deal with the fragment expansion separately before the rest of the operation.
this allows us to simplify ufs_balloc_range() by not worrying about implicit
fragment expansion.
call VOP_PUTPAGES() directly for vnodes instead of
going through the UVM pager "put" vector.
- calculate the offset and length of the postbl before byteswapping.
problem noted by der Mouse.
- use offsetof() to determine # of fields to calculate in initial
loop, rather than hard-coding in `52 fields'
- improve comments.
fragment synchronously, update the vnode's size before doing the flush.
otherwise we might only write part of the data and cause softdep's
accounting to get out of sync. fixes PR 14201.
many thanks to enami for figuring out what was going on.
spotted by Nathan Williams.
while I'm here, move an splbio() so that we don't return without
splx()ing it if there's an error, and don't bother calling the
pager put if the vnode has no pages.
- don't both updating fs->fs_cgrotor, since it's actually not used in
the kernel. from Manuel Bouyer in [kern/3389]
- when examining cylinder groups from startcg to startcg-1 (wrapping
at fs->fs_ncg), there's no need to check startcg at the end as well
as the start...
- highlight in the struct fs declaration that fs_cgrotor is UNUSED
- remove special treatment of pager_map mappings in pmaps. this is
required now, since I've removed the globals that expose the address range.
pager_map now uses pmap_kenter_pa() instead of pmap_enter(), so there's
no longer any need to special-case it.
- eliminate struct uvm_vnode by moving its fields into struct vnode.
- rewrite the pageout path. the pager is now responsible for handling the
high-level requests instead of only getting control after a bunch of work
has already been done on its behalf. this will allow us to UBCify LFS,
which needs tighter control over its pages than other filesystems do.
writing a page to disk no longer requires making it read-only, which
allows us to write wired pages without causing all kinds of havoc.
- use a new PG_PAGEOUT flag to indicate that a page should be freed
on behalf of the pagedaemon when it's unlocked. this flag is very similar
to PG_RELEASED, but unlike PG_RELEASED, PG_PAGEOUT can be cleared if the
pageout fails due to eg. an indirect-block buffer being locked.
this allows us to remove the "version" field from struct vm_page,
and together with shrinking "loan_count" from 32 bits to 16,
struct vm_page is now 4 bytes smaller.
- no longer use PG_RELEASED for swap-backed pages. if the page is busy
because it's being paged out, we can't release the swap slot to be
reallocated until that write is complete, but unlike with vnodes we
don't keep a count of in-progress writes so there's no good way to
know when the write is done. instead, when we need to free a busy
swap-backed page, just sleep until we can get it busy ourselves.
- implement a fast-path for extending writes which allows us to avoid
zeroing new pages. this substantially reduces cpu usage.
- encapsulate the data used by the genfs code in a struct genfs_node,
which must be the first element of the filesystem-specific vnode data
for filesystems which use genfs_{get,put}pages().
- eliminate many of the UVM pagerops, since they aren't needed anymore
now that the pager "put" operation is a higher-level operation.
- enhance the genfs code to allow NFS to use the genfs_{get,put}pages
instead of a modified copy.
- clean up struct vnode by removing all the fields that used to be used by
the vfs_cluster.c code (which we don't use anymore with UBC).
- remove kmem_object and mb_object since they were useless.
instead of allocating pages to these objects, we now just allocate
pages with no object. such pages are mapped in the kernel until they
are freed, so we can use the mapping to find the page to free it.
this allows us to remove splvm() protection in several places.
The sum of all these changes improves write throughput on my
decstation 5000/200 to within 1% of the rate of NetBSD 1.5
and reduces the elapsed time for "make release" of a NetBSD 1.5
source tree on my 128MB pc to 10% less than a 1.5 kernel took.
allocate memory from kernel_map but some of the objects are freed from
interrupt context, we put objects on a queue instead of freeing them
immediately. then in softdep_process_worklist() (which is called at
least once per second from the syncer), we process that queue and
free all the objects. allocating from kernel_map instead of from kmem_map
allows us to have a much larger number of softdeps pending even in
configurations where kmem_map is relatively small.
adjusted via sysctl. file systems that have hash tables which are
sized based on the value of this variable now resize those hash tables
using the new value. the max number of FFS softdeps is also recalculated.
convert various file systems to use the <sys/queue.h> macros for
their hash tables.
