When we queue data buffers for ordered write, the buffers are added
to the head of the ordered write list. When the log needs to push
these buffers to disk, it also walks the list from the head. The
result is that the the ordered buffers are submitted to disk in
reverse order.
For large writes, this means that whenever the log flushes large
streams of reverse sequential order buffers are pushed down into the
block layers. The elevators don't handle this particularly well, so
IO rates tend to be significantly lower than if the IO was issued in
ascending block order.
Queue new ordered buffers to the tail of the ordered buffer list to
ensure that IO is dispatched in the order it was submitted. This
should significantly improve large sequential write speeds. On a
disk capable of 85MB/s, speeds increase from 50MB/s to 65MB/s for
noop and from 38MB/s to 50MB/s for cfq.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
There are two spare field in the header common to all GFS2
metadata. One is just the right size to fit a journal id
in it, and this patch updates the journal code so that each
time a metadata block is modified, we tag it with the journal
id of the node which is performing the modification.
The reason for this is that it should make it much easier to
debug issues which arise if we can tell which node was the
last to modify a particular metadata block.
Since the field is updated before the block is written into
the journal, each journal should only contain metadata which
is tagged with its own journal id. The one exception to this
is the journal header block, which might have a different node's
id in it, if that journal was recovered by another node in the
cluster.
Thus each journal will contain a record of which nodes recovered
it, via the journal header.
The other field in the metadata header could potentially be
used to hold information about what kind of operation was
performed, but for the time being we just zero it on each
transaction so that if we use it for that in future, we'll
know that the information (where it exists) is reliable.
I did consider using the other field to hold the journal
sequence number, however since in GFS2's journaling we write
the modified data into the journal and not the original
data, this gives no information as to what action caused the
modification, so I think we can probably come up with a better
use for those 64 bits in the future.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch adds the ability to trace various aspects of the GFS2
filesystem. The trace points are divided into three groups,
glocks, logging and bmap. These points have been chosen because
they allow inspection of the major internal functions of GFS2
and they are also generic enough that they are unlikely to need
any major changes as the filesystem evolves.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
After Jens recent updates:
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=a1f242524c3c1f5d40f1c9c343427e34d1aadd6e
et al. this is a patch to bring gfs2 uptodate with the core
code. Also I've managed to squash another call to ll_rw_block()
along the way.
There is still one part of the GFS2 I/O paths which are not correctly
annotated and that is due to the sharing of the writeback code between
the data and metadata address spaces. I would like to change that too,
but this patch is still worth doing on its own, I think.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This is the big patch that I've been working on for some time
now. There are many reasons for wanting to make this change
such as:
o Reducing overhead by eliminating duplicated fields between structures
o Simplifcation of the code (reduces the code size by a fair bit)
o The locking interface is now the DLM interface itself as proposed
some time ago.
o Fewer lookups of glocks when processing replies from the DLM
o Fewer memory allocations/deallocations for each glock
o Scope to do further optimisations in the future (but this patch is
more than big enough for now!)
Please note that (a) this patch relates to the lock_dlm module and
not the DLM itself, that is still a separate module; and (b) that
we retain the ability to build GFS2 as a standalone single node
filesystem with out requiring the DLM.
This patch needs a lot of testing, hence my keeping it I restarted
my -git tree after the last merge window. That way, this has the maximum
exposure before its merged. This is (modulo a few minor bug fixes) the
same patch that I've been posting on and off the the last three months
and its passed a number of different tests so far.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch is performance related. When we're doing a log flush,
I noticed we were calling buf_lo_incore_commit twice: once for
data bufs and once for metadata bufs. Since this is the same
function and does the same thing in both cases, there should be
no reason to call it twice. Since we only need to call it once,
we can also make it faster by removing it from the generic "lops"
code and making it a stand-along static function.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The only reason for adding glocks to the journal was to keep track
of which locks required a log flush prior to release. We add a
flag to the glock to allow this check to be made in a simpler way.
This reduces the size of a glock (by 12 bytes on i386, 24 on x86_64)
and means that we can avoid extra work during the journal flush.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch splits gfs2_writepage into separate functions for each of
the three cases: writeback, ordered and journalled. As a result
it becomes a lot easier to see what each one is doing. The common
code is moved into gfs2_writepage_common.
This fixes a performance bug where we were doing more work than
strictly required in the ordered write case.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch cleans up the code for writing journaled data into the log.
It also removes the need to allocate a small "tag" structure for each
block written into the log. Instead we just keep count of the outstanding
I/O so that we can be sure that its all been written at the correct time.
Another result of this patch is that a number of ll_rw_block() calls
have become submit_bh() calls, closing some races at the same time.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The following alters gfs2_trans_add_revoke() to take a struct
gfs2_bufdata as an argument. This eliminates the memory allocation which
was previously required by making use of the already existing struct
gfs2_bufdata. It makes some sanity checks to ensure that the
gfs2_bufdata has been removed from all the lists before its recycled as
a revoke structure. This saves one memory allocation and one free per
revoke structure.
