Whamcloud - gitweb
wangdi [Fri, 1 Aug 2003 03:36:44 +0000 (03:36 +0000)]
update changes of lookup_intent for hp kernel
rread [Fri, 1 Aug 2003 00:19:57 +0000 (00:19 +0000)]
* doh! (thanks, mike)
rread [Thu, 31 Jul 2003 23:55:56 +0000 (23:55 +0000)]
* simple cleanups to get replay-single.sh to run again, also added
a trap ERR.
zab [Thu, 31 Jul 2003 20:29:35 +0000 (20:29 +0000)]
make sanity and iozone run from llite over the lov in the new ocp
io model.
- add ocp_off to the ocp so the lov can remap it as it passes through
- fix up the page accounting a bit 'cause truncate doesn't really know
right know to subtract pages from the dirty or io counts
- pass our completion struct, _not_ a pointer to the complete function
- make sure to always free the ocp in commit_write_sync
- disable the watermark io until it actually does something
- make sure to alter the filehandle/o_id as we pass through the lov
- initialize our loi bits in lov_unpackmd
zab [Thu, 31 Jul 2003 17:15:41 +0000 (17:15 +0000)]
- introduce tunables for the aggregate amount of cached data that llite will
allow on a machine. When we have more than 'dirty_high_water_MB' meg of
data dirty we start async ocp io on the amount of data between the high
water mark and 'dirty_low_water_MB'. If we ever have more than
'max_dirty_MB' we write synchronously in commit_write, batching with other
dirty data in the inode. these watermarks are checked every time we pass
through commit_write and we account seperately for dirty pages and pages
that are under io.
- add ocp_count to the ocp so that we can write partial pages at the end
of the file. make the callback func a proper type.
- remove iod.c, the few code paths we cared about in there are moved to
rw{,24,26}.c
- turn off generic file read-ahead by setting filp->f_ramax to 0.
- have writepage and commit_write when writing sync use the plugging to batch
their io with io started from other dirty pages in the inode's dirty list
- woops, missed a hardcoded '16'. move to max_pags_per_rpc
There's still much to do. The 2.6 side needs work on its many-page variants,
kiovec->bio, etc. We need to actually start watermark io from the dirty inode
list.
adilger [Wed, 30 Jul 2003 20:35:18 +0000 (20:35 +0000)]
Ensure OST objids are unique even after a crash by using the mount count as
a generation number. Remove old "superhack" code.
Add a "force_sync" parameter to filter_update_server_data() and
fsfilt_write_record() so that we ensure the mount_count is safe on disk
when the OST has started.
b=294
r=phil
eeb [Wed, 30 Jul 2003 04:13:55 +0000 (04:13 +0000)]
* Optimized qswnal GETs use EKC rpcs to "teleport" payload (as far as
portals is concerned).
* allowed qswnal to have knowledge of lib_msg_t.
* lib_fake_reply_msg() written to allow qswnal to create replies and
deliver them in the normal fashion (with lib_finalize()).
zab [Mon, 28 Jul 2003 20:49:56 +0000 (20:49 +0000)]
- put in initial lov obd plug/async/unplug APIs. for now its giving
the stripe to the osc via an explicit loi argument. this will be
refactored when the refactoring described in 314 arrives.
alex [Mon, 28 Jul 2003 12:18:00 +0000 (12:18 +0000)]
- simple optimization for EXT3:
ext3_write_super() doesn't need to start commit as kupdated wants. this
improves things significantly in some cases
alex [Mon, 28 Jul 2003 12:02:04 +0000 (12:02 +0000)]
- minor changes in Andreas'es ext3 direct-io patch to be applicable for
2.4.18-chaos
- ext3-o_direct-1-2.4.18-chaos.patch have been added to chaos-2.4.18 series
adilger [Sun, 27 Jul 2003 20:31:08 +0000 (20:31 +0000)]
Config for 2-mounts-on-local-host-with LOV (currently not-usable until
bug 938 is fixed).
