Use the lnet_magic_accept() function to compare 'magic' against
LNET_PROTO_TCP_MAGIC for the appropriate string for an error message.
The original fix removed an unneeded byte-ordering cast because the define
was already CPU byte-ordered and it was assumed that 'magic' was CPU
byte-ordered, too.
Now modify the if-statement to use the appropriate lnet_accept_magic()
function in order to be consistent with similar tests. This will allow
the code to be consistent with the general understanding that 'magic'
should be in host-byte-order for the peer that sent the message.
Fixes: 80782927e3 ("staging: lustre: Fix unneeded byte-ordering cast")
Signed-off-by: Justin Skists <j.skists@gmail.com>
Reviewed-by: NeilBrown <neil@brown.name>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Fix sparse warning:
CHECK drivers/staging//lustre/lnet/lnet/acceptor.c
drivers/staging//lustre/lnet/lnet/acceptor.c:243:30: warning: cast to
restricted __le32
LNET_PROTO_TCP_MAGIC, as a define, is already CPU byte-ordered when
compared to 'magic', so no need for a cast.
Signed-off-by: Justin Skists <j.skists@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
One of Neil's recent cleanups apparently has led the code to get
to a state where gcc tracks the 'seqnr' variable just enough to
see that it is sometimes initialized in seq_client_alloc_seq(),
but not enough that it can prove this initialization to be reliable
before the use of that variable:
drivers/staging/lustre/lustre/fid/fid_request.c: In function 'seq_client_alloc_fid':
drivers/staging/lustre/lustre/fid/fid_request.c:245:22: error: 'seqnr' may be used uninitialized in this function [-Werror=maybe-uninitialized]
The code seems to be otherwise correct, and I could not come
up with a good way to simplify it further, so this adds a fake
initialization to shut up that warning.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
We now allow lustre to be built when CONFIG_MODULES is disabled,
but that causes a build failure:
In file included from drivers/staging/lustre/include/linux/libcfs/libcfs.h:42,
from drivers/staging/lustre/lustre/obdclass/lu_object.c:44:
drivers/staging/lustre/lustre/obdclass/lu_object.c: In function 'lu_context_key_degister':
drivers/staging/lustre/lustre/obdclass/lu_object.c:1410:51: error: dereferencing pointer to incomplete type 'struct module'
This particular case can be avoided by using the module_name()
macro that was designed exactly to handle printing the name of
a module in all configurations.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The kernel would like to have all stack VLA usage removed[1]. This switches
to a simple kasprintf() instead, and in the process fixes an off-by-one
between the allocation and the sprintf (allocation did not include NULL
byte in calculation).
[1] https://lkml.org/lkml/2018/3/7/621
Signed-off-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Remove yield() call. In this case it's use is considered broken, since
it is being assumed that yield() will let another process run that will
make the event true.
Signed-off-by: Thomas Avery <tavery321@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This reverts commit 16f1eeb660.
The reason for this patch was that lustre used copy_from_user_page.
Commit 76133e66b1 ("staging/lustre: Replace jobid acquiring with per
node setting") removed that usage.
So the arch restrictions can go.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Remove restriction the lustre must be built
as modules. It now works as a monolithic build.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
When the ptlrpc module is loaded, it starts the pinger thread and
calls LNetNIInit which starts various threads.
We don't need these threads until the module is actually being
used, such as when a lustre filesystem is mounted.
So move the thread creation into new ptlrpc_inc_ref() (modeled on
ptlrpcd_inc_ref()), and call that when needed, such as at mount time.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Rather than allocating a ptlrpc_thread for the
stat-ahead thread, just use the task_struct provided
by kthreads directly.
As nothing ever waits for the sai_task, it must call do_exit()
directly rather than simply return from the function.
Also it cannot use kthread_should_stop() to know when to stop.
There is one caller which can ask it to stop so we need a simple
signaling mechanism. I've chosen to set ->sai_task to NULL
when the thread should finish up. The thread notices this and
cleans up and exits.
lli_sa_lock is used to avoid races between waking up the process
and the process exiting.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Lustre has a 'struct ptlrpc_thread' which provides
control functionality wrapped around kthreads.
None of the functionality used in statahead.c requires
ptlrcp_thread - it can all be done directly with kthreads.
So discard the ptlrpc_thread and just use a task_struct directly.
