When a lot of IO, with some distributed mmap IO, is run on a GFS2 filesystem in
a cluster, it will deadlock. The reason is that do_no_page() will repeatedly
call gfs2_sharewrite_nopage(), because each node keeps giving up the glock
too early, and is forced to call unmap_mapping_range(). This bumps the
mapping->truncate_count sequence count, forcing do_no_page() to retry. This
patch institutes a minimum glock hold time a tenth a second. This insures
that even in heavy contention cases, the node has enough time to get some
useful work done before it gives up the glock.
A second issue is that when gfs2_glock_dq() is called from within a page fault
to demote a lock, and the associated page needs to be written out, it will
try to acqire a lock on it, but it has already been locked at a higher level.
This patch puts makes gfs2_glock_dq() use the work queue as well, to avoid this
issue. This is the same patch as Steve Whitehouse originally proposed to fix
this issue, execpt that gfs2_glock_dq() now grabs a reference to the glock
before it queues up the work on it.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
We only need a single gfs2_scand process rather than the one
per filesystem which we had previously. As a result the parameter
determining the frequency of gfs2_scand runs becomes a module
parameter rather than a mount parameter as it was before.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch passes all my nasty tests that were causing the code to
fail under one circumstance or another. Here is a complete summary
of all changes from today's git tree, in order of appearance:
1. There are now separate variables for metadata buffer accounting.
2. Variable sd_log_num_hdrs is no longer needed, since the header
accounting is taken care of by the reserve/refund sequence.
3. Fixed a tiny grammatical problem in a comment.
4. Added a new function "calc_reserved" to calculate the reserved
log space. This isn't entirely necessary, but it has two benefits:
First, it simplifies the gfs2_log_refund function greatly.
Second, it allows for easier debugging because I could sprinkle the
code with calls to this function to make sure the accounting is
proper (by adding asserts and printks) at strategic point of the code.
5. In log_pull_tail there apparently was a kludge to fix up the
accounting based on a "pull" parameter. The buffer accounting is
now done properly, so the kludge was removed.
6. File sync operations were making a call to gfs2_log_flush that
writes another journal header. Since that header was unplanned
for (reserved) by the reserve/refund sequence, the free space had
to be decremented so that when log_pull_tail gets called, the free
space is be adjusted properly. (Did I hear you call that a kludge?
well, maybe, but a lot more justifiable than the one I removed).
7. In the gfs2_log_shutdown code, it optionally syncs the log by
specifying the PULL parameter to log_write_header. I'm not sure
this is necessary anymore. It just seems to me there could be
cases where shutdown is called while there are outstanding log
buffers.
8. In the (data)buf_lo_before_commit functions, I changed some offset
values from being calculated on the fly to being constants. That
simplified some code and we might as well let the compiler do the
calculation once rather than redoing those cycles at run time.
9. This version has my rewritten databuf_lo_add function.
This version is much more like its predecessor, buf_lo_add, which
makes it easier to understand. Again, this might not be necessary,
but it seems as if this one works as well as the previous one,
maybe even better, so I decided to leave it in.
10. In databuf_lo_before_commit, a previous data corruption problem
was caused by going off the end of the buffer. The proper solution
is to have the proper limit in place, rather than stopping earlier.
(Thus my previous attempt to fix it is wrong).
If you don't wrap the buffer, you're stopping too early and that
causes more log buffer accounting problems.
11. In lops.h there are two new (previously mentioned) constants for
figuring out the data offset for the journal buffers.
12. There are also two new functions, buf_limit and databuf_limit to
calculate how many entries will fit in the buffer.
13. In function gfs2_meta_wipe, it needs to distinguish between pinned
metadata buffers and journaled data buffers for proper journal buffer
accounting. It can't use the JDATA gfs2_inode flag because it's
sometimes passed the "real" inode and sometimes the "metadata
inode" and the inode flags will be random bits in a metadata
gfs2_inode. It needs to base its decision on which was passed in.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Under certain circumstances its possible (though rather unlikely) that
inodes which were unlinked by one node while still open on another might
get "lost" in the sense that they don't get deallocated if the node
which held the inode open crashed before it was unlinked.
