filesystem-specific vnode data to the struct vnode. Provide the
default implementation for the vop_advlock and vop_advlockasync.
Purge the locks on the vnode reclaim by using the lf_purgelocks().
The default implementation is augmented for the nfs and smbfs.
In the nfs_advlock, push the Giant inside the nfs_dolock.
Before the change, the vop_advlock and vop_advlockasync have taken the
unlocked vnode and dereferenced the fs-private inode data, racing with
with the vnode reclamation due to forced unmount. Now, the vop_getattr
under the shared vnode lock is used to obtain the inode size, and
later, in the lf_advlockasync, after locking the vnode interlock, the
VI_DOOMED flag is checked to prevent an operation on the doomed vnode.
The implementation of the lf_purgelocks() is submitted by dfr.
Reported by: kris
Tested by: kris, pho
Discussed with: jeff, dfr
MFC after: 2 weeks
user-mode lock manager, build a kernel with the NFSLOCKD option and
add '-k' to 'rpc_lockd_flags' in rc.conf.
Highlights include:
* Thread-safe kernel RPC client - many threads can use the same RPC
client handle safely with replies being de-multiplexed at the socket
upcall (typically driven directly by the NIC interrupt) and handed
off to whichever thread matches the reply. For UDP sockets, many RPC
clients can share the same socket. This allows the use of a single
privileged UDP port number to talk to an arbitrary number of remote
hosts.
* Single-threaded kernel RPC server. Adding support for multi-threaded
server would be relatively straightforward and would follow
approximately the Solaris KPI. A single thread should be sufficient
for the NLM since it should rarely block in normal operation.
* Kernel mode NLM server supporting cancel requests and granted
callbacks. I've tested the NLM server reasonably extensively - it
passes both my own tests and the NFS Connectathon locking tests
running on Solaris, Mac OS X and Ubuntu Linux.
* Userland NLM client supported. While the NLM server doesn't have
support for the local NFS client's locking needs, it does have to
field async replies and granted callbacks from remote NLMs that the
local client has contacted. We relay these replies to the userland
rpc.lockd over a local domain RPC socket.
* Robust deadlock detection for the local lock manager. In particular
it will detect deadlocks caused by a lock request that covers more
than one blocking request. As required by the NLM protocol, all
deadlock detection happens synchronously - a user is guaranteed that
if a lock request isn't rejected immediately, the lock will
eventually be granted. The old system allowed for a 'deferred
deadlock' condition where a blocked lock request could wake up and
find that some other deadlock-causing lock owner had beaten them to
the lock.
* Since both local and remote locks are managed by the same kernel
locking code, local and remote processes can safely use file locks
for mutual exclusion. Local processes have no fairness advantage
compared to remote processes when contending to lock a region that
has just been unlocked - the local lock manager enforces a strict
first-come first-served model for both local and remote lockers.
Sponsored by: Isilon Systems
PR: 95247 107555 115524 116679
MFC after: 2 weeks
BO_LOCK/UNLOCK/MTX when manipulating the bufobj.
- Create a new lock in the bufobj to lock bufobj fields independently.
This leaves the vnode interlock as an 'identity' lock while the bufobj
is an io lock. The bufobj lock is ordered before the vnode interlock
and also before the mnt ilock.
- Exploit this new lock order to simplify softdep_check_suspend().
- A few sync related functions are marked with a new XXX to note that
we may not properly interlock against a non-zero bv_cnt when
attempting to sync all vnodes on a mountlist. I do not believe this
race is important. If I'm wrong this will make these locations easier
to find.
Reviewed by: kib (earlier diff)
Tested by: kris, pho (earlier diff)
- Remove the "thread" argument from the lockmgr() function as it is
always curthread now
- Axe lockcount() function as it is no longer used
- Axe LOCKMGR_ASSERT() as it is bogus really and no currently used.
Hopefully this will be soonly replaced by something suitable for it.
- Remove the prototype for dumplockinfo() as the function is no longer
present
Addictionally:
- Introduce a KASSERT() in lockstatus() in order to let it accept only
curthread or NULL as they should only be passed
- Do a little bit of style(9) cleanup on lockmgr.h
KPI results heavilly broken by this change, so manpages and
FreeBSD_version will be modified accordingly by further commits.
Tested by: matteo
conjuction with 'thread' argument passing which is always curthread.
Remove the unuseful extra-argument and pass explicitly curthread to lower
layer functions, when necessary.
KPI results broken by this change, which should affect several ports, so
version bumping and manpage update will be further committed.
