When creating a hard link to a file, the MDT/MDD/OSD code does
not verify whether the target link name already exists in the
directory. The ZFS ZAP code checks for duplicate entries. The
add_dirent_to_buf() function in ldiskfs only checks entries for
duplicates while it is traversing the leaf block looking for free
space. Even if it scanned the whole leaf block, this would not
work for non-htree directories since there is no guarantee that
the name is being inserted into the same leaf block.
To fix this, link should check target object doesn't exist as
other creat operations.
Add sanity.sh test_31o with multiple threads racing to link a new
name into the directory, while ensuring that there is a free entry
in the leaf block that is large enough to hold the duplicate name.
This needs to be racy, because otherwise the client VFS will see
the existing name and not send the RPC to the MDS, hiding the bug.
Add DLDLMRES/PLDLMRES macros for printing the whole lock resource
name (including the name hash) in LDLM_DEBUG() messages in a format
similar to DFID/PFID so they can be found in debug logs more easily.
The patch pickes client side change of the original patch, which only
contains the DLM printk part.
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-2901
Lustre-change: http://review.whamcloud.com/6591
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Lai Siyao <[email protected]>
Reviewed-by: Alex Zhuravlev <[email protected]>
Reviewed-by: Mike Pershin <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
When ptlrpc_start_thread fails to create a new thread, it will
finalize and free a struct ptlrpc_thread created and used here.
Considering this, it can be a problem when ptlrpc_svcpt_stop_thread
is driven and handles the struct ptlrpc_thread right after or right
before failure of cfs_create_thread. Because this situation let
the both of ptlrpc_start_thread and ptlrpc_svcpt_stop_threads
access the freed ptlrpc_thread and cause OS panic. Or, it may
happen that ptlrpc_svcpt_stop_threads waits forever holding an
already-freed waitq.
This patch adds an error handling into ptlrpc_start_thread to fix
this problem.
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-2889
Lustre-change: http://review.whamcloud.com/5552
Signed-off-by: Hiroya Nozaki <[email protected]>
Reviewed-by: Liang Zhen <[email protected]>
Reviewed-by: Nikitas Angelinas <[email protected]>
Reviewed-by: Keith Mannthey <[email protected]>
Reviewed-by: Oleg Drokin <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
* lustre/include/lustre_dlm.h: Remove all bit fields and the unused
weighing callback procedure. respell LDLM_AST_DISCARD_DATA as
LDLM_FL_AST_DISCARD_DATA to match other flags.
* .gitignore: ignore emacs temporary files
* autogen.sh: rebuild the lock bits, if autogen is available.
* contrib/bit-masks/lustre_dlm_flags.def: define the ldlm_lock flags
* contrib/bit-masks/lustre_dlm_flags.tpl: template for emitting text
* contrib/bit-masks/Makefile: construct the .c and .h files
The .c file is for constructing a crash extension and is not
preserved.
* contrib/bit-masks/.gitignore: ignore built products
* lustre/contrib/wireshark/packet-lustre.c: use built files instead
of local versions of the defines.
In the rest of the modified sources, replace flag field references
with bit mask references.
* lustre/osc/osc_lock.c: removed osc_lock_weigh, too
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-2771
Lustre-change: http://review.whamcloud.com/5312
Signed-off-by: Bruce Korb <[email protected]>
Reviewed-by: Andreas Dilger <[email protected]>
Reviewed-by: Oleg Drokin <[email protected]>
Reviewed-by: Keith Mannthey <[email protected]>
Reviewed-by: Keith Mannthey <[email protected]>
Reviewed-by: <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
It seems quota namespace is needlessly referenced on connect,
but that's not necessary as it could not go away until entire
obd goes away.
On the other hand this extra reference disturbs other logic
depending on empty namespace having zero refcount, so this patch
drops such extra referencing.
This picks client side change of the original patch.
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-2924
Lustre-change: http://review.whamcloud.com/6234
Signed-off-by: Oleg Drokin <[email protected]>
Reviewed-by: Niu Yawei <[email protected]>
Reviewed-by: Johann Lombardi <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
The main reason behind this is ldlm_poold walks all namespaces currently
no matter if there are any locks or not. On large systems this could take
quite a bit of time, esp. since ldlm_poold is currently woken up once per
second.
Now every time a client namespace loses it's last resource it is placed
into an inactive list that is not touched by ldlm_poold as pointless.