eeb [Sun, 27 Jul 2003 02:38:15 +0000 (02:38 +0000)]
* Merged b_devel updates
eeb [Sun, 27 Jul 2003 02:05:48 +0000 (02:05 +0000)]
* Portals MTU now 512K (honest)
* lconf configures an EKC tunable on elan nodes to help bulk comms
braam [Sat, 26 Jul 2003 23:14:53 +0000 (23:14 +0000)]
- do not send unused uid and gid fields in mdc/mds update packets
- correctly handle subdirectories of sgid directories (owner set to
parent id - test 6g added for this)
- fix for 1607 (test 19c)
- remove unused kernel configuration files
shaver [Sat, 26 Jul 2003 17:59:00 +0000 (17:59 +0000)]
b=1590, (partial 948):
Stop flushing on close, and instead flush all outstanding client data
atomically with the destruction of the final stripe.
And when flushing for destroy, or other cases of truncate-to-zero, send a
hint with the AST to indicate that dirty data should be discarded, instead of
being written back.
r=zab/phil
eeb [Sat, 26 Jul 2003 16:59:43 +0000 (16:59 +0000)]
* Guarded #define EXPORT_SYMTAB
* Bit the bullet and #defined PTL_MTU
NB currently defined in <portals/lib-types.h>; maybe should move...
shaver [Sat, 26 Jul 2003 15:21:26 +0000 (15:21 +0000)]
Import new ONLY/EXCEPT mechanism from sanity, and check in my trivial upcall.
shaver [Sat, 26 Jul 2003 03:28:16 +0000 (03:28 +0000)]
Jacob schooled me.
shaver [Sat, 26 Jul 2003 03:05:34 +0000 (03:05 +0000)]
Signals are hard.
shaver [Sat, 26 Jul 2003 02:57:06 +0000 (02:57 +0000)]
Add multi-operation microshell to aid in the testing of different replay
patterns.
phil [Sat, 26 Jul 2003 02:20:46 +0000 (02:20 +0000)]
clobber b_filterio patches with those from b_devel.
Until future notice, kernel patches must be changed only on b_devel,
and only with permission. Kernel patch conflicts can never be
resolved cleanly.
zab [Sat, 26 Jul 2003 01:38:11 +0000 (01:38 +0000)]
- add the /proc tunables to control the number of pages per rpc and the
number of concurrent rpcs
- make sure to check to fire new rpcs as an rpc completes
- push the rpc concurrency checks into the right place
- rearrange the rpc firing logic so it doesn't get stuck :/
braam [Sat, 26 Jul 2003 00:05:32 +0000 (00:05 +0000)]
define the it_flags correctly
shaver [Fri, 25 Jul 2003 23:57:15 +0000 (23:57 +0000)]
don't leak mcd when shutting down --failover
shaver [Fri, 25 Jul 2003 23:04:39 +0000 (23:04 +0000)]
b=1600: my version of Andreas' patch to use write_record to ensure that the
MDS server data is written at creation time
b=1597: cope with entering recovery/replay without a preceding timeout
We pass this initial single-node/single-mount replay test, yay.
adilger [Fri, 25 Jul 2003 22:35:27 +0000 (22:35 +0000)]
Recovery test script for dual-mount testing.
zab [Fri, 25 Jul 2003 22:13:12 +0000 (22:13 +0000)]
- move the osc rpc managing thread into the osc and rename it appropriately
- osc_rpcd's set is internal to the osc instead of passed in
- evolve the temporary 'barrier' api into the proper 'plug/unplug'
- move the 'ocp' management helpers from llite into genops.c
- start and stop osc rpc thread based on the presence of osc imports
- give test_rbw/echo_client the ability to use plug/unplug
- mirror some of the b_filterio test_brw/echo_client changes
- add an osc_internal.h for the osc rpcd
adilger [Fri, 25 Jul 2003 22:07:31 +0000 (22:07 +0000)]
Fix llobdstat usage message.
jacob [Fri, 25 Jul 2003 21:37:41 +0000 (21:37 +0000)]
Use $(CC) instead of gcc (fixes cross compile build).