Tested by: kris, pho, Diego Sardina <siarodx at gmail dot com>
Remove this argument and pass curthread directly to underlying
VOP_LOCK1() VFS method. This modify makes the code cleaner and in
particular remove an annoying dependence helping next lockmgr() cleanup.
KPI results, obviously, changed.
Manpage and FreeBSD_version will be updated through further commits.
As a side note, would be valuable to say that next commits will address
a similar cleanup about VFS methods, in particular vop_lock1 and
vop_unlock.
Tested by: Diego Sardina <siarodx at gmail dot com>,
Andrea Di Pasquale <whyx dot it at gmail dot com>
- markvoldirty() needs to write to underlying GEOM provider. We
have to do that *before* g_access() which sets the GEOM provider
to read-only.
- Remove dirty flag before free'ing iconv related resources. The
dirty flag removal could fail, and it is hard to revert the
iconv-free after the fail.
- Mark volume as dirty if we have failed to mark it clean for safe.
- Other style fixes to the touched functions.
This is much simpler than for ffs since there are many fewer places
where we need to choose between a delayed write and a sync write --
just 5 in msdosfs and more than 30 in ffs.
This is more complete and correct than in ffs. Several places in ffs
are are still missing the choice. ffs_update() has a layering violation
that breaks callers which want to force a sync update (mainly fsync(2)
and O_SYNC write(2)).
However, fsync(2) and O_SYNC write(2) are still more broken than in
ffs, since they are broken for default (non-sync non-async) mounts
too. Both fail to sync the FAT in all cases, and both fail to sync
the directory entry in some cases after losing a race. Async everything
is probably safer than the half-baked sync of metadata given by default
mounts.
leaving space for adding missing options. Negative options are sorted
after removing their "no" prefix, and generic options are sorted before
msdosfs-specific ones.
(except indirectly for the size pseudo-attribute). If anything deserves
a sync update, then it is ids and immutable flags, since these are
related to security, but ffs never synced these and msdosfs doesn't
support them. (ufs_setattr() only does an update in one case where
it is least needed (for timestamps); it did pessimal sync updates for
timestamps until 1998/03/08 but was changed for unlogged reasons related
to soft updates.)
Now msdosfs calls deupdat() with waitfor == 0, which normally gives a
delayed update to disk but always gives a sync update of timestamps
in core, while for ffs everything is delayed until the syncer daemon
or other activity causes an update (except for timestamps).
This gives a large optimization mainly for things like cp -p, where
attribute adjustment could easily triple the number of physical I/O's
if it is done synchronously (but cp -p to msdosfs is not as bad as
that, since msdosfs doesn't support many attributes so null adjustments
are more common, and msdosfs doesn't support ctimes so even if cp
doesn't weed out null adjustments they don't become non-null after
clobbering the ctime).
All active fields in fsi are advisory/optional, so we shouldn't do
extra work to make them valid at all times, but instead we write to
the fsi too often (we still do), and we searched for a free cluster
for fsinxtfree too often.
This commit just removes the whole search and its results, so that we
write out our in-core copy of fsinxtfree instead of writing a "fixed"
copy and clobbering our in-core copy. This saves fixing 3 bugs:
- off-by-1 error for the end of the search, resulting in fsinxtfree
not actually being adjusted iff only the last cluster is free.
- missing adjustment when no clusters are free.
- off-by-many error for the start of the search. Starting the search
at 0 instead of at (the in-core copy of) fsinxtfree did more than
defeat the reasons for existence of fsinxtfree. fsinxtfree exists
mainly to avoid having to start at 0 for just the first search per
mount, but has the side effect of reducing bias towards allocating
near cluster 0. The bias would normally only be generated by the
first search per mount (if fsinxtfree is not supported), but since
we also adjusted the in-core copy of fsinxtfree here, we were doing
extra work to maximize the bias.
Approved by: re (kensmith)
can easily block in bread(), and then there was nothing to prevent the
static buffer (nambuf_{ptr,len,last_id}) being clobbered by another
thread.
The effects of the bug seem to have been limited to failed lookups and
mangled names in readdir(), since Giant locking provides enough
serialization to prevent concurrent calls to the functions that access
the buffer. They were very obvious for multiple concurrent tree walks,
especially with a small cluster size.
The bug was introduced in msdosfs_conv.c 1.34 and associated changes,
and is in all releases starting with 5.2.
The fix is to allocate the buffer as a local variable and pass around
pointers to it like "_r" functions in libc do. Stack use from this
is large but not too large. This also fixes a memory leak on module
unload.