On creation of a first resource in a namespace it is placed back into
the active list.
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-2924
Lustre-change: http://review.whamcloud.com/5624
Signed-off-by: Oleg Drokin <[email protected]>
Reviewed-by: Hiroya Nozaki <[email protected]>
Reviewed-by: Niu Yawei <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Lustre puts system errors (e.g., ENOTCONN) on wire as numbers
essentially specific to senders' architectures. While this is fine
for x86-only sites, where receivers share the same error number
definition with senders, problems will arise, however, for sites
involving multiple architectures with different error number
definitions. For instance, an ENOTCONN reply from a sparc server will
be put on wire as -57, which, for an x86 client, means EBADSLT
instead.
To solve the problem, this patch defines a set of network errors for
on-wire or on-disk uses. These errors correspond to a subset of the
x86 system errors and share the same number definition, maintaining
compatibility with existing x86 clients and servers.
Then, either error numbers could be translated at run time, or all
host errors going on wire could be replaced with network errors in the
code. This patch does the former by introducing both generic and
field-specific translation routines and calling them at proper places,
so that translations for existing fields are transparent.
(Personally, I tend to think the latter way might be worthwhile, as it
is more straightforward conceptually. Do we really need so many
different errors? Should errors returned by kernel routines really be
passed up and eventually put on wire? There could even be security
implications in that.)
Thank Fujitsu for the original idea and their contributions that make
this available upstream.
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-2743
Lustre-change: http://review.whamcloud.com/5577
Signed-off-by: Li Wei <[email protected]>
Reviewed-by: Andreas Dilger <[email protected]>
Reviewed-by: Hiroya Nozaki <[email protected]>
Reviewed-by: Oleg Drokin <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
The race is result of use-after-free situation:
~ ptlrpc_stop_pinger() ~ ptlrpc_pinger_main()
---------------------------------------------------------------
thread_set_flags(SVC_STOPPING)
cfs_waitq_signal(pinger_thread) ...
... thread_set_flags(SVC_STOPPED)
l_wait_event(thread_is_stopped)
OBD_FREE_PTR(pinger_thread)
... cfs_waitq_signal(pinger_thread)
---------------------------------------------------------------
The memory used by pinger_thread might have been freed and
reallocated to something else, when ptlrpc_pinger_main()
used it in cvs_waitq_signal().
Signed-off-by: Li Wei <[email protected]>
Signed-off-by: Dmitry Eremin <[email protected]>
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-3032
Lustre-change: http://review.whamcloud.com/6040
Reviewed-by: Faccini Bruno <[email protected]>
Reviewed-by: Mike Pershin <[email protected]>
Reviewed-by: Andreas Dilger <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Print the namespace and OBD device name, as well as the first two
lock resource fields (typically the FID) if there is an error with
loading the object from disk. This will be more important with
FID-on-OST and also the MDS. Using fid_extract_from_res_name() isn't
possible in the LDLM code, since the lock resource may not be a FID.
Make fid_extract_quota_resid() argument order and name consistent
with other fid_*_res() functions, with FID first and resource second.
Fix a bug in ofd_lvbo_init() where NULL lvb is accessed on error.
Print FID in ofd_lvbo_update() CDEBUG() and CERROR() messages.
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-2193
Lustre-change: http://review.whamcloud.com/4501
Signed-off-by: Andreas Dilger <[email protected]>
Reviewed-by: Jinshan Xiong <[email protected]>
Reviewed-by: Bobi Jam <[email protected]>
Reviewed-by: Oleg Drokin <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Commit 4b5b4c7222 ("staging/lustre/libcfs: restore LINVRNT") added
"default false" to this Kconfig file. It was obviously meant to use
"default n" here. But we might as well drop this line, as a Kconfig bool
defaults to 'n' anyway.
Signed-off-by: Paul Bolle <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Confirmed by cscope that the functions are not used anymore. A fresh compilation does not yield any errors.
Signed-off-by: Dragos Foianu <[email protected]>
Reviewed-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
kmap_atomic allows only one argument now, just remove the second.
Signed-off-by: Zhao Hongjiang <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
This reverts commit 0ad1ea6954
I didn't use git revert because it can not be done cleanly.
Hopefully it will be the last time we do it...