braam [Fri, 25 Jul 2003 21:28:52 +0000 (21:28 +0000)]
- backport a few name changes in the patches to 2.4.18
- include some new I/O functions.
phil [Fri, 25 Jul 2003 21:18:41 +0000 (21:18 +0000)]
land b_unify on b_devel
- patches for 2.6
- linus-approved APIs
- lots of little reorg and fixes for 2.4/2.6 compatibility
- better code to manage MDC lock cancellation and dentry memory pressure
braam [Fri, 25 Jul 2003 18:08:36 +0000 (18:08 +0000)]
- remove quilt again.]+|[-+]\d{4}) \s (\d\d\d\d) \s - \s (.+) \
wangdi [Fri, 25 Jul 2003 09:29:51 +0000 (09:29 +0000)]
add quilt
shaver [Thu, 24 Jul 2003 19:07:10 +0000 (19:07 +0000)]
Initial single-node replay test framework, based on readonly and notransno
lctls and multi-facet configuration.
Next: an upcall and actual replay tests.
mmex [Wed, 23 Jul 2003 19:47:38 +0000 (19:47 +0000)]
- change configuration to use modern AC_DEFINE, otherwise /usr/bin/autoheader
fails when trying to compile RPMS. Single arg AC_DEFINE() is no
longer supported fully.
zab [Wed, 23 Jul 2003 16:16:08 +0000 (16:16 +0000)]
- put the oa in the loi and feed it to the rpcs, for now
- disable more old-guard code for the time being
- walk the right req list to find completed reqs
- init some things, deref some things, you know.
- appease lprocfs
- fix up the cleanup path of rpc building
wangdi [Wed, 23 Jul 2003 09:44:13 +0000 (09:44 +0000)]
add kernel.spec.in
wangdi [Wed, 23 Jul 2003 06:47:03 +0000 (06:47 +0000)]
add sysrq and crash support in lkcd-cvs-2.6.0-test1
wangdi [Wed, 23 Jul 2003 05:27:11 +0000 (05:27 +0000)]
add netdump in 2.6
wangdi [Wed, 23 Jul 2003 05:17:34 +0000 (05:17 +0000)]
fix flaw in 2.6
wangdi [Wed, 23 Jul 2003 05:12:40 +0000 (05:12 +0000)]
fix some flaws
braam [Wed, 23 Jul 2003 04:31:30 +0000 (04:31 +0000)]
- THIS BREAKS b_unify
2.6.0 patches now fixed to Linus specs. 2.6.0 compiles, untested.
beginning of similar changes for 2.4 changes
braam [Tue, 22 Jul 2003 16:45:53 +0000 (16:45 +0000)]
- this patch needs to be re-added with -ko because it contains rcs strings.
phil [Tue, 22 Jul 2003 16:29:21 +0000 (16:29 +0000)]
- in ll_umount_begin, protect against NULL obd
- fix debug subsystem in obdclass/obdo.c
- remove zero-length buffer check from lustre_msg_buf; we already have
the minimum-size check, and zero length is common in the "no EA yet"
getattr case
wangdi [Tue, 22 Jul 2003 15:09:34 +0000 (15:09 +0000)]
fix netdump error in 2.6
wangdi [Tue, 22 Jul 2003 15:04:14 +0000 (15:04 +0000)]
fix some flaws
wangdi [Tue, 22 Jul 2003 15:03:02 +0000 (15:03 +0000)]
add vfs_nointent_2.6.0-test1.patch
wangdi [Tue, 22 Jul 2003 15:01:56 +0000 (15:01 +0000)]
add vfs_nointent_2.6.0-test1.pc
wangdi [Tue, 22 Jul 2003 15:00:19 +0000 (15:00 +0000)]
add lkcd-cvs-2.6.0-test1.pc
wangdi [Tue, 22 Jul 2003 14:59:18 +0000 (14:59 +0000)]
add lkcd-cvs-2.6.0-test1.patch
phil [Tue, 22 Jul 2003 14:05:22 +0000 (14:05 +0000)]
partial merge of b_unify into b_devel (20030722_0202)
phil [Tue, 22 Jul 2003 13:51:01 +0000 (13:51 +0000)]
partial merge of b_unify into b_devel (20030722_0202)
braam [Tue, 22 Jul 2003 07:20:21 +0000 (07:20 +0000)]
- these should apply, but they don't quite compile (something with netdump)
braam [Tue, 22 Jul 2003 07:06:31 +0000 (07:06 +0000)]
and the series file
braam [Tue, 22 Jul 2003 07:05:32 +0000 (07:05 +0000)]
- 2.6.0-test1 patches
braam [Tue, 22 Jul 2003 07:05:00 +0000 (07:05 +0000)]
- 2.6.0 patches
zab [Tue, 22 Jul 2003 06:38:59 +0000 (06:38 +0000)]
Work-in-progress commit of the IO path refactoring.