Reviewed by: kib
Approved by: re (kensmith)
% mount | grep home
/dev/ad4s1e on /home (ufs, local, noatime, soft-updates)
% mount -u -o atime /home
% mount | grep home
/dev/ad4s1e on /home (ufs, local, soft-updates)
Restore this behavior for on 7.x for the following mount options:
noatime, noclusterr, noclusterw, noexec, nosuid, nosymfollow
In addition, on 7.x, the following are equivalent:
mount -u -o atime /home
mount -u -o nonoatime /home
Ideally, when we introduce new mount options, we should avoid
options starting with "no". :)
Requested by: jhb
Reported by: Karol Kwiat <karol.kwiat gmail com>, Scott Hetzel <swhetzel gmail com>
Approved by: re (bmah)
Proxy commit for: rodrigc
(uio_offset < 0) since this can't happen. If this happens, then the
general code handles the problem safely (better than before for reading,
returning 0 (EOF) instead of the bogus errno EINVAL, and the same as
before for writing, returning EFBIG).
In msdosfs_read(), don't check for (uio_resid < 0). msdosfs_write()
already didn't check.
In msdosfs_read(), document in a comment our assumptions that the caller
passed a valid uio_offset and uio_resid. ffs checks using KASSERT(),
and that is enough sanity checking. In the same comment, partly document
there is no need to check for the EOVERFLOW case, unlike in ffs where this
case can happen at least in theory.
In msdosfs_write(), add a comment about why the checking of
(uio_resid == 0) is explicit, unlike in ffs.
In msdosfs_write(), check for impossibly large final offsets before
checking if the file size rlimit would be exceeded, so that we don't
have an overflow bug in the rlimit check and are consistent with ffs.
We now return EFBIG instead of EFBIG plus a SIGXFSZ signal if the final
offset would be impossibly large but not so large as to cause overflow.
Overflow normally gave the benign behaviour of no signal.
Approved by: re (kensmith) (blanket)
remove some parentheses; fix some whitespace errors; fix only one case of
a boolean comparison of a non-boolean).
Improve an error message by quoting ".", and by not printing large positive
values as negative ones.
Approved by: re (kensmith) (blanket)
namespace pollution in <sys/vnode.h>.
Sort the include of <sys/mutex.h> instead of unsorting it after
<sys/vnode.h> and depending on the pollution there.
Approved by: re (kensmith) (blanket)
sector, instead of failing the whole mount if it is garbage. Fields
in the fsinfo sector are only advisory, so there are better sanity
checks than this, and we already silently fix up the only other advisory
field in the fsinfo (the free cluster count).
This wasn't handled quite right in rev.1.92, 1.117, or in NetBSD. 1.92
also failed the whole mount for the non-garbage magic value 0xffffffff
1.117 fixed this well enough in practice since garbage values shouldn't
occur in practice, but left the error handling larger and more convoluted
than necessary. Now we handle the magic value as a special case of
fixing up all out of bounds values.
Also fix up the estimated next free cluster number when there is no
fsinfo sector. We were using 0, but CLUST_FIRST is safer.
Approved by: re (kensmith)
message explained why the size is 1 sector, but the code used a
size of 1 cluster.
I/o sizes larger than necessary may cause serious coherency problems
in the buffer cache. Here I think there were only minor efficiency
problems, since a too-large fsinfo buffer could only get far enough
to overlap buffers for the same vnode (the device vnode), so mappings
are coherent at the page level although not at the buffer level, and
the former is probably enough due to our limited use of the fsinfo
buffer.
Approved by: re (kensmith)
o Initialize ownerships and permissions. They were garbage (0) for
root mounts since vfs_mountroot_try() doesn't ask for them to be set
and msdosfs's old incomplete code to set them was removed. The
garbage happened to give the correct ownerships root:wheel, but it
gave permissions 000 so init could not be execed. Use the macros
for root: wheel and 0755. (The removed code gave 0:0 and 0777. 0755
is more normal and secure, thought wrong for /tmp.)
o Check the readonly flag for initial (non-MNT_UPDATE) mounts in the
correct place, as in ffs. For root mounts, it is only passed in
mp->mnt_flags, since vfs_mountroot_try() only passes it as a flag
and nothing translates the flag to the "ro" option string. msdosfs
only looked for it in the string, so it gave a rw mount for root
mounts without even clearing the flag in mp->mnt_flags, so the final
state was inconsistent. Checking the flag only in mp->mnt_flags
works for initial userland mounts too. The MNT_UPDATE case is
messier.
The main point that should work but doesn't is fsck of msdosfs root
while it is mounted ro. This needs mainly MNT_RELOAD support to work.