This patch adds the recovery code which allows automatic deallocation of
the inode if its found during block allocation (the sensible time to
look for such inodes since we are scanning the rgrp's bitmaps anyway at
this time, so it adds no overhead to do this).
Since the inode will have had its i_nlink set to zero, all we need to
trigger recovery is a lookup and an iput(), and the normal deallocation
code takes care of the rest.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch fixes some sign issues which were accidentally introduced
into the quota & statfs code during the endianess annotation process.
Also included is a general clean up which moves all of the _host
structures out of gfs2_ondisk.h (where they should not have been to
start with) and into the places where they are actually used (often only
one place). Also those _host structures which are not required any more
are removed entirely (which is the eventual plan for all of them).
The conversion routines from ondisk.c are also moved into the places
where they are actually used, which for almost every one, was just one
single place, so all those are now static functions. This also cleans up
the end of gfs2_ondisk.h which no longer needs the #ifdef __KERNEL__.
The net result is a reduction of about 100 lines of code, many functions
now marked static plus the bug fixes as mentioned above. For good
measure I ran the code through sparse after making these changes to
check that there are no warnings generated.
This fixes Red Hat bz #239686
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch fixes an error in the quota code where a 'struct
gfs2_quota_lvb*' was being passed to gfs2_adjust_quota() instead of a
'struct gfs2_quota_data*'. Also moved 'struct gfs2_quota_lvb' from
fs/gfs2/incore.h to include/linux/gfs2_ondisk.h as per Steve's suggestion.
Signed-off-by: Abhijith Das <adas@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch cleans up the inode number handling code. The main difference
is that instead of looking up the inodes using a struct gfs2_inum_host
we now use just the no_addr member of this structure. The tests relating
to no_formal_ino can then be done by the calling code. This has
advantages in that we want to do different things in different code
paths if the no_formal_ino doesn't match. In the NFS patch we want to
return -ESTALE, but in the ->lookup() path, its a bug in the fs if the
no_formal_ino doesn't match and thus we can withdraw in this case.
In order to later fix bz #201012, we need to be able to look up an inode
without knowing no_formal_ino, as the only information that is known to
us is the on-disk location of the inode in question.
This patch will also help us to fix bz #236099 at a later date by
cleaning up a lot of the code in that area.
There are no user visible changes as a result of this patch and there
are no changes to the on-disk format either.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The patch below consists of the following changes (in code order):
1. I fixed a minor compiler warning regarding the printing of
a kernel symbol address.
2. I implemented a suggestion from Dave Teigland that moves
the debugfs information for gfs2 into a subdirectory so
we can easily expand our use of debugfs in the future.
The current code keeps the glock information in:
/debug/gfs2/<fs>
With the patch, the new code keeps the glock information in:
/debug/gfs2/<fs>/glock
That will allow us to create more debugfs files in the future.
3. This fixes a bug whereby a failed mount attempt causes the
debugfs file to not be deleted. Failed mount attempts should
always clean up after themselves, including deleting the
debugfs file and/or directory.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
In Testing the previously posted and accepted patch for
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=228540
I uncovered some gfs2 badness. It turns out that the current
gfs2 code saves off a process pointer when glocks is taken
in both the glock and glock holder structures. Those
structures will persist in memory long after the process has
ended; pointers to poisoned memory.
This problem isn't caused by the 228540 fix; the new capability
introduced by the fix just uncovered the problem.
I wrote this patch that avoids saving process pointers
and instead saves off the process pid. Rather than
referencing the bad pointers, it now does process lookups.
There is special code that makes the output nicer for
printing holder information for processes that have ended.
This patch also adds a stub for the new "sprint_symbol"
function that exists in Andrew Morton's -mm patch set, but
won't go into the base kernel until 2.6.22, since it adds
functionality but doesn't fix a bug.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch prevents the printing of a warning message in cases where
the fs is functioning normally by handing off responsibility for
unlinked, but still open inodes, to another node for eventual deallocation.
Also, there is now an improved system for ensuring that such requests
to other nodes do not get lost. The callback on the iopen lock is
only ever called when i_nlink == 0 and when a node is unable to deallocate
it due to it still being in use on another node. When a node receives
the callback therefore, it knows that i_nlink must be zero, so we mark
it as such (in gfs2_drop_inode) in order that it will then attempt
deallocation of the inode itself.