- intitiate brw on pages by pushing down through the obd api
- add some data structures for passing and collecting the pages
- batch the pages into rpcs in the osc
- refactor ptlrpc a little to export some ptlrpc_set_wait internals
- temporarily build a
- add a kernel patch that adds page->private to 2.4
- split up some page cache stuff into 2.4 and 2.6 files
- make some progress in tearing out the 2.5 thread brw stuff
braam [Tue, 22 Jul 2003 05:43:03 +0000 (05:43 +0000)]
- remove old patches....E(inode->i_sb);
braam [Tue, 22 Jul 2003 05:30:45 +0000 (05:30 +0000)]
- chaos-2.4.18 patch files
braam [Tue, 22 Jul 2003 05:28:53 +0000 (05:28 +0000)]
- include a chaos-2.4.18 patch again.
adilger [Mon, 21 Jul 2003 23:08:57 +0000 (23:08 +0000)]
Move gigantic inline obdo_* functions to obdclass.
phil [Mon, 21 Jul 2003 16:31:36 +0000 (16:31 +0000)]
b=1505
- Demote some overly-common CERRORs to CDEBUGs
- Remove unused MDS-connection bits from filter_connect()
wangdi [Mon, 21 Jul 2003 13:46:36 +0000 (13:46 +0000)]
add some features for crash
phil [Sat, 19 Jul 2003 18:10:51 +0000 (18:10 +0000)]
Get patches out of my working tree; change to the ENABLE_ORPHANS constant
and fix a couple of warnings
phil [Fri, 18 Jul 2003 18:03:04 +0000 (18:03 +0000)]
More debug cleanup related to the subsystem; the old array is gone,
it's a mask just like the debug level, everything is simpler (and
works now, too).
phil [Fri, 18 Jul 2003 17:14:55 +0000 (17:14 +0000)]
b=1322
Commit the Lustre-side fix for the common "looping in revalidate after
recovery" case--unhash aliases before returning 0 in revalidate_it.
Also move d_unhash_aliases into llite/ and call it ll_unhash_aliases.
phil [Fri, 18 Jul 2003 16:42:00 +0000 (16:42 +0000)]
This fixes the bug that new S_* debug subsystems were not being
printed in debug logs. The problem was a desynchronization between
the kernel and userspace tools.
The old code would silently drop unknown subsystem messages. Now a
warning is printed and they are included.
The subsystems are no longer a crazy twist of left-shifted integers, but
a normal bitmask. There may have been a good reason for the craziness,
to try to fit a subsystem and type into one integer, but it's long gone.
adilger [Fri, 18 Jul 2003 08:38:22 +0000 (08:38 +0000)]
Increment objids on OST after a restart.
adilger [Fri, 18 Jul 2003 06:50:15 +0000 (06:50 +0000)]
Fix POSIX write/getattr failure.
adilger [Fri, 18 Jul 2003 01:55:53 +0000 (01:55 +0000)]
Commit small_write.c test.
Move munlink into installed files so we can use it to delete bad files.
mmex [Fri, 18 Jul 2003 01:28:37 +0000 (01:28 +0000)]
- update rh-2.4.20 configuration to compile correctly
adilger [Thu, 17 Jul 2003 21:07:19 +0000 (21:07 +0000)]
Set mtime and ctime on old inode if truncating by thread.