FreeBSD (three commits; the initial work, man page updates, and a fix
to ffs_reload()), with the following differences:
- Be consistent between newfs(8) and tunefs(8) as to the options which
set and control the tuning parameters for this work (avgfilesize & avgfpdir)
- Use u_int16_t instead of u_int8_t to keep track of the number of
contiguous directories (suggested by Chuck Silvers)
- Work within our FFS_EI framework
- Ensure that fs->fs_maxclusters and fs->fs_contigdirs don't point to
the same area of memory
The new algorithm has a marked performance increase, especially when
performing tasks such as untarring pkgsrc.tar.gz, etc.
The original FreeBSD commit messages are attached:
=====
mckusick 2001/04/10 01:39:00 PDT
Directory layout preference improvements from Grigoriy Orlov <gluk@ptci.ru>.
His description of the problem and solution follow. My own tests show
speedups on typical filesystem intensive workloads of 5% to 12% which
is very impressive considering the small amount of code change involved.
------
One day I noticed that some file operations run much faster on
small file systems then on big ones. I've looked at the ffs
algorithms, thought about them, and redesigned the dirpref algorithm.
First I want to describe the results of my tests. These results are old
and I have improved the algorithm after these tests were done. Nevertheless
they show how big the perfomance speedup may be. I have done two file/directory
intensive tests on a two OpenBSD systems with old and new dirpref algorithm.
The first test is "tar -xzf ports.tar.gz", the second is "rm -rf ports".
The ports.tar.gz file is the ports collection from the OpenBSD 2.8 release.
It contains 6596 directories and 13868 files. The test systems are:
1. Celeron-450, 128Mb, two IDE drives, the system at wd0, file system for
test is at wd1. Size of test file system is 8 Gb, number of cg=991,
size of cg is 8m, block size = 8k, fragment size = 1k OpenBSD-current
from Dec 2000 with BUFCACHEPERCENT=35
2. PIII-600, 128Mb, two IBM DTLA-307045 IDE drives at i815e, the system
at wd0, file system for test is at wd1. Size of test file system is 40 Gb,
number of cg=5324, size of cg is 8m, block size = 8k, fragment size = 1k
OpenBSD-current from Dec 2000 with BUFCACHEPERCENT=50
You can get more info about the test systems and methods at:
http://www.ptci.ru/gluk/dirpref/old/dirpref.html
Test Results
tar -xzf ports.tar.gz rm -rf ports
mode old dirpref new dirpref speedup old dirprefnew dirpref speedup
First system
normal 667 472 1.41 477 331 1.44
async 285 144 1.98 130 14 9.29
sync 768 616 1.25 477 334 1.43
softdep 413 252 1.64 241 38 6.34
Second system
normal 329 81 4.06 263.5 93.5 2.81
async 302 25.7 11.75 112 2.26 49.56
sync 281 57.0 4.93 263 90.5 2.9
softdep 341 40.6 8.4 284 4.76 59.66
"old dirpref" and "new dirpref" columns give a test time in seconds.
speedup - speed increasement in times, ie. old dirpref / new dirpref.
------
Algorithm description
The old dirpref algorithm is described in comments:
/*
* Find a cylinder to place a directory.
*
* The policy implemented by this algorithm is to select from
* among those cylinder groups with above the average number of
* free inodes, the one with the smallest number of directories.
*/
A new directory is allocated in a different cylinder groups than its
parent directory resulting in a directory tree that is spreaded across
all the cylinder groups. This spreading out results in a non-optimal
access to the directories and files. When we have a small filesystem
it is not a problem but when the filesystem is big then perfomance
degradation becomes very apparent.
What I mean by a big file system ?
1. A big filesystem is a filesystem which occupy 20-30 or more percent
of total drive space, i.e. first and last cylinder are physically
located relatively far from each other.
2. It has a relatively large number of cylinder groups, for example
more cylinder groups than 50% of the buffers in the buffer cache.
The first results in long access times, while the second results in
many buffers being used by metadata operations. Such operations use
cylinder group blocks and on-disk inode blocks. The cylinder group
block (fs->fs_cblkno) contains struct cg, inode and block bit maps.
It is 2k in size for the default filesystem parameters. If new and
parent directories are located in different cylinder groups then the
system performs more input/output operations and uses more buffers.
On filesystems with many cylinder groups, lots of cache buffers are
used for metadata operations.
My solution for this problem is very simple. I allocate many directories
in one cylinder group. I also do some things, so that the new allocation
method does not cause excessive fragmentation and all directory inodes
will not be located at a location far from its file's inodes and data.
The algorithm is:
/*
* Find a cylinder group to place a directory.
*
* The policy implemented by this algorithm is to allocate a
* directory inode in the same cylinder group as its parent
* directory, but also to reserve space for its files inodes
* and data. Restrict the number of directories which may be
* allocated one after another in the same cylinder group
* without intervening allocation of files.
*
* If we allocate a first level directory then force allocation
* in another cylinder group.
