ASSERT for filename validity check.
ericm [Fri, 24 Nov 2006 04:36:51 +0000 (04:36 +0000)]
branch: b_new_cmd
back port from b1_8_gss:
- don't taint passed-in obj until we know everything is fine.
- multi-process client gss daemon, to prevent one unresponsive server
block other context negotiation. nfs-utils patch not updated though.
- fix regression of exporting context of des3-hmac-sha1 on old kerberos.
nikita [Thu, 23 Nov 2006 18:45:22 +0000 (18:45 +0000)]
osd: disable back expensive fid_is_local() assertion.
nikita [Thu, 23 Nov 2006 18:44:45 +0000 (18:44 +0000)]
lmv/iam: new "scalable" readdir.
nikita [Thu, 23 Nov 2006 18:06:59 +0000 (18:06 +0000)]
fid: factor common function out.
nikita [Thu, 23 Nov 2006 18:05:40 +0000 (18:05 +0000)]
liblustre: add hash_long() from linux/hash.c.
fanyong [Thu, 23 Nov 2006 14:01:52 +0000 (14:01 +0000)]
Do NOT count the terminating '\0' of filename for length.
huanghua [Thu, 23 Nov 2006 10:09:12 +0000 (10:09 +0000)]
remote operation should set MDS_CROSS_REF flag.
fanyong [Thu, 23 Nov 2006 06:39:59 +0000 (06:39 +0000)]
More sanity check for filename length.
huanghua [Thu, 23 Nov 2006 05:17:28 +0000 (05:17 +0000)]
fix a typo and clear out the fid.
huanghua [Thu, 23 Nov 2006 02:20:44 +0000 (02:20 +0000)]
more precise debugging code.
wangdi [Wed, 22 Nov 2006 16:34:15 +0000 (16:34 +0000)]
Branch: b_new_cmd
miss kunmap for those reset offset page. found by umka
fanyong [Wed, 22 Nov 2006 14:55:39 +0000 (14:55 +0000)]
More sanity check for nlink of file.
nikita [Wed, 22 Nov 2006 14:43:48 +0000 (14:43 +0000)]
jbd-checkpoint-on-commit: fix locking.
tappro [Wed, 22 Nov 2006 13:38:51 +0000 (13:38 +0000)]
- fix recovery issue. dropping obd_recovering flag before notify() cause loosing
the objects on OST due to orphan destroying. The obd_recovering should be
zeroed after notify()
tappro [Wed, 22 Nov 2006 13:36:06 +0000 (13:36 +0000)]
- make sure there is transno for close/done_write() to keep them in replay queue
tappro [Wed, 22 Nov 2006 13:22:55 +0000 (13:22 +0000)]
- count -EAGAIN during close() as valid answer
tappro [Wed, 22 Nov 2006 13:17:03 +0000 (13:17 +0000)]
- fix objid update call. It should be done before transaction stop. Separate it
from freeing lov_create data.
- encapsulate checking no_lov_create flag inside mdd_lov_create
huanghua [Wed, 22 Nov 2006 13:06:18 +0000 (13:06 +0000)]
add debug code to catch b11241
huanghua [Wed, 22 Nov 2006 07:37:30 +0000 (07:37 +0000)]
let export Lustre over NFS work.
basic test works now. More tests needed.
huanghua [Wed, 22 Nov 2006 07:36:34 +0000 (07:36 +0000)]
fix oops while accessing NULL pointer
fanyong [Wed, 22 Nov 2006 03:31:53 +0000 (03:31 +0000)]
(1) Drop unnecessary permission check for name_{insert,remove}.
(2) Do local permission check for name_{insert,remove} before remote ops.
(3) Add some comment.
fanyong [Wed, 22 Nov 2006 02:23:20 +0000 (02:23 +0000)]
Add symlink target length check.
nikita [Tue, 21 Nov 2006 21:12:00 +0000 (21:12 +0000)]
iam: revert changes to IAM_LOOKUP_LAST that might caused spurious -ENOENT.
nikita [Tue, 21 Nov 2006 17:15:20 +0000 (17:15 +0000)]
lu: ultimately hash function that works(tm)... famous last words.
nikita [Tue, 21 Nov 2006 14:24:40 +0000 (14:24 +0000)]
add per-thread debugging flags. Use then to control CDEBUG().
nikita [Tue, 21 Nov 2006 14:22:45 +0000 (14:22 +0000)]
lu/mdt: more site stats.
yujian [Tue, 21 Nov 2006 08:45:49 +0000 (08:45 +0000)]
The sixth field in the csv entry is "fsname", so just move fsname option to its position.
fanyong [Tue, 21 Nov 2006 07:40:33 +0000 (07:40 +0000)]
(1) For remote create, only continue when lookup return ENOENT
(2) For mdd_lookup, do NOT return ENOENT when parent is dead obj.