One particular change worth noting is that in the current
code, the thread performs some start-up actions and then
signals that it is ready to go. In the new code, the thread
is first created, then the startup actions are perform, then
the thread is woken up. This means there is no need to wait
any more than kthread_create() already waits.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
lustre has a "Pinger" kthread which periodically pings peers
to ensure all hosts are functioning.
This can more easily be done using a work queue.
As maintaining contact with other peers is import for
keeping the filesystem running, and as the filesystem might
be involved in freeing memory, it is safest to have a
separate WQ_MEM_RECLAIM workqueue.
The SVC_EVENT functionality to wake up the thread can be
replaced with mod_delayed_work().
Also use round_jiffies_up_relative() rather than setting a
minimum of 1 second delay. The PING_INTERVAL is measured in
seconds so this meets the need is allow the workqueue to
keep wakeups synchronized.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The garbage collection for security contexts currently
has a dedicated kthread which wakes up every 30 minutes
to discard old garbage.
Replace this with a simple delayed_work item on the
system work queue.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
ldlm currenty has a kthread which wakes up every so often
and calls ldlm_pools_recalc().
The thread is started and stopped, but no other external interactions
happen.
This can trivially be replaced by a delayed_work if we have
ldlm_pools_recalc() reschedule the work rather than just report
when to do that.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
obdclass currently maintains two lists of data structures
(imports and exports), and a kthread which will free
anything on either list. The thread is woken whenever
anything is added to either list.
This is exactly the sort of thing that workqueues exist for.
So discard the zombie kthread and the lists and locks, and
create a single workqueue. Each obd_import and obd_export
gets a work_struct to attach to this workqueue.
This requires a small change to import_sec_validate_get()
which was testing if an obd_import was on the zombie
list. This cannot have every safely found it to be
on the list (as it could be freed asynchronously)
so it must be dead code.
We could use system_wq instead of creating a dedicated
zombie_wq, but as we occasionally want to flush all pending
work, it is a little nicer to only have to wait for our own
work items.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
These allocations are performed during initialization,
so they don't need GFP_NOFS.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
When the 'lustre' module is loaded, it gets a list of
net devices and uses the node ids to add entropy
to the prng. This means that the network interfaces need
to be configured before the module is loaded, which prevents
the module from being compiled into a monolithic kernel.
So move this entropy addition to the moment when
the interface is imported to LNet and the node id is first known.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
ln_nportals should be zero when no portals have
been allocated. This ensures that memory allocation failure
is handled correctly elsewhere.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Some places in lu_object.c allow lct_owner to be NULL, implying
that the code is built in to the kernel (not a module), but
two places don't. This prevents us from building lustre into
the kernel.
So remove the requirement and always allow lct_owner to be NULL.
This requires removing an "assert" that the module count is positive,
but this is redundant as module_put() already does the necessary test.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Including agl_list_empty() in the wait_event_idle() condition
is pointless as the body of the loop doesn't do anything
about the agl list.
So if the list wasn't empty, the while loop would spin
indefinitely.
The test was removed in the lustre-release commit
672ab0e00d61 ("LU-3270 statahead: small fixes and cleanup"),
but not in the Linux commit 5231f7651c ("staging: lustre:
statahead: small fixes and cleanup").
Fixes: 5231f7651c ("staging: lustre: statahead: small fixes and cleanup")
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The lustre-release patch commit bdc5bb52c554 ("LU-4933 osc:
Automatically increase the max_dirty_mb") changed
- if (cli->cl_dirty + PAGE_CACHE_SIZE <= cli->cl_dirty_max &&
+ if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
When this patch landed in Linux a couple of years later, it landed as
- if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
+ if (cli->cl_dirty_pages <= cli->cl_dirty_max_pages &&
which is clearly different ('<=' vs '<'), and allows cl_dirty_pages to
increase beyond cl_dirty_max_pages - which causes a latter assertion
to fails.
Fixes: 3147b26840 ("staging: lustre: osc: Automatically increase the max_dirty_mb")
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit 4f016420d3 ("Staging: lustre: obdclass: Use kasprintf") moved
some sprintf() calls earlier in the code to combine them with
memory allocation and create kasprintf() calls.
In one case, this code movement moved the sprintf to a location where the
values being formatter were different.
In particular
sprintf(niduuid, "%s_%x", mgcname, i);
was move from *after* the line
i = 0;
to a location where the value of 'i' was at least 1.
This cause the wrong name to be formatted, and triggers
CERROR("del MDC UUID %s failed: rc = %d\n",
niduuid, rc);
at unmount time.