*/
My early versions of dirpref give me a good results for a wide range of
file operations and different filesystem capacities except one case:
those applications that create their entire directory structure first
and only later fill this structure with files.
My solution for such and similar cases is to limit a number of
directories which may be created one after another in the same cylinder
group without intervening file creations. For this purpose, I allocate
an array of counters at mount time. This array is linked to the superblock
fs->fs_contigdirs[cg]. Each time a directory is created the counter
increases and each time a file is created the counter decreases. A 60Gb
filesystem with 8mb/cg requires 10kb of memory for the counters array.
The maxcontigdirs is a maximum number of directories which may be created
without an intervening file creation. I found in my tests that the best
performance occurs when I restrict the number of directories in one cylinder
group such that all its files may be located in the same cylinder group.
There may be some deterioration in performance if all the file inodes
are in the same cylinder group as its containing directory, but their
data partially resides in a different cylinder group. The maxcontigdirs
value is calculated to try to prevent this condition. Since there is
no way to know how many files and directories will be allocated later
I added two optimization parameters in superblock/tunefs. They are:
int32_t fs_avgfilesize; /* expected average file size */
int32_t fs_avgfpdir; /* expected # of files per directory */
These parameters have reasonable defaults but may be tweeked for special
uses of a filesystem. They are only necessary in rare cases like better
tuning a filesystem being used to store a squid cache.
I have been using this algorithm for about 3 months. I have done
a lot of testing on filesystems with different capacities, average
filesize, average number of files per directory, and so on. I think
this algorithm has no negative impact on filesystem perfomance. It
works better than the default one in all cases. The new dirpref
will greatly improve untarring/removing/coping of big directories,
decrease load on cvs servers and much more. The new dirpref doesn't
speedup a compilation process, but also doesn't slow it down.
Obtained from: Grigoriy Orlov <gluk@ptci.ru>
=====
=====
iedowse 2001/04/23 17:37:17 PDT
Pre-dirpref versions of fsck may zero out the new superblock fields
fs_contigdirs, fs_avgfilesize and fs_avgfpdir. This could cause
panics if these fields were zeroed while a filesystem was mounted
read-only, and then remounted read-write.
Add code to ffs_reload() which copies the fs_contigdirs pointer
from the previous superblock, and reinitialises fs_avgf* if necessary.
Reviewed by: mckusick
=====
=====
nik 2001/04/10 03:36:44 PDT
Add information about the new options to newfs and tunefs which set the
expected average file size and number of files per directory. Could do
with some fleshing out.
=====
in an effort to maintain compatibility with freebsd/openbsd/whatever,
i'm attempting to get the superblock format in sync, and freebsd uses
the int32_t at this position for `fs_pendinginodes'.
if we ever decide to implement fscktime functionality, we'll:
a) make sure to liaise with the other projects to reserve the same
spare field
b) actually implement the code this time ...
(this is also preparing us for other changes, like the new dirpref code)
cylinder groups to work correctly, with minor modifications by me to work
with our FFS_EI code. From the FreeBSD commit message:
The ffs superblock includes a 128-byte region for use by temporary
in-core pointers to summary information. An array in this region
(fs_csp) could overflow on filesystems with a very large number of
cylinder groups (~16000 on i386 with 8k blocks). When this happens,
other fields in the superblock get corrupted, and fsck refuses to
check the filesystem.
Solve this problem by replacing the fs_csp array in 'struct fs'
with a single pointer, and add padding to keep the length of the
128-byte region fixed. Update the kernel and userland utilities
to use just this single pointer.
With this change, the kernel no longer makes use of the superblock
fields 'fs_csshift' and 'fs_csmask'. Add a comment to newfs/mkfs.c
to indicate that these fields must be calculated for compatibility
with older kernels.
Reviewed by: mckusick
- Cast blk argument to lblktosize() to (off_t), to prevent 32 bit overflow.
whilst almost every use in ffs used this for small blknos, there are
potential issues, and it's safer this way. (as discussed with chuq)
- Use 64bit (off_t) math to calculate if we have hit our freespace() limit.
Necessary for coherent results on filesystems bigger than 0.5Tb.
