conf/param.c, and move the initialisation of the sb_max
variable from kern/uipc_socket2.c to conf/param.c. Now
everthing that includes sys/socketvar.h doesn't get
recompiled when SB_MAX's value changes.
The problem is that if "sl" is a symbolic link, a lookup on "sl/"
will be flagged as the last component. Thus VOP_LOOKUP will lock
the parent directory if LOCKPARENT is set. In order for the symbolic
link to be resolved, this lock needs to be released. namei() would
test for this by checking if ni_pathlen == 1, which it wouldn't as
"/" is left in the name, and namei() would not unlock the parent.
The next call to lookup() to resolve the symbolic link would fail
as the parent was still locked.
initialized). This lock also protects the "next drain candidate" pointer.
XXX There is still one locking protocol problem, which should not be
a problem in practice, but is still marked as an issue in the code anyhow.
a data structure after it was freed. This wasn't actually a problem,
and the change caused the wrong pool_item_header to be freed
in the non-PR_PHINPAGE case.
Call configure() directly immediately after config_init().
This causes autoconfiguration to happen at the same time as before, but
creates some kernel submaps earlier, so that e.g. mbinit() can now
allocate memory.
- Protect userspace from unnecessary header inclusions (as noted on
current-users).
- Some const poisioning.
- GREATLY simplify the locking protocol, and fix potential deadlock
scenarios. In particular, assume that the back-end page allocator
provides its own locking mechanism (this is currently true for all
such allocators in the NetBSD kernel). Doing so allows us to simply
use one spin lock for serialized access to all r/w members of the pool
descriptor. The spin lock is released before calling the back-end
allocator, and re-acquired upon return from it.
- Fix a problem in pr_rmpage() where a data structure was referenced
after it was freed.
- Minor tweak to page manaement. Migrate both idle and empty pages
to the end of the page list. As soon as a page becomes un-empty
(by a pool_put()), place it at the head of the page list, and set
curpage to point to it. This reduces fragmentation as well as the
time required to find a non-empty page as soon as curpage becomes
empty again.
- Use mono_time throughout, and protect access to it w/ splclock().
- In pool_reclaim(), if freeing an idle page would reduce the number
of allocatable items to below the low water mark, don't.
NMBCLUSTERS for the mbuf cluster pool. On platforms which use direct-mapped
segments for pool pages (MIPS and Alpha), this makes NMBCLUSTERS actually
meaningful (such ports don't even allocate mb_map, as it is not used to
map mbuf cluster pages).
Improve the message logged at a maximum rate of once per second. The
new message: "WARNING: mclpool limit reached; increase NMBCLUSTERS".
In the back-end pool page allocator, remove the message about mb_map
being full. The message was not necessarily correct as the allocator
may have been starved for pages, rather than for space in the map. Also,
the hard limit on the mbuf cluster pool will be reached before the map
fills (the last cluster will always fit into the map), so the message
is redundant.
Add a comment in mbinit() about considering setting low water marks on
the mbuf and mbuf cluster pools.
- Add support for hard limits, with optional rate-limited logging of
a warning message when the pool limit is reached. (This will be used
to fix a bug in mbuf cluster allocation on the MIPS and Alpha ports.)
- Fix some locking protocol errors. This required splitting pr_flags
into pr_flags (which is protected by the spin lock) and pr_roflags (which
are `read only' flags, set when the pool is initialized, and never changed
again; these do not need to be protected by a mutex).
- Make the low water support actually mean something. When a low water
mark is set, add free items to the pool until the low water mark is
reached. When an item allocation causes the number of free items to
drop below the low water mark, make the pool catch up to it. This can
make the pool allocator more useful for several applications (e.g.
pmap `pv entry' management) and more robust for others (for e.g. mbuf
and mbuf cluster allocation, so that the pagedaemon can use NFS to clean
pages on diskless systems without completely running dry on buffers to
receive packets in during extreme memory shoratages).
- Add a comment where we sleep waiting for more pages for the back-end
page allocator. Specifically, instead of sleeping potentially forever,
perhaps we should just wake up once a second to try allocating a page
again. XXX Revisit this soon.
and system call now just return EFAULT). A complete fix will
presumably have to wait for UBC and/or for vnode locking protocols to
be revamped to allow use of shared locks.
same uid or by root.