Also, mark the inodes dirty _after_ we modify them (done in orphan_add/del),
in case of crash before the real truncate is called.
I thought this might be the cause of bug 1530 (deleted files come back to
life), but that would not be the case even if there is a problem in the
delete thread as the name unlink is done before the delete thread gets
involved.
adilger [Thu, 17 Jul 2003 07:58:01 +0000 (07:58 +0000)]
Landing statfs cache code.
- preliminary for prealloc (bug 974)
- preliminary for QOS (bug 855)
- speeds up parallel df
- speeds up lprocfs
Lprocfs underwent a bunch of cleanup (mostly from Intel, partly from me)
to remove lots of macros and strangeness w.r.t. per-target statfs values.
Still open for discussion is how old the cached statfs values can be.
I arbitrarily picked 1 second for client statfs, lprocfs, and target
statfs caches. We may want to change that, or allow it to be tunable.
Note that this code is different than that in bug 855, in that the
max_age parameter is an absolute age in jiffies, whereas the patch
in bug 855 used a relative time in seconds. The thing I'm trying
to avoid is if we say "I want a result up to 1 second old", but we
get delayed during an RPC and we don't know what "1 second old" is
related to.
adilger [Thu, 17 Jul 2003 05:54:13 +0000 (05:54 +0000)]
Quiet some compiler warnings.
jacob [Thu, 17 Jul 2003 03:22:42 +0000 (03:22 +0000)]
Also include the gpl_header patch in rh-2.4.20 series.
mmex [Wed, 16 Jul 2003 23:51:34 +0000 (23:51 +0000)]
- updated series to apply cleanly. uses chaos series patchs instead of
vanilla ones.
adilger [Wed, 16 Jul 2003 22:59:35 +0000 (22:59 +0000)]
Misc changes from PNNL lester tree.
braam [Wed, 16 Jul 2003 22:40:45 +0000 (22:40 +0000)]
1486, 1456, 1546
- Fixes for these bugs included
- Includes substantial cleanup of lock teardown in llite.
- Ignore mountpoints for ll_revalidate_it
- Upgrade to uml-6
Fixes for
adilger [Wed, 16 Jul 2003 20:47:59 +0000 (20:47 +0000)]
With fixing of EA leak, sanity.sh and sanityN.sh both pass all the way through.
We also pass acceptance-small.sh all the way through (even on a single 72MB
UML) for THREADS=2 SIZE=10240, and will likely pass for the default if run on
a system with more memory.
Sadly (very sadly), nobody noticed that I broke sanityN.sh last week, until
I did this morning. Nobody ever noticed that "opendevunlink" was actually
compiled from opendirunlink.c until today (not me either).
Updated sanityN.sh to allow single tests to be run, like sanity.sh.
Added checks to both to ensure we have correct config for current mount.
It could be smarter about not needing a config if running with START=:
CLEAN=:, but I leave that for another day.
shaver [Wed, 16 Jul 2003 11:00:43 +0000 (11:00 +0000)]
fix GPL_HEADER in chaos-2.4.20
shaver [Wed, 16 Jul 2003 10:57:05 +0000 (10:57 +0000)]
patch: fix for totally unhelpful GPL_HEADER syntax error (chaos-2.4.20)
shaver [Wed, 16 Jul 2003 10:43:37 +0000 (10:43 +0000)]
Fix build bustage from SIGNAL_MASK_ASSERT. Works in chaos-2.4.20 (which
doesn't report as redhat-2.4.20, oddly), looks correct for 2.5/rh-2.4.20
as well.
shaver [Wed, 16 Jul 2003 09:50:53 +0000 (09:50 +0000)]
b=1534: Avoid g_f_w and friends for updating last_rcvd records, and instead
perform simple buffer-cache operations. Fixes an OST (and possible,
though rare and possibly never seen, MDS) deadlock case.
phil [Wed, 16 Jul 2003 06:21:36 +0000 (06:21 +0000)]
b=1511
Add a SIGNAL_MASK_ASSERT() function to try to nail down whether the
lock magic is already trashed before we enter Lustre
zab [Tue, 15 Jul 2003 23:02:59 +0000 (23:02 +0000)]
- teach the filter to flip write pages into the read cache after it has
written them with o_direct. passes sanity and doesn't even leak pages
or nothin'.
