introduce a modified allocation mechanism for mbufs and mbuf clusters; one
which can scale under SMP and which offers the possibility of resource
reclamation to be implemented in the future. Notable advantages:
o Reduce contention for SMP by offering per-CPU pools and locks.
o Better use of data cache due to per-CPU pools.
o Much less code cache pollution due to excessively large allocation macros.
o Framework for `grouping' objects from same page together so as to be able
to possibly free wired-down pages back to the system if they are no longer
needed by the network stacks.
Additional things changed with this addition:
- Moved some mbuf specific declarations and initializations from
sys/conf/param.c into mbuf-specific code where they belong.
- m_getclr() has been renamed to m_get_clrd() because the old name is really
confusing. m_getclr() HAS been preserved though and is defined to the new
name. No tree sweep has been done "to change the interface," as the old
name will continue to be supported and is not depracated. The change was
merely done because m_getclr() sounds too much like "m_get a cluster."
- TEMPORARILY disabled mbtypes statistics displaying in netstat(1) and
systat(1) (see TODO below).
- Fixed systat(1) to display number of "free mbufs" based on new per-CPU
stat structures.
- Fixed netstat(1) to display new per-CPU stats based on sysctl-exported
per-CPU stat structures. All infos are fetched via sysctl.
TODO (in order of priority):
- Re-enable mbtypes statistics in both netstat(1) and systat(1) after
introducing an SMP friendly way to collect the mbtypes stats under the
already introduced per-CPU locks (i.e. hopefully don't use atomic() - it
seems too costly for a mere stat update, especially when other locks are
already present).
- Optionally have systat(1) display not only "total free mbufs" but also
"total free mbufs per CPU pool."
- Fix minor length-fetching issues in netstat(1) related to recently
re-enabled option to read mbuf stats from a core file.
- Move reference counters at least for mbuf clusters into an unused portion
of the cluster itself, to save space and need to allocate a counter.
- Look into introducing resource freeing possibly from a kproc.
Reviewed by (in parts): jlemon, jake, silby, terry
Tested by: jlemon (Intel & Alpha), mjacob (Intel & Alpha)
Preliminary performance measurements: jlemon (and me, obviously)
URL: http://people.freebsd.org/~bmilekic/mb_alloc/
processes a little earlier to avoid a deadlock. Second, when calculating
the 'largest process' do not just count RSS. Instead count the RSS + SWAP
used by the process. Without this the code tended to kill small
inconsequential processes like, oh, sshd, rather then one of the many
'eatmem 200MB' I run on a whim :-). This fix has been extensively tested on
-stable and somewhat tested on -current and will be MFCd in a few days.
Shamed into fixing this by: ps
canonical: define a versioned struct xswdev, and add a sysctl node
handler that allows the user to get this structure for a certain device
index by specifying this index as last element of the MIB.
This new node handler, vm.swap_info, replaces the old vm.nswapdev
and vm.swapdevX.* (where X was the index) sysctls.
- move the sysctl code to kern_intr.c
- do not use INTRCNT_COUNT, but rather eintrcnt - intrcnt to determine
the length of the intrcnt array
- move the declarations of intrnames, eintrnames, intrcnt and eintrcnt
from machine-dependent include files to sys/interrupt.h
- remove the hw.nintr sysctl, it is not needed.
- fix various style bugs
Requested by: bde
Reviewed by: bde (some time ago)
Tor created a while ago, removes the raw I/O piece (that has cache coherency
problems), and adds a buffer cache / VM freeing piece.
Essentially this patch causes O_DIRECT I/O to not be left in the cache, but
does not prevent it from going through the cache, hence the 80%. For
the last 20% we need a method by which the I/O can be issued directly to
buffer supplied by the user process and bypass the buffer cache entirely,
but still maintain cache coherency.
I also have the code working under -stable but the changes made to sys/file.h
may not be MFCable, so an MFC is not on the table yet.
Submitted by: tegge, dillon
- Assert Giant in vm_pageout_scan() for the vnode hacking that it does.
- Don't hold vm_mtx around vget() or vput().
- Lock Giant when calling vm_pageout_scan() from the pagedaemon. Also,
lock curproc while setting the P_BUFEXHAUST flag.
- For now we still hold Giant for all of the vm_daemon. When process
limits are locked we will be only need Giant for swapout_procs().
- Restore the previous order of setting up a new vm_object. The previous
had a small bug where we zero'd out the flags after we set the
OBJ_ONEMAPPING flag.
- Add several asserts of vm_mtx.
- Assert Giant is held rather than locking and unlocking it in a few
places.
- Add in some #ifdef objlocks code to lock individual vm objects when
vm objects each have their own lock someday.
- Don't bother acquiring the allproc lock for a ddb command. If DDB
blocked on the lock, that would be worse than having an inconsistent
allproc list.
- Add a few KTR tracepoints to track the addition and removal of
vm_map_entry's and the creation adn free'ing of vmspace's.
- Adjust a few portions of code so that we update the process' vmspace
pointer to its new vmspace before freeing the old vmspace.
- Don't lock Giant in the scheduler() function except for when calling
faultin().
- In swapout_procs(), lock the VM before the proccess to avoid a lock order
violation.
- In swapout_procs(), release the allproc lock before calling swapout().
We restart the process scan after swapping out a process.
- In swapout_procs(), un #if 0 the code to bump the vmspace reference count
and lock the process' vm structures. This bug was introduced by me and
could result in the vmspace being free'd out from under a running
process.
- Fix an old bug where the vmspace reference count was not free'd if we
failed the swap_idle_threshold2 test.
acquired.
- Assert Giant is held in the strategy, getpages, and putpages methods and
the getchainbuf, flushchainbuf, and waitchainbuf functions.
- Always call flushchainbuf() w/o the VM lock.
vnodes.
- Fix an old bug that would leak a reference to a fd if the vnode being
mmap'd wasn't of type VREG or VCHR.
- Lock Giant in vm_mmap() around calls into the VM that can call into
pager routines that need Giant or into other VM routines that need
Giant.
- Replace code that used a goto to jump around the else branch of a test
to use an else branch instead.
vm_mtx does not recurse and is required for most low level
vm operations.
faults can not be taken without holding Giant.
Memory subsystems can now call the base page allocators safely.
Almost all atomic ops were removed as they are covered under the
vm mutex.
Alpha and ia64 now need to catch up to i386's trap handlers.
FFS and NFS have been tested, other filesystems will need minor
changes (grabbing the vm lock when twiddling page properties).
Reviewed (partially) by: jake, jhb
wakeup proc0 by hand to enforce the timeout.
- When swapping out a process, keep the process locked via the proc lock
from the first checks up until we clear PS_INMEM and set PS_SWAPPING in
swapout(). The swapout() function now must be called with the proc lock
held and releases it before returning.
- Comment out the code to attempt to lock a process' VM structures