ENOENT means child does not exist.
fanyong [Tue, 21 Nov 2006 06:04:40 +0000 (06:04 +0000)]
Miss to set spec->u.sp_ea.fid for remote object_create with DEFAULT_ACL before.
huanghua [Tue, 21 Nov 2006 02:36:42 +0000 (02:36 +0000)]
(1) do not return any entry to client for a dead directory;
(2) add nlink overflow checking for mkdir. more checking will be added soon.
brian [Tue, 21 Nov 2006 02:16:01 +0000 (02:16 +0000)]
Change the key for objects to their name.
Support "--lmv <lmvname>" for CMD. Using --lmv is how LMVs were built with
lmc in 1.4 and since we are trying to source lmc, we should simply use the
same syntax.
find_obj()->find_objs() now returns all objects that match.
So callers need to expect lists and deal with them. In a lot of cases a
caller can *know* that the list returned will be only one item if the lmc
file is contructed correctly.
Support specification of multiple MDTs on the same MDS using different
devices.
For this, the mountpoint of the MDT needs to use the fsname to make it
unique.
The MDTs and OSTs also need to be created with fsnames.
Use the --node <name> from the lmc mtpt object as the fsname.
ericm [Tue, 21 Nov 2006 01:01:31 +0000 (01:01 +0000)]
branch: b_new_cmd
more fix to gss module loading: waiting init channel be opened in gss
module itself.
nikita [Mon, 20 Nov 2006 23:17:56 +0000 (23:17 +0000)]
__find_get_block_slow-scale.patch: attempt to reduce contention on bd_mapping->private_lock.
nikita [Mon, 20 Nov 2006 23:16:22 +0000 (23:16 +0000)]
iam: 1. more correctness checks, 2. fixes to readdir locking.
ericm [Mon, 20 Nov 2006 23:07:18 +0000 (23:07 +0000)]
branch: b_new_cmd
gss daemon debugging message.
yury [Mon, 20 Nov 2006 15:00:15 +0000 (15:00 +0000)]
- in cmm_split_check() passed to OBD_FREE correct size (though this is not bug, as size does not matter in linux memory manager);
- cleanups in mdd_lov.c
nikita [Mon, 20 Nov 2006 14:35:12 +0000 (14:35 +0000)]
lu: now that we have larger hash table, switch to the simpler hash.
nikita [Mon, 20 Nov 2006 13:44:52 +0000 (13:44 +0000)]
iam: remove wrong assertion introduced by previous commit.
nikita [Mon, 20 Nov 2006 13:40:11 +0000 (13:40 +0000)]
iam: 0. do not unlock leaf while advancing index part. 1. avoid binary search when possible.
nikita [Mon, 20 Nov 2006 13:38:01 +0000 (13:38 +0000)]
lu/mdt: more stat stats.
yury [Mon, 20 Nov 2006 13:05:20 +0000 (13:05 +0000)]
- simplify and optimize lmv_placement_policy(). By now it does not do additional raw_name2idx()
wangdi [Mon, 20 Nov 2006 10:43:48 +0000 (10:43 +0000)]
Branch: b_new_cmd
add debug_info in ll_file_read for test36
nikita [Mon, 20 Nov 2006 06:26:13 +0000 (06:26 +0000)]
lu: use vmalloc() to allocate hash table.
nikita [Mon, 20 Nov 2006 06:25:38 +0000 (06:25 +0000)]
log_do_checkpoint(): fix return value.
huanghua [Mon, 20 Nov 2006 02:57:14 +0000 (02:57 +0000)]
(1) use hash value 0xfffffffeUL to mark end of a directory instead of -1;
(2) add a TODO to handle readdir from removed dir. mdd_object is set to be dead when its nlink drops to zero.
nikita [Mon, 20 Nov 2006 01:02:10 +0000 (01:02 +0000)]
iam fixes:
- defect: iam: split_index_node(): adjust ->curidx of split frame (45).
- defect: iam: ipd was shared in iam-uapi (240).