It should be possible to run fsck -p and succeed provided the fs is
consistent, not just for msdosfs, but this fails because fsck -p always
tries to open the device rw. The hack that allows open for writing
in ffs is not implemented in msdosfs, since without MNT_RELOAD support
writing could only be harmful. So fsck must be turned off to use
msdosfs as root. This is quite dangerous, since msdosfs is still missing
actually using its fs-dirty flag internally, so it is happy to mount
dirty fileystems rw.
Unrelated changes:
- Fix missing error handling for MNT_UPDATE from rw to ro.
- Catch up with renaming msdos to msdosfs in a string.
Approved by: re (kensmith)
This gives a very large speedup for small block sizes (in my tests,
about 5 times for write and 3 times for read with a block size of 512,
if clustering is possible) and a moderate speedup for the moderatatly
large block sizes that should be used on non-small media (4K is the
best size in most cases, and the speedup for that is about 1.3 times
for write and 1.2 times for read). mmap() should benefit from clustering
like read()/write(), but the current implementation of vm only supports
clustering (at least for getpages) if the fs block size is >= PAGE SIZE.
msdosfs is now only slightly slower than ffs with soft updates for
writing and slightly faster for reading when both use their best block
sizes. Writing is slower for msdosfs because of more sync writes.
Reading is faster for msdosfs because indirect blocks interfere with
clustering in ffs.
The changes in msdosfs_read() and msdosfs_write() are simpler merges
of corresponding code in ffs (after fixing some style bugs in ffs).
msdosfs_bmap() needs fs-specific code. This implementation loops
calling a lower level bmap function to do the hard parts. This is a
bit inefficient, but is efficient enough since msdsfs_bmap() is only
called when there is physical i/o to do.
Approved by: re (hrs)
In msdosfs_read(), mainly reorder the main loop to the same order as in
ffs_read().
In msdosfs_write() and extendfile(), use vfs_bio_clrbuf() instead of
clrbuf(). I think this just just a bogus optimization, but ffs always
does it and msdosfs already did it in one place, and it is what I've
tested.
In msdosfs_write(), merge good bits from a comment in ffs_write(), and
fix 1 style bug.
In the main comment for msdosfs_pcbmap(), improve wording and catch
up with 13 years of changes in the function. This comment belongs in
VOP_BMAP.9 but that doesn't exist.
In msdosfs_bmap(), return EFBIG if the requested cluster number is out
of bounds instead of blindly truncating it, and fix many style bugs.
Approved by: re (hrs)
to the FAT is possible.
Make the FAT block size less arbitrary before it is rounded up:
- for FAT12, default to 3*512 instead of to 3 sectors. The magic 3 is
the default number of 512-byte FAT sectors on a floppy drive. That
many sectors is too many if the sector size is larger.
- for !FAT12, default to PAGE_SIZE instead of to 4096. Remove
MSDOSFS_DFLTBSIZE since it only obfuscated this 4096.
For reading the BPB, use a block size of 8192 instead of 2048 so that
sector sizes up to 8192 can work. We should try several sizes, or just
try the maximum supported size (MAXBSIZE = 64K). I use 8192 because
that is enough for DVD-RW's (even 2048 is enough) and 8192 has been
tested a lot in use by ffs.
This completes fixing msdosfs for some large sector sizes (up to 8K
for read and 64K for write). Microsoft documents support for sector
sizes up to 4K in mdosfs. ffs is currently limited to 8K for both
read and write.
Approved by: re (kensmith)
Approved by: nyan (several years ago)
part of fixing msdosfs for large sector sizes. One of the fixed bugs
was fatal for large sector sizes.
1. The fsinfo block has size 512, but it was misunderstood and declared
as having size 1024, with nothing in the second 512 bytes except a
signature at the end. The second 512 bytes actually normally (if
the file system was created by Windows) consist of a second boot
sector which is normally (in WinXP) empty except for a signature --
the normal layout is one boot sector, one fsinfo sector, another
boot sector, then these 3 sectors duplicated. However, other
layouts are valid. newfs_msdos produces a valid layout with one
boot sector, one fsinfo sector, then these 2 sectors duplicated.
The signature check for the extra part of the fsinfo was thus
normally checking the signature in either the second boot sector
or the first boot sector in the copy, and thus accidentally
succeeding. The extra signature check would just fail for weirder
layouts with 512-byte sectors, and for normal layouts with any other
sector size.
Remove the extra bytes and the extra signature check.