- Use lblktosize() in blksize() and dblksize(), to make it obvious what's
happening
- Remove sblksize() - nothing uses it
- replace the unused fs_headswitch and fs_trkseek with fs_id[2], bringing
our struct fs closer to that in freebsd & openbsd (& solaris FWIW)
- dumpfs: improve warning message when cpc == 0
determine the endianness of the `struct fs *o' superblock from o->fs_magic
and set needswap as necessary, rather than trusting the caller to get
it right. invariably, almost every caller of ffs_sb_swap() was calling it
with ns set to the wrong value for ns anyway!
ansi KNF ffs_bswap.c declarations whilst here.
this fixes all sorts of problems when trying to use other-endian file systems,
notably the kernel trying to access memory *way* off, possibly corrupting or
panicing, and userland programs SEGVing and/or corrupting things (e.g,
"fsck_ffs -B" to swap a file system endianness).
whilst the previous rev of ffs_bswap.c (1.10, 2000/12/23) made this problem
worse, i suspect that the problem was always there and previous versions
just happened not to trash things at the wrong time.
FFS_EI should now be a lot more stable.
Kernels and tools understand both v1 and v2 filesystems; newfs_lfs
generates v2 by default. Changes for the v2 layout include:
- Segments of non-PO2 size and arbitrary block offset, so these can be
matched to convenient physical characteristics of the partition (e.g.,
stripe or track size and offset).
- Address by fragment instead of by disk sector, paving the way for
non-512-byte-sector devices. In theory fragments can be as large
as you like, though in reality they must be smaller than MAXBSIZE in size.
- Use serial number and filesystem identifier to ensure that roll-forward
doesn't get old data and think it's new. Roll-forward is enabled for
v2 filesystems, though not for v1 filesystems by default.
- The inode free list is now a tailq, paving the way for undelete (undelete
is not yet implemented, but can be without further non-backwards-compatible
changes to disk structures).
- Inode atime information is kept in the Ifile, instead of on the inode;
that is, the inode is never written *just* because atime was changed.
Because of this the inodes remain near the file data on the disk, rather
than wandering all over as the disk is read repeatedly. This speeds up
repeated reads by a small but noticeable amount.
Other changes of note include:
- The ifile written by newfs_lfs can now be of arbitrary length, it is no
longer restricted to a single indirect block.
- Fixed an old bug where ctime was changed every time a vnode was created.
I need to look more closely to make sure that the times are only updated
during write(2) and friends, not after-the-fact during a segment write,
and certainly not by the cleaner.
vfs_busy'ing just before the dounmount() call. This is to avoid
sleeping with the mountlist_slock held -- but we must acquire
syncer_lock before vfs_busy because the syncer itself uses
syncer_lock -> vfs_busy locking order.
space before deciding which cylinder group should contain a new directory
inode.
Fixes kern/11983; works around some, but not all, of the side effects
of kern/11989.
Tested by me for well over a month on my laptop; preliminary versions of
the fix were tested by Frank van der Linden and Herb Peyerl.
in effect cosmetic). Original FreeBSD commit messages:
==
date: 2000/03/15 07:18:15; author: mckusick; state: Exp; lines: +4 -4
Bug fixes for currently harmless bugs that could rise to bite
the unwary if the code were called in slightly different ways.
[...]
2) In ufs_lookup() there is an off-by-one error in the test that checks
if dp->i_diroff is outside the range of the the current directory size.
This is completely harmless, since the following while-loop condition
'dp->i_offset < endsearch' is never met, so the code immediately
does a second pass starting at dp->i_offset = 0.
3) Again in ufs_lookup(), the condition in a sanity check is wrong
for directories that are longer than one block. This bug means that
the sanity check is only effective for small directories.
Submitted by: Ian Dowse <iedowse@maths.tcd.ie>
==
date: 2000/03/09 18:54:59; author: dillon; state: Exp; lines: +2 -2
branches: 1.33.2;
In the 'found' case for ufs_lookup() the underlying bp's data was
being accessed after the bp had been releaed. A simple move of the
brelse() solves the problem.
Approved by: jkh
Submitted by: Ian Dowse <iedowse@maths.tcd.ie>
==
don't update UVM's notion of the file size before the VOP_FSYNC() when
we're partially truncating a file with softdeps enabled. doing so could
free pages without updating the dependency info, which would result in
"panic: softdep_write_inodeblock: direct pointer #1 mismatch 0 != N".
lfs_writeseg (possibly after they had been freed).
If MALLOCLOG is defined, make lfs_newbuf and lfs_freebuf pass along the
caller's file and line to _malloc and _free.
pages we've allocated past the real EOF when we fail to allocate a block.
we used to play games with the VM notion of the file size but we don't do
that anymore, so uvm_vnp_setsize() doesn't do what we want anymore.
call the pager flush op instead.
aby bad symptoms any more, fix for bug causing problems with this
option was in BSD4.4-Lite2 and pulled in together with softdep changes
See also Keith Smith & Margo Seltzer's paper on the topic at
http://www.eecs.harvard.edu/~keith/papers/realloc.ps.gz
on mount, through the newer checkpoint and on through any newer
partial-segments that may have been written but not checkpointed because
of an intervening crash.