So use '0' instead of 'i'.
Fixes: 4f016420d3 ("Staging: lustre: obdclass: Use kasprintf")
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Remove else after a return statement as it is not useful. Issue found
using checkpatch.
Signed-off-by: Santha Meena Ramamoorthy <santhameena13@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Replace 'unsigned' with 'unsigned int' to improve readability.
Issues found with checkpatch.pl
Signed-off-by: Dafna Hirschfeld <dafna3@gmail.com>
Acked-by: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Replace 'const char*' arrays with 'const char * const'
since the values in the arrays are not changed.
Issues found with checkpatch.pl
Signed-off-by: Dafna Hirschfeld <dafna3@gmail.com>
Acked-by: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Replace comparison to NULL with a 'not' operator.
Issue found with checkpatch.
Signed-off-by: Dafna Hirschfeld <dafna3@gmail.com>
Acked-by: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Replace "dont" with "don't".
"Dont" is not same as "Do not" or "Don't".
Signed-off-by: Arushi Singhal <arushisinghal19971997@gmail.com>
Acked-by: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Add trailing */ on a separate line for block comments to conform to
the Linux kernel coding style. Issue found using checkpatch.
Signed-off-by: Santha Meena Ramamoorthy <santhameena13@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
If lstcon_test_add sets 'ret' (passed by reference) to 1,
then lst_test_add_ioctl() ignores the return value.
This isn't justified - the return value must be zero for 'ret'
to be meaningful.
Signed-off-by: NeilBrown <neilb@suse.com>
Reviewed-by: James Simmons <jsimmons@infradead.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit 8f18c8a48b ("staging: lustre: lmv: separate master object
with master stripe") changed how lmo_root inodes were managed,
particularly when LMV_HASH_FLAG_MIGRATION is not set.
Previously lsm_md_oinfo[0].lmo_root was always a borrowed
inode reference and didn't need to by iput().
Since the change, that special case only applies when
LMV_HASH_FLAG_MIGRATION is set
In the upstream (lustre-release) version of this patch [Commit
60e07b972114 ("LU-4690 lod: separate master object with master
stripe")] the for loop in the lmv_unpack_md() was changed to count
from 0 and to ignore entry 0 if LMV_HASH_FLAG_MIGRATION is set.
In the patch that got applied to Linux, that change was missing,
so lsm_md_oinfo[0].lmo_root is never iput().
This results in a "VFS: Busy inodes" warning at unmount.
Fixes: 8f18c8a48b ("staging: lustre: lmv: separate master object with master stripe")
Signed-off-by: NeilBrown <neilb@suse.com>
Reviewed-by: James Simmons <jsimmons@infradead.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
lov - the logical object volume manager - is responsible for
striping data across multiple volumes.
So when it is given a request, it creates one or more
sub-requests, one for each target volume. Each sub_io
request has a sub_env environment which it operates in.
When lov_io_data_version_end() calls lov_io_end_wrapper() to
wait for and close off a sub_io, it passes the wrong
environment.
This causes an LINVRNT() to fail in cl2osc_io(), and may
cause other problems.
This patch changes the call to use ->sub_env, much like
other code in the same file.
Fixes: f0cf21abcc ("staging: lustre: clio: add CIT_DATA_VERSION and remove IOC_LOV_GETINFO")
Signed-off-by: NeilBrown <neilb@suse.com>
Reviewed-by: James Simmons <jsimmons@infradead.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
We should just return directly if memdup_user() fails. The current code
tries to free "param" which is an error pointer so it will Oops.
Fixes: 2baddf262e ("staging: lustre: use memdup_user to allocate memory and copy from user")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
ksnc_rx_iov_space is currently a union of two arrays,
one of 'struct kvec', the other of 'struct bio_vec'.
The 'struct bio_vec' option is never used. The
array of kvec is used to read in a packet header, or
to read data that needs to be skipped so as to synchronize
with a packet boundary.
In each case the target memory location is a virtual address,
never a page, so 'struct bio_vec' is never needed.
When we read into a page, different code steps up a separate
array of 'struct bio_vec'.
So remove the bio_vec option, and remove the union ksock_rxiovspace..
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
When fid fetches a new range from the server, it commits
to it (*output = *out) *before* performing sanity checks.
This looks backwards.
Don't commit to a value until it has been found to be sane.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
lcs_space can change while the lock is not held
if an RPC in underway. This can be detected by
seq->ls_update being set.