2. Old versions did i/o to the fsinfo block using size 1024, with the
second half only used for the extra signature check on read. This
was harmless for sector size 512, and worked accidentally for sector
size 1024. The i/o just failed for larger sector sizes.
The version being fixed did i/o to the fsinfo block using size
fsi_size(pmp) = (1024 << ((pmp)->pm_BlkPerSec >> 2)). This
expression makes no sense. It happens to work for sector small
sector sizes, but for sector size 32K it gives the preposterous
value of 64M and thus causes panics. A sector size of 32768 is
necessary for at least some DVD-RW's (where the minimum write size
is 32768 although the minimum read size is 2048).
Now that the size of the fsinfo block is 512, it always fits in
one sector so there is no need for a macro to express it. Just
use the sector size where the old code uses 1024.
Approved by: re (kensmith)
Approved by: nyan (several years ago for a different version of (2))
of the the first cluster in a file (and, if the allocation cannot be
continued contiguously, for subsequent clusters in a file) was randomized
in an attempt to leave space for contiguous allocation of subsequent
clusters in each file when there are multiple writers. This reduced
internal fragmentation by a few percent, but it increased external
fragmentation by up to a few thousand percent.
Use simple sequential allocation instead. Actually maintain the fsinfo
sequence index for this. The read and write of this index from/to
disk still have many non-critical bugs, but we now write an index that
has something to do with our allocations instead of being modified
garbage. If there is no fsinfo on the disk, then we maintain the index
internally and don't go near the bugs for writing it.
Allocating the first free cluster gives a layout that is almost as good
(better in some cases), but takes too much CPU if the FAT is large and
the first free cluster is not near the beginning.
The effect of this change for untar and tar of a slightly reduced copy
of /usr/src on a new file system was:
Before (msdosfs 4K-clusters):
untar: 459.57 real untar from cached file (actually a pipe)
tar: 342.50 real tar from uncached tree to /dev/zero
Before (ffs2 soft updates 4K-blocks 4K-frags)
untar: 39.18 real
tar: 29.94 real
Before (ffs2 soft updates 16K-blocks 2K-frags)
untar: 31.35 real
tar: 18.30 real
After (msdosfs 4K-clusters):
untar 54.83 real
tar 16.18 real
All of these times can be improved further.
With multiple concurrent writers or readers (especially readers), the
improvement is smaller, but I couldn't find any case where it is
negative. 342 seconds for tarring up about 342 MB on a ~47MB/S partition
is just hard to unimprove on. (This operation would take about 7.3
seconds with reasonably localized allocation and perfect read-ahead.)
However, for active file systems, 342 seconds is closer to normal than
the 16+ seconds above or the 11 seconds with other changes (best I've
measured -- won easily by msdosfs!). E.g., my active /usr/src on ffs1
is quite old and fragmented, so reading to prepare for the above
benchmark takes about 6 times longer than reading back the fresh copies
of it.
Approved by: re (kensmith)
some cases, move to priv_check() if it was an operation on a thread and
no other flags were present.
Eliminate caller-side jail exception checking (also now-unused); jail
privilege exception code now goes solely in kern_jail.c.
We can't yet eliminate suser() due to some cases in the KAME code where
a privilege check is performed and then used in many different deferred
paths. Do, however, move those prototypes to priv.h.
Reviewed by: csjp
Obtained from: TrustedBSD Project
late stages of unmount). On failure, the vnode is recycled.
Add insmntque1(), to allow for file system specific cleanup when
recycling vnode on failure.
Change getnewvnode() to no longer call insmntque(). Previously,
embryonic vnodes were put onto the list of vnode belonging to a file
system, which is unsafe for a file system marked MPSAFE.
Change vfs_hash_insert() to no longer lock the vnode. The caller now
has that responsibility.
Change most file systems to lock the vnode and call insmntque() or
insmntque1() after a new vnode has been sufficiently setup. Handle
failed insmntque*() calls by propagating errors to callers, possibly
after some file system specific cleanup.
Approved by: re (kensmith)
Reviewed by: kib
In collaboration with: kib
This way we may support multiple structures in v_data vnode field within
one file system without using black magic.
Vnode-to-file-handle should be VOP in the first place, but was made VFS
operation to keep interface as compatible as possible with SUN's VFS.
BTW. Now Solaris also implements vnode-to-file-handle as VOP operation.
VFS_VPTOFH() was left for API backward compatibility, but is marked for
removal before 8.0-RELEASE.
Approved by: mckusick
Discussed with: many (on IRC)
Tested with: ufs, msdosfs, cd9660, nullfs and zfs