Also as a result, and to simplify the locking, since there is no longer
any blocking code in gfs2_trans_add_revoke() we must hold the log lock
whenever this function is called. This reduces the amount of times we
take and unlock the log lock.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The old revoke structure was allocated using kalloc/kfree but
there is a slab cache for gfs2_bufdata, so we should use that
now that the structures have been converted.
This is part two of the patch series to merge the revoke
and gfs2_bufdata structures.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Both the revoke structure and the bufdata structure are quite similar.
They are basically small tags which are put on lists. In addition to
which the revoke structure is always allocated when there is a bufdata
structure which is (or can be) freed. As such it should be possible to
reduce the number of frees and allocations by using the same structure
for both purposes.
This patch is the first step along that path. It replaces existing uses
of the revoke structure with the bufdata structure.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The following patch removes the ordered write processing from
databuf_lo_before_commit() and moves it to log.c. This has the effect of
greatly simplyfying databuf_lo_before_commit() and well as potentially
making the ordered write code more efficient.
As a side effect of this, its now possible to remove ordered buffers
from the ordered buffer list at any time, so we now make use of this in
invalidatepage and releasepage to ensure timely release of these
buffers.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
gfs2_pin and gfs2_unpin are only used in lops.c, despite being
defined in meta_io.c, so this patch moves them into lops.c and
makes them static. At the same time, its possible to clean up
the locking in the buf and databuf _lo_add() functions so that
we only need to grab the spinlock once. Also we have to move
lock_buffer() around the _lo_add() functions since we can't
do that in gfs2_pin() any more since we hold the spinlock
for the duration of that function.
As a result, the code shrinks by 12 lines and we do far fewer
operations when adding buffers to the log. It also makes the
code somewhat easier to read & understand.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This is a patch to GFS2 to protect sd_log_num_jdata with the
gfs2_log_lock. Without this patch, there is a timing window
where you can get hit the following assert from function
gfs2_log_flush():
gfs2_assert_withdraw(sdp,
sdp->sd_log_num_buf + sdp->sd_log_num_jdata ==
sdp->sd_log_commited_buf +
sdp->sd_log_commited_databuf);
I've tested it on my roth cluster and it fixes the problem.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This is the first of five patches for bug #248176:
There were still some critical variables being manipulated outside
the log_lock spinlock. That usually resulted in a hang.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This is part 2 of the patch for bug #245832, part 1 of which is already
in the git tree.
The problem was that sdp->sd_log_num_databuf was not always being
protected by the gfs2_log_lock spinlock, but the sd_log_le_databuf
(which it is supposed to reflect) was protected. That meant there
was a timing window during which gfs2_log_flush called
databuf_lo_before_commit and the count didn't match what was
really on the linked list in that window. So when it ran out of
items on the linked list, it decremented total_dbuf from 0 to -1 and
thus never left the "while(total_dbuf)" loop.
The solution is to protect the variable sdp->sd_log_num_databuf so
that the value will always match the contents of the linked list,
and therefore the number will never go negative, and therefore, the
loop will be exited properly.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch is an addendum to the previous journaled file/unmount patch.
It fixes a problem discovered during testing.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
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>
This patch is for bugzilla bug 283162, which uncovered a number of
bugs pertaining to writing to files that have the journaled bit on.
These bugs happen most often when writing to the meta_fs because
the files are always journaled. So operations like gfs2_grow were
particularly vulnerable, although many of the problems could be
recreated with normal files after setting the journaled bit on.
The problems fixed are:
-GFS2 wasn't ever writing unstuffed journaled data blocks to their
in-place location on disk. Now it does.
-If you unmounted too quickly after doing IO to a journaled file,
GFS2 was crashing because you would discard a buffer whose bufdata
was still on the active items list. GFS2 now deals with this
gracefully.
-GFS2 was losing track of the bufdata for journaled data blocks,
and it wasn't getting freed, causing an error when you tried to
unmount the module. GFS2 now frees all the bufdata structures.
-There was a memory corruption occurring because GFS2 wrote
twice as many log entries for journaled buffers.
-It was occasionally trying to write journal headers in buffers
that weren't currently mapped.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This is a patch for the first three issues of RHBZ #238162
The first issue is that when you allocate a new page for a file, it will not
start off uptodate. This makes sense, since you haven't written anything to that
part of the file yet. Unfortunately, gfs2_pin() checks to make sure that the
buffers are uptodate. The solution to this is to mark the buffers uptodate in
gfs2_commit_write(), after they have been zeroed out and have the data written
into them. I'm pretty confident with this fix, although it's not completely
obvious that there is no problem with marking the buffers uptodate here.
The second issue is simply that you can try to pin a data buffer that is already
on the incore log, and thus, already pinned. This patch checks to see if this
buffer is already on the log, and exits databuf_lo_add() if it is, just like
buf_lo_add() does.