LFS_DO_ROLLFORWARD is not defined by default.
in an error case in lfs_markv. Change the vfs_getvfs() error to return
ENOENT, for consistency with failure of vfs_busy().
99% of this patch was from Jesse Off <joff@gci-net.com> (PR #11547).
(PR #11468). In the case of fragment allocation, check to see if enough
space is available before extending a fragment already scheduled for writing.
The locked_queue_* variables indicate the number of buffer headers and bytes,
respectively, that are unavailable to getnewbuf() because they are locked up
waiting for LFS to flush them; make sure that that is actually what we're
counting, i.e., never count malloced buffers, and always use b_bufsize instead
of b_bcount.
If DEBUG is defined, the periodic calls to lfs_countlocked will now complain
if either counter is incorrect. (In the future lfs_countlocked will not need
to be called at all if DEBUG is not defined.)
when deallocating a fragment that has not made it to disk yet.
Also, during dirops, give the directory vnode an extra reference in
SET_DIROP, to ensure its continued existence during SET_ENDOP, preventing
a possible NULL-dereference there.
These two changes should close PR #11064.
Kernel:
* Add runtime quantity lfs_ravail, the number of disk-blocks reserved
for writing. Writes to the filesystem first reserve a maximum amount
of blocks before their write is allowed to proceed; after the blocks
are allocated the reserved total is reduced by a corresponding amount.
If the lfs_reserve function cannot immediately reserve the requested
number of blocks, the inode is unlocked, and the thread sleeps until
the cleaner has made enough space available for the blocks to be
reserved. In this way large files can be written to the filesystem
(or, smaller files can be written to a nearly-full but thoroughly
clean filesystem) and the cleaner can still function properly.
* Remove explicit switching on dlfs_minfreeseg from the kernel code; it
is now merely a fs-creation parameter used to compute dlfs_avail and
dlfs_bfree (and used by fsck_lfs(8) to check their accuracy). Its
former role is better assumed by a properly computed dlfs_avail.
* Bounds-check inode numbers submitted through lfs_bmapv and lfs_markv.
This prevents a panic, but, if the cleaner is feeding the filesystem
the wrong data, you are still in a world of hurt.
* Cleanup: remove explicit references of DEV_BSIZE in favor of
btodb()/dbtob().
lfs_cleanerd:
* Make -n mean "send N segments' blocks through a single call to
lfs_markv". Previously it had meant "clean N segments though N calls
to lfs_markv, before looking again to see if more need to be cleaned".
The new behavior gives better packing of direct data on disk with as
little metadata as possible, largely alleviating the problem that the
cleaner can consume more disk through inefficient use of metadata than
it frees by moving dirty data away from clean "holes" to produce
entirely clean segments.
* Make -b mean "read as many segments as necessary to write N segments
of dirty data back to disk", rather than its former meaning of "read
as many segments as necessary to free N segments worth of space". The
new meaning, combined with the new -n behavior described above,
further aids in cleaning storage efficiency as entire segments can be
written at once, using as few blocks as possible for segment summaries
and inode blocks.
* Make the cleaner take note of segments which could not be cleaned due
to error, and not attempt to clean them until they are entirely free
of dirty blocks. This prevents the case in which a cleanerd running
with -n 1 and without -b (formerly the default) would spin trying
repeatedly to clean a corrupt segment, while the remaining space
filled and deadlocked the filesystem.
* Update the lfs_cleanerd manual page to describe all the options,
including the changes mentioned here (in particular, the -b and -n
flags were previously undocumented).
fsck_lfs:
* Check, and optionally fix, lfs_avail (to an exact figure) and
lfs_bfree (within a margin of error) in pass 5.
newfs_lfs:
* Reduce the default dlfs_minfreeseg to 1/20 of the total segments.
* Add a warning if the sgs disklabel field is 16 (the default for FFS'
cpg, but not usually desirable for LFS' sgs: 5--8 is a better range).
* Change the calculation of lfs_avail and lfs_bfree, corresponding to
the kernel changes mentioned above.
mount_lfs:
* Add -N and -b options to pass corresponding -n and -b options to
lfs_cleanerd.
* Default to calling lfs_cleanerd with "-b -n 4".
[All of these changes were largely tested in the 1.5 branch, with the
idea that they (along with previous un-pulled-up work) could be applied
to the branch while it was still in ALPHA2; however my test system has
experienced corruption on another filesystem (/dev/console has gone
missing :^), and, while I believe this unrelated to the LFS changes, I
cannot with good conscience request that the changes be pulled up.]