Cc: Stephen Rothwell <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Building on 32bit system, I got warnings like below:
drivers/staging/lustre/lustre/llite/../include/lprocfs_status.h:666:7: note: expected ‘long unsigned int *’ but argument is of type ‘size_t *’
char *lprocfs_find_named_value(const char *buffer, const char *name,
drivers/staging/lustre/lustre/lov/lov_io.c: In function ‘lov_io_rw_iter_init’:
include/asm-generic/div64.h:43:28: warning: comparison of distinct pointer types lacks a cast [enabled by default]
(void)(((typeof((n)) *)0) == ((uint64_t *)0)); \
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
dump_trace() is only available on X86. Without it, Lustre's own
watchdog is broken. We can only dump current task's stack.
The client-side this code is much less likely to hit deadlocks and
it's probably OK to drop this altogether, since we hardly have any
ptlrpc threads on clients, most notable ones are ldlm cb threads
that should not really be blocking on the client anyway.
Remove libcfs watchdog for now, until the upstream kernel watchdog
can detect distributed deadlocks and dump other kernel threads.
Cc: Oleg Drokin <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
kuid_t/kgid_t are wrappered when CONFIG_UIDGID_STRICT_TYPE_CHECKS is on.
Lustre build is broken because we always treat them as plain __u32.
The patch fixes it. Internally, Lustre always use __u32 uid/gid, and
convert to kuid_t/kgid_t when necessary.
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Got below errors on s390 build:
CC [M] drivers/staging/lustre/lustre/llite/dir.o
drivers/staging/lustre/lustre/llite/dir.c: In function 'll_dir_filler':
drivers/staging/lustre/lustre/llite/dir.c:225:3: error: implicit declaration of function 'prefetchw' [-Werror=implicit-function-declaration]
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
As reported by Fengguang:
In file included from drivers/staging/lustre/lustre/obdclass/../include/lustre/lustre_idl.h:99:0,
from drivers/staging/lustre/lustre/obdclass/../include/lprocfs_status.h:46,
from drivers/staging/lustre/lustre/obdclass/../include/obd_support.h:42,
from drivers/staging/lustre/lustre/obdclass/../include/obd_class.h:40,
from drivers/staging/lustre/lustre/obdclass/lu_object.c:53:
drivers/staging/lustre/lustre/obdclass/../include/lustre/lustre_user.h:356:10: error: field 'lmd_st' has incomplete type
drivers/staging/lustre/lustre/obdclass/../include/lustre/lustre_user.h:361:10: error: field 'lmd_st' has incomplete type
Reported-by: Fengguang Wu <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Three functions cfs_cpu_ht_nsiblings, cfs_cpt_cpumask and
cfs_cpt_table_print are missing if !CONFIG_SMP.
cpumask_t/nodemask_t/__read_mostly/____cacheline_aligned
are redefined.
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Stephen Rothwell reported below error on powerpc:
In file included from drivers/staging/lustre/include/linux/libcfs/libcfs.h:203:0,
from drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd.h:67,
from drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd.c:41:
drivers/staging/lustre/lnet/klnds/o2iblnd/o2iblnd.c: In function 'kiblnd_dev_need_failover':
drivers/staging/lustre/include/linux/libcfs/libcfs_debug.h:215:16: error: implicit declaration of function 'NIPQUAD' [-Werror=implicit-function-declaration]
static struct libcfs_debug_msg_data msgdata; \
^
We should just remove HIPQUAD and replace it with %pI4h.
Reported-by: Stephen Rothwell <[email protected]>
Cc: Joe Perches <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Lustre internal dependency needs to be cleaned up. Currently,
libcfs is acting as a basis of all other modules, while other
modules in lustre/ directory in turn depend on lnet modules.
It creates a dependency loop that need to be fixed. Hopefully
we will remove libcfs in the end. So just disable buildin for
now.
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
If LNetNIInit() fails, we'll get zero ln_refcount. So fail
LNetGetId() properly instead of asserting.
We can get to it when socklnd fails to scan network interfaces,
which is possible if Lustre is builtin.
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
Change Makefiles to keep link order in match with Lustre module
dependency, so that when Lustre is built in kernel, we'll have
the same dependency. Otherwise we'll crash kernel if Lustre is
builtin due to missing internal dependency.
Reported-by: Fengguang Wu <[email protected]>
Signed-off-by: Peng Tao <[email protected]>
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