- defect: iam: dx_check_full_path(): renew ->at (60).
nikita [Sun, 19 Nov 2006 14:50:21 +0000 (14:50 +0000)]
jbd-checkpoint-on-commit: dix possible deadlock.
tappro [Sun, 19 Nov 2006 13:46:14 +0000 (13:46 +0000)]
add 'exist' info in lu_object_header_print()
tappro [Sun, 19 Nov 2006 13:38:08 +0000 (13:38 +0000)]
debug for failed object
nikita [Sun, 19 Nov 2006 12:39:55 +0000 (12:39 +0000)]
use nr_free_buffer_pages() to calculate hash table size. It has to be exported.
tappro [Sun, 19 Nov 2006 11:36:30 +0000 (11:36 +0000)]
- led_lock is correct here
tappro [Sun, 19 Nov 2006 08:46:15 +0000 (08:46 +0000)]
- wrong lock was used to protect namespace operation
nikita [Sun, 19 Nov 2006 00:55:23 +0000 (00:55 +0000)]
fix possible deadlock in jbd-checkpoint-on-commit.
nikita [Sun, 19 Nov 2006 00:54:40 +0000 (00:54 +0000)]
iam: remove unreliable correctness check.
nikita [Sat, 18 Nov 2006 18:52:00 +0000 (18:52 +0000)]
mdt/lu: add lu_site statistics at /proc/fs/lustre/mdt/*/site_stats
tappro [Sat, 18 Nov 2006 10:06:31 +0000 (10:06 +0000)]
- more debug
tappro [Sat, 18 Nov 2006 08:11:17 +0000 (08:11 +0000)]
- raw_lookup() is not intent operation, no need to handle disposion
- add debug
wangdi [Sat, 18 Nov 2006 03:55:43 +0000 (03:55 +0000)]
Branch: b_new_cmd
add this assert for debuging b11123
ericm [Sat, 18 Nov 2006 03:29:06 +0000 (03:29 +0000)]
branch: b_new_cmd
in case of gss module newly loaded, gss service upcall might be issued
before channel be opened by gss daemon, nfsv4 cache will directly drop
the upcall. workaround it by sleep after module loaded; and in gss daemon
try to keep channel always open.
yury [Fri, 17 Nov 2006 17:28:42 +0000 (17:28 +0000)]
- in lmv_placement_policy() add missed lmv_object_put()
nikita [Fri, 17 Nov 2006 15:57:35 +0000 (15:57 +0000)]
obd: use ({ ... }) instead of do { ... } while (0) in OBD macros.
nikita [Fri, 17 Nov 2006 15:56:53 +0000 (15:56 +0000)]
lu: make site hash table dynamically adjustable.
yury [Fri, 17 Nov 2006 15:08:49 +0000 (15:08 +0000)]
- fixed one more race in LMV which made it posssible to exist cross-ref files.
vitaly [Fri, 17 Nov 2006 13:08:42 +0000 (13:08 +0000)]
Branch b_new_cmd
b=11224
let dentry to be already hashed at the time of ll_create_it
fanyong [Fri, 17 Nov 2006 09:24:34 +0000 (09:24 +0000)]
Fix about remote acl operation:
(1) push_ctxt can not switch user space context (uid / gid),
do not use it any more.
(2) ship user info (uid / gid) as argv[] to user space by upcall,
switch user by setuid / setgid in user space.
(3) do not cache for remote acl upcall,
do not use fid as key for remote acl upcall,
it maybe cause wrong result when concurrent remote acl operations.
(4) remove unused code.
fanyong [Fri, 17 Nov 2006 03:14:29 +0000 (03:14 +0000)]
For change file time operation (utime),
file owner and CAP_FOWNER privilege are preferential.
vitaly [Thu, 16 Nov 2006 10:54:39 +0000 (10:54 +0000)]
Branch b_new_cmd
b=11103
Do not let dentry to be unhashed after create,
some callers, e.g. do_coredump, expect it to be hashed.
fanyong [Thu, 16 Nov 2006 10:22:06 +0000 (10:22 +0000)]
Do UCRED_OLD authentication when gss disabled.
fanyong [Thu, 16 Nov 2006 07:41:42 +0000 (07:41 +0000)]
(1) Remove fs privilege (CAP_FS_MASK) for non-root user, it can result in
invalid authorization.
(2) Do UCRED_OLD authentication for inter-mdt ops and for the case that
client and mdt on the same node. (gss can not distinguish the two cases)
(3) Fix Bug of authority promote for inter-mdt ops.