As an additional benefit, queuing a demote request no longer requires
a memory allocation. This simplifies the code for dealing with gfs2_holders
as it removes one special case.
There are two new fields in struct gfs2_glock. gl_demote_state is the
state which the remote node has requested and gl_demote_time is the
time when the request came in. Both fields are only valid when the
GLF_DEMOTE flag is set in gl_flags.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The attached patch resolves bz 228540. This adds the capability
for gfs2 to dump gfs2 locks through the debugfs file system.
This used to exist in gfs1 as "gfs_tool lockdump" but it's missing from
gfs2 because all the ioctls were stripped out. Please see the bugzilla
for more history about the fix. This patch is also attached to the bugzilla
record.
The patch is against Steve Whitehouse's latest nmw git tree kernel
(2.6.21-rc1) and has been tested on system trin-10.
Signed-off-by: Robert Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch doesn't make any changes to the ordering of the various
operations related to glocking, but it does tidy up the calls to the
glops.c functions to make the structure more obvious.
The two functions: gfs2_glock_xmote_th() and gfs2_glock_drop_th() can be
made static within glock.c since they are called by every set of glock
operations. The xmote_th and drop_th glock operations are then made
conditional upon those two routines existing and called from the
previously mentioned functions in glock.c respectively.
Also it can be seen that the go_sync operation isn't needed since it can
easily be replaced by calls to xmote_bh and drop_bh respectively. This
results in no longer (confusingly) calling back into routines in glock.c
from glops.c and also reducing the glock operations by one member.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The "greedy" code was an attempt to retain glocks for a minimum length
of time when they relate to mmap()ed files. The current implementation
of this feature is not, however, ideal in that it required allocating
memory in order to do this and its overly complicated.
It also misses the mark by ignoring the other I/O operations which are
just as likely to suffer from the same problem. So the plan is to remove
this now and then add the functionality back as part of the glock state
machine at a later date (and thus take into account all the possible
users of this feature)
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Here is something I spotted (while looking for something entirely
different) the other day.
Rather than using a completion in each and every struct gfs2_holder,
this removes it in favour of hashed wait queues, thus saving a
considerable amount of memory both on the stack (where a number of
gfs2_holder structures are allocated) and in particular in the
gfs2_inode which has 8 gfs2_holder structures embedded within it.
As a result on x86_64 the gfs2_inode shrinks from 2488 bytes to
1912 bytes, a saving of 576 bytes per inode (no thats not a typo!).
In actual practice we get a much better result than that since
now that a gfs2_inode is under the 2048 byte barrier, we get two
per 4k slab page effectively halving the amount of memory required
to store gfs2_inodes.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This removes the extra filldir callback which gfs2 was using to
enclose an attempt at readahead for inodes during readdir. The
code was too complicated and also hurts performance badly in the
case that the getdents64/readdir call isn't being followed by
stat() and it wasn't even getting it right all the time when it
was.
As a result, on my test box an "ls" of a directory containing 250000
files fell from about 7mins (freshly mounted, so nothing cached) to
between about 15 to 25 seconds. When the directory content was cached,
the time taken fell from about 3mins to about 4 or 5 seconds.
Interestingly in the cached case, running "ls -l" once reduced the time
taken for subsequent runs of "ls" to about 6 secs even without this
patch. Now it turns out that there was a special case of glocks being
used for prefetching the metadata, but because of the timeouts for these
locks (set to 10 secs) the metadata was being timed out before it was
being used and this the prefetch code was constantly trying to prefetch
the same data over and over.
Calling "ls -l" meant that the inodes were brought into memory and once
the inodes are cached, the glocks are not disposed of until the inodes
are pushed out of the cache, thus extending the lifetime of the glocks,
and thus bringing down the time for subsequent runs of "ls"
considerably.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The go_sync callback took two flags, but one of them was set on every
call, so this patch removes once of the flags and makes the previously
conditional operations (on this flag), unconditional.
The go_inval callback took three flags, each of which was set on every
call to it. This patch removes the flags and makes the operations
unconditional, which makes the logic rather more obvious.