- exports some symbols for the page fliping to work with
- some minor fixes of the commitrw_write cleanup path, we should get more eyes
looking over this
shaver [Tue, 15 Jul 2003 08:54:14 +0000 (08:54 +0000)]
Land b_unify on b_devel.
Highlights include:
- intent rework from Peter and Phil
- unification of VFS intent strategy for 2.4 and 2.5/2.6
- select 2.5 kernel patches from b_ad
- all the whitespace fixes you can imagine.
Complete diff for sanity-checking at http://off.net/~shaver/unify-landing.diff.
braam [Tue, 15 Jul 2003 08:33:13 +0000 (08:33 +0000)]
- missing patches for b_unify from the 2.5 series
braam [Tue, 15 Jul 2003 08:27:57 +0000 (08:27 +0000)]
- add the unhash method to the kernel
braam [Tue, 15 Jul 2003 03:51:12 +0000 (03:51 +0000)]
- 2.5 kernel patches from ad for unify
mmex [Tue, 15 Jul 2003 02:42:06 +0000 (02:42 +0000)]
- add ext3-map_inode_page.patch to all series
- add ext3-error-export.patch to all series and add new export symbols that
will be needed in last_rcvd_records_via_metadata patch.
braam [Tue, 15 Jul 2003 00:01:21 +0000 (00:01 +0000)]
- remaining b_unify ---> b_ad merge
adilger [Mon, 14 Jul 2003 22:37:20 +0000 (22:37 +0000)]
Instead of setting all of the valid flags, and then clearing them for
regular files if an EA is found, we instead leave them unset for regular
files and only set them if we find we do not find an EA. This avoids
sending back bad data to the client if we abort early because of error
and do not clear the valid flags (POSIX.os/files/open/T.open 31).
shaver [Mon, 14 Jul 2003 22:32:59 +0000 (22:32 +0000)]
revert mmex's accidental checkin
mmex [Mon, 14 Jul 2003 22:22:28 +0000 (22:22 +0000)]
- cleanup come of the configuration files to be more compatible and
define some new things to help building rpms.
braam [Sat, 12 Jul 2003 21:41:29 +0000 (21:41 +0000)]
- add a chaos 2.4.20 patch series to b_intent
mmex [Sat, 12 Jul 2003 21:21:46 +0000 (21:21 +0000)]
- fix my typo
- include string.h so we don't get a warning everytime dcache.h is included,
which is often.
mmex [Sat, 12 Jul 2003 20:51:27 +0000 (20:51 +0000)]
- resolve conflict between 2.4.20 vanilla kernel and hp kernel. The XFS
patches cause us problems.
alex_tomas [Sat, 12 Jul 2003 18:30:39 +0000 (18:30 +0000)]
- ll_fill_super() calls lprocfs_register_mountpoint() _before_
*iget
- lprocfs_counter_add() shows backtrace in counter var isn't initialized yet
alex_tomas [Sat, 12 Jul 2003 09:51:50 +0000 (09:51 +0000)]
- inode-protection-from-pdflush.patch: an inode may have I_SKIP_PDFLUSH
flag which protects it from to be processed by pdflush's writeback
mechanism. this allows us to work locking problem with lustre_fwrite()
around
- MDS sets I_SKIP_PDFLUSH flag on 'last record' file
- filter_prep() sets I_SKIP_PDFLUSH flag
- inode-protection-from-pdflush.patch has been included in kgdb-2.5.73
series
alex_tomas [Sat, 12 Jul 2003 09:36:56 +0000 (09:36 +0000)]
- ext3-noread-inode.patch added