This code is from FreeBSD. (Whilst it was originally obtained from OpenBSD,
FreeBSD fixed it to work with multicast. To quote the commit message:
- Don't bother checking for conflicting sockets if we're binding to a
multicast address.
- Don't return an error if we're binding to INADDR_ANY, the conflicting
socket is bound to INADDR_ANY, and the conflicting socket has
SO_REUSEPORT set.
)
directories which aren't under the recipient's root.
Clean up of many error conditions involving descriptor passing, to
eliminate infinite loops, panics, premature garbage collection of
sockets, and descriptor leaks:
- Avoid letting unp_gc() see descriptors with a refcount of zero by
removing them from the socket's queue before releasing them.
- Avoid socket leak in PRU_ABORT (this will also gc descriptors queued
on a not-yet accepted socket when the accepting socket goes away).
- Put in block comment explaining how unp_gc() should work.
- Correctly manage unp_defer count so we don't get stuck in an infinite
loop with nothing to do.
- Don't tie MARK and DEFER bits so closely together.
- Mark descriptors queued on not-yet-accepted sockets as well.
- Don't call sorflush on non-socket, it doesn't work very well.
- Deal with discard of NULL file pointer.
- Hopefully cause GC to converge faster by only deferring sockets in
unp_mark().
limit into account when checking against the limit; fdp->fd_nfiles may
be greater than the current descriptor limit, and there may be space
in fdp->fd_ofiles beyond the limit. If we say it's available,
unp_externalize will get confused and panic when fdalloc fails.
deadlock in VOP_FSYNC() if the unreferenced vnode picked for
reclamation happened to be stacked on top of a vnode the process
already had locked. This could happen if the same filesystem was
accessed both through a union mount and directly; it seemed to happen
most frequently when the direct access was through NFS.
Avoid this deadlock by changing vinvalbuf to pass a new FSYNC_RECLAIM
flag bit to VOP_FSYNC() to indicate that a reclaim is in progress and
only a `shallow' fsync is necessary.
Do nothing in *_fsync() in umapfs, nullfs, and unionfs when
FSYNC_RECLAIM is set; the underlying vnodes will shortly be released
in *_reclaim and may be reclaimed (and fsync'ed) later.
was changes to comments only, but..)
Build vfs_getcwd.c as standard part of kernel.
Add implementation of fchroot(), since two emulations already had it.
Call vn_isunder() in fchdir(), chroot(), and fchroot() to make it harder
to escape chroot().
Add kernel implementation of getcwd() which uses this cache, falling
back to reading the filesystem on a cache miss.
Along for the ride: add new VOP_FSYNC flag FSYNC_RECLAIM indicating
that a reclaim is being done, so only a "shallow" fsync is needed.
some things (e.g. unionfs) may depend on it. It's currently ok
for vnodecovered to be set already; it's not for v_mountedhere in
the vnode, though.
From John Darrow.
XXX should probably just extend VFS_MOUNT to take the vnode pointer as
an argument.
since the lock may be taken again. This was the intention of the CANRECURSE
lock already there, but didn't work.
Only fill in the vnode<->mountpoint links (mountedhere and vnodecovered)
after VFS_MOUNT returned succesfully. It might happen that something called
from VFS_MOUNT mistook the vnode for an already successfully mounted on
one because of this.
could be done in one of 2 ways:
* call lk_init with LK_CANRECURSE, resulting in a lock that
always can be used recursively.
* call lockmgr with LK_CANRECURSE, meaning that it's ok if this
lock is already held by us.
Sometimes we need a locking type that says: take this lock now, exclusively,
but while I am holding it, I may go through a code path which could attempt
to get the lock again, and which is unaware that the lock might already
be taken. Implement LK_SETRECURSE for this purpose. Assume that locks and
unlocks come in matching pairs (they should), and check for this 'level'
using SETRECURSE locks.
to pass down a locked node. Modify union_copyup() to call VOP_CLOSE
locked nodes.
Also fix a bug in union_copyup() where a lock on the lower vnode would
only be released if VOP_OPEN didn't fail.
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