Two now unused flags are also removed from incore.h.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This shrinks the size of the gfs2_inode by 8 bytes by
replacing the version counter with a one bit valid/invalid
flag.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Annotated scalar fields, dropped unused ones. Note that
it's not at all obvious that we want to convert all of them
to host-endian...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Fix a bug in the directory reading code, where we might have dereferenced
a NULL pointer in case of OOM. Updated the directory code to use the new
& improved version of gfs2_meta_ra() which now returns the first block
that was being read. Previously it was releasing it requiring following
code to grab the block again at each point it was called.
Also turned off readahead on directory lookups since we are reading a
hash table, and therefore reading the entries in order is very
unlikely. Readahead is still used for all other calls to the
directory reading function (e.g. when growing the hash table).
Removed the DIO_START constant. Everywhere this was used, it was
used to unconditionally start i/o aside from a couple of places, so
I've removed it and made the couple of exceptions to this rule into
separate functions.
Also hunted through the other DIO flags and removed them as arguments
from functions which were always called with the same combination of
arguments.
Updated gfs2_meta_indirect_buffer to be a bit more efficient and
hopefully also be a bit easier to read.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Use atomic_t as the ref count in glocks rather than a kref.
This is another step towards using RCU for the glock hash.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This results in smaller list heads, so that we can have more chains
in the same amount of memory (twice as many). I've multiplied the
size of the table by four though - this is because we are saving
memory by not having one lock per chain any more. So we land up
using about the same amount of memory for the hash table as we
did before I started these changes, the difference being that we
now have four times as many hash chains.
The reason that I say "about the same amount of memory" is that the
actual amount now depends upon the NR_CPUS and some of the config
variables, so that its not exact and in some cases we do use more
memory. Eventually we might want to scale the hash table size
according to the size of physical ram as measured on module load.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The existing implementation of this function in glock.c was not
very efficient as it relied upon keeping a cursor element upon the
hash chain in question and moving it along. This new version improves
upon this by using the current element as a cursor. This is possible
since we only look at the "next" element in the list after we've
taken the read_lock() subsequent to calling the examiner function.
Obviously we have to eventually drop the ref count that we are then
left with and we cannot do that while holding the read_lock, so we
do that next time we drop the lock. That means either just before
we examine another glock, or when the loop has terminated.
The new implementation has several advantages: it uses only a
read_lock() rather than a write_lock(), so it can run simnultaneously
with other code, it doesn't need a "plug" element, so that it removes
a test not only from this list iterator, but from all the other glock
list iterators too. So it makes things faster and smaller.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Make the number of locks used for hash chains in glock.c
proportional to NR_CPUS. Also move constants for the number
of hash chains into glock.c from incore.h since they are
not used outside of glock.c.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This splits the rwlocks guarding the hash chains of the glock hash
table into their own array. This will reduce memory usage in some
cases due to better alignment, although the real reason for doing it
is to allow the two tables to be different sizes in future (i.e.
the locks will be sized proportionally with the max number of CPUs
and the hash chains sized proportinally with the size of physical memory)
In order to allow this, the gl_bucket member of struct gfs2_glock has
now become gl_hash, so we record the hash rather than a pointer to the
bucket itself.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
As requested by Jan Engelhardt, this removes the typedefs in the
locking module interface and replaces them with void *. Also
since we are changing the interface, I've added a few consts
as well.
Cc: Jan Engelhardt <jengelh@linux01.gwdg.de>
Cc: David Teigland <teigland@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
There are several reasons why we want to do this:
- Firstly its large and thus we'll scale better with multiple
GFS2 fs mounted at the same time
- Secondly its easier to scale its size as required (thats a plan
for later patches)
- Thirdly, we can use kzalloc rather than vmalloc when allocating
the superblock (its now only 4888 bytes)
- Fourth its all part of my plan to eventually be able to use RCU
with the glock hash.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
As per Jan Engelhardt's comments, this should make all the headers
compile on their own by including and/or declaring structures
early.
Cc: Jan Engelhardt <jengelh@linux01.gwdg.de>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This makes all fixed size types have consistent names.
Cc: Jan Engelhardt <jengelh@linux01.gwdg.de>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
As per comments from Jan Engelhardt <jengelh@linux01.gwdg.de> this
updates the copyright message to say "version" in full rather than
"v.2". Also incore.h has been updated to remove forward structure
declarations which are not required.