The third issue is that gfs2_log_flush() doesn't do it's block accounting
correctly. Both metadata and journaled data are logged, but gfs2_log_flush()
only compares the number of metadata blocks with the number of blocks to commit
to the ondisk journal. This patch also counts the journaled data blocks.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
When glock_lo_add and rg_lo_add attempt to add an element to the log, they
check to see if has already been added before locking the log. If another
process adds that element to the log in this window between the check and
locking the log, the element will be added to the list twice. This causes
the log element list to become corrupted in such a way that the log element
can never be successfully removed from the list. This patch pulls the
list_empty() check inside the log lock, to remove this window.
Signed-off-by: Benjamin E. Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The patch below appears to fix the list corruption that we are seeing on
occasion. Although the transaction structure is private to a single
thread, when the queued structures are dismantled during an in-core
commit, its possible for a different thread to be trying to add the same
structure to another, new, transaction at the same time.
To avoid this, this patch takes the log spinlock during this operation.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The log lock is dropped prior to io submittion, but
this exposes a hole in which the log data structures
may be going away due to a truncate.
Store the buffer head in a local pointer prior to
dropping the lock and relay on the buffer_head lock
for consitency on the buffer head.
Signed-Off-By: Russell Cattelan <cattelan@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Pass kaddr rather than (incorrect) struct page to kunmap_atomic.
Signed-off-by: Russell Cattelan <cattelan@redhat.com>
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>
lm_interface.h has a few out of the tree clients such as GFS1
and userland tools.
Right now, these clients keeps a copy of the file in their build tree
that can go out of sync.
Move lm_interface.h to include/linux, export it to userland and
clean up fs/gfs2 to use the new location.
Signed-off-by: Fabio M. Di Nitto <fabbione@ubuntu.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
As per Jan Engelhardt's comments, removed some unused code and
removed some brackets which were not required.
Cc: Jan Engelhardt <jengelh@linux01.gwdg.de>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
As per Jan Engelhardt's request, I've added a ',' to the end of
each of the multi-line structures which didn't already have
one (most already did).
Cc: Jan Engelhardt <jengelh@linux01.gwdg.de>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
As per Jan Engelhardt's fourth email, this is the first part of the
change set with a few minor style points.
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>
This patch fixes three main bugs. Firstly the direct i/o get_block
was returning the wrong return code in certain cases. Secondly, the
GFS2's releasepage function was not dealing with cases when clean,
ordered buffers were found still queued on a transaction (which can
happen depending on the ordering of journal flushes). Thirdly, the
journaling code itself needed altering to take account of the
after effects of removing the clean ordered buffers from the transactions
before a journal flush.
The releasepage bug did also show up under "normal" buffered i/o
as well, so its not just a fix for direct i/o. In fact its not
normally used in the direct i/o path at all, except when flushing
existing buffers after performing a direct i/o write, but that was
the code path that led us to spot this.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
A list_del should have been a list_del_init in lops.c which was
resulting in incorrect status returns from list_empty().
Signed-off-by: Steven Whitheouse <swhiteho@redhat.com>
This fixes a memory leak of struct gfs2_bufdata and also some
problems in the ordered write handling code. It needs a bit
more testing, but I believe that the reference counting of
ordered write buffers should now be correct.
This is aimed at fixing Red Hat bugzilla: #201028 and #201082
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Include the glock in the transaction, even when not journaling
data in order that ordered write data will be correctly flushed
when the lock is released.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
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 no longer use semaphores, everything has been converted to
mutex or rwsem, so we don't need to include this header any more.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This fixes a ref count bug that sometimes showed up a umount time
(causing it to hang) but it otherwise mostly harmless. At the same
time there are some clean ups including making the log operations
structures const, moving a memory allocation so that its not done
in the fast path of checking to see if there is an outstanding
transaction related to a particular glock.
Removes the sd_log_wrap varaible which was updated, but never actually
used anywhere. Updates the gfs2 ioctl() to run without the kernel lock
(which it never needed anyway). Removes the "invalidate inodes" loop
from GFS2's put_super routine. This is done in kill super anyway so
we don't need to do it here. The loop was also bogus in that if there
are any inodes "stuck" at this point its a bug and we need to know
about it rather than hide it by hanging forever.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Fix a bug I introduced earlier with a kfree() and usage of
a structure in the wrong order. Also try and get the counts
of the journaled data buffers "more correct". Still some work
to do in this area though.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
As suggested by Pekka Enberg <penberg@cs.helsinki.fi>.
The DIV_RU macro is renamed DIV_ROUND_UP and and moved to kernel.h
The other macros are gone from gfs2.h as (although not requested
by Pekka Enberg) are a number of included header file which are now
included individually. The inode number comparison function is
now an inline function.
The DT2IF and IF2DT may be addressed in a future patch.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