In this case, reading or writing the value should return
-EBUSY.
Also, the D_INFO CDEBUG() which reports the lcs_space being
updated never fires, as it tests the wrong value -
ldebugfs_fid_write_common() returns 'count' on success.
Finally, this return value should be returned from
ldebugfs_fid_space_seq_write(), rather than always returning 'count',
so that errors can be detected.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
seq_fid_alloc_fini() is tiny and only called
from two places in the one function. We can move
both those calls earlier and merge them so only
one call is needed. At that point, there is no
value added by having a separate function.
Also instead of using ++ and -- on ->lcs_update to
toggle between 0 and 1, explicitly set to 0 or 1
as appropriate.
Moving the locking earlier means that the code which updates
seq->lcs_fid is now protected, so
ldebugfs_fid_fid_seq_show() now cannot see a torn value.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Rather than open-coding a wait event loop twice,
use wait_event_cmd() to wait, dropping the spinlock
over schedule().
This does require duplicating part of the wait
condition, but that is just three tests on values that
are in registers or in cache, so the cost is small
and the increased readability is large.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
There is only one place where this lock is held
while the task might sleep - in
ldebugfs_fid_space_seq_write()
while ldebugfs_fid_write_common() is called.
This call can easily be taken out of the locked region
by asking it to parse the user data into a local variable,
and then copying that variable into ->lcs_space while
holding the lock.
Note that ldebugfs_gid_write_common returns >0 on
success, so use that to gate updating ->lcs_space.
So make that change, and convert lcs_mutex to a spinlock
named lcs_lock. spinlocks are slightly cheaper than mutexes
and using one makes is clear that the lock is only held for
a short time.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
1/ use list_for_each_entry_safe() instead of
list_for_each_safe() and similar.
2/ use list_first_entry() and list_last_entry() where appropriate.
3/ When removing everything from a list, use
while ((x = list_first_entry_or_null()) {
as it makes the intent clear
4/ No need to take a spinlock in a structure that is about
to be freed - we must have exclusive access at this stage.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Possibly the most interesting is the for-loop with no body.
Rearranging and initializing end_dirent on each iteration of
the outer while, makes the intent clearer.
Reviewed-by: "Eremin, Dmitry" <dmitry.eremin@intel.com>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
There is no "struct cfs_crypto_hash_desc" structure. There
are only pointers to this structure, which are cast back and
forth to struct ahash_request.
So discard cfs_crypto_hash_desc, and just use ahash_request directly.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The only interesting difference between libcfs_kvzalloc()
and kvzalloc() is that the former appears to work
with GFP_NOFS, which the latter gives a WARN_ON_ONCE()
when that is attempted.
Each libcfs_kvzalloc() should really be analysed
and either converted to a kzalloc() call if the size is never
more than a page, or to use GFP_KERNEL if no locks are held.
If there is ever a case where locks are held and a large allocation
is needed, then some other technique should be used.
It might be nice to not always blindly zero pages too...
For now, just convert libcfs_kvzalloc() calls to
kvzalloc(), and let the warning remind us that there is work to do.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Using vmalloc with GFP_NOFS is not supported as vmalloc
performs some internal allocations with GFP_KERNEL.
So in cases where the size passed to libcfs_kvzalloc()
is clearly at most 1 page, convert to kzalloc().
In cases where the call clearly doesn't hold any
filesystem locks, convert to GFP_KERNEL.
Unfortunately there are many more that are not easy to fix.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This library code is unnecessarily generic, but also
not generic enough. Library code that performs
allocations should always take a gfp_flags argument.
So discard the library and in the one file where it is used,
just use kzalloc or krealloc as needed.
In this context, it is clear that vmalloc is never needed.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This function is used precisely once, and is sufficiently
trivial that it may as well be open-coded.
Doing so helpfully highlights the similarity
between the new kvzalloc_node() call and the already existing
kzalloc_node() call in the same function.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
cfs_block_sigsinv() and cfs_restore_sigs() are now
simple enough to inline them.
This means we can discard linux-prim.c
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
According to comment for set_current_blocked() in
kernel/signal.c, changing ->blocked directly is wrong.
sigprocmask() should be called instead.
So change cfs_block_sigsinv() and cfs_restore_sigs()
to use sigprocmask().
For consistency, change them to pass the sigset_t by reference
rather than by value.
Also fix cfs_block_sigsinv() so that it correctly blocks
signals above 32 on a 32bit host.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Justin Skists