The gfs2_quota_lvb structure has now had endianess annotations added
to it. Also quota.c has been updated so that we now store the
lvb data locally in endian independant format to avoid needing
a structure in host endianess too. As a result the endianess
conversions are done as required at various points and thus the
conversion routines in lvb.[ch] are no longer required. I've
moved the one remaining constant in lvb.h thats used into lm.h
and removed the unused lvb.[ch].
I have not changed the HIF_ constants. That is left to a later patch
which I hope will unify the gh_flags and gh_iflags fields of the
struct gfs2_holder.
Cc: Jan Engelhardt <jengelh@linux01.gwdg.de>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Remove the unused sync feature from glocks. This is currently done by
calling the required functions to sync pages/blocks directly so this
code isn't needed.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
For all the usual reasons of enforcing correctness and potentially
reducing code size, this patch makes the glock operations const.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch allows the simultaneous mounting of gfs2meta and gfs2
filesystems. A restriction however is that a gfs2meta fs may only be
mounted if its corresponding gfs2 filesystem is also mounted. Also, a
gfs2 filesystem cannot be unmounted before its gfs2meta filesystem.
Signed-off-by: Abhijith Das <adas@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
I noticed the gfs2_scand seemed to be taking a lot of CPU,
so in order to cut that down a bit, here is a patch. Firstly
the type of a glock is a constant during its lifetime, so that
its possible to check this without needing locking. I've moved
the (common) case of testing for an inode glock outside of
the glmutex lock.
Also there was a mutex left over from when the glock cache was
master of the inode cache. That isn't required any more so I've
removed that too.
There is probably scope for further speed ups in the future
in this area.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Mmapped files were able to trigger a lock ordering bug. Private
maps do not need to take the glock so early on. Shared maps do
unfortunately, however we can get around that by adding a flag
into the flags for the struct gfs2_file. This only works because
we are taking an exclusive lock at this point, so we know that
nobody else can be racing with us.
Fixes Red Hat bugzilla: #201196
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
fs/gfs2/glock.c: In function ‘gfs2_holder_get’:
fs/gfs2/glock.c:439: warning: passing argument 2 of ‘set_bit’ from incompatible pointer type
fs/gfs2/glock.c: In function ‘rq_promote’:
fs/gfs2/glock.c:512: warning: passing argument 2 of ‘set_bit’ from incompatible pointer type
fs/gfs2/glock.c:526: warning: passing argument 2 of ‘set_bit’ from incompatible pointer type
...
Signed-off-by: David Woodhouse <dwmw2@infradead.org>
This patch fixes the way we have been dealing with unlinked,
but still open files. It removes all limits (other than memory
for inodes, as per every other filesystem) on numbers of these
which we can support on GFS2. It also means that (like other
fs) its the responsibility of the last process to close the file
to deallocate the storage, rather than the person who did the
unlinking. Note that with GFS2, those two events might take place
on different nodes.
Also there are a number of other changes:
o We use the Linux inode subsystem as it was intended to be
used, wrt allocating GFS2 inodes
o The Linux inode cache is now the point which we use for
local enforcement of only holding one copy of the inode in
core at once (previous to this we used the glock layer).
o We no longer use the unlinked "special" file. We just ignore it
completely. This makes unlinking more efficient.
o We now use the 4th block allocation state. The previously unused
state is used to track unlinked but still open inodes.
o gfs2_inoded is no longer needed
o Several fields are now no longer needed (and removed) from the in
core struct gfs2_inode
o Several fields are no longer needed (and removed) from the in core
superblock
There are a number of future possible optimisations and clean ups
which have been made possible by this patch.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
We can reclaim some space by moving fields in some structures
in order to allow them to pack better on 64 bit architectures.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This adds some extra debugging to glock.c and changes
inode.c's deallocation code to call the debugging code
at a suitable moment. I'm chasing down a particular bug
to do with deallocation at the moment and the code can
go again once the bug is fixed.
Also this includes the first part of some changes to unify
the Linux struct inode and GFS2's struct gfs2_inode. This
transformation will happen in small parts over the next short
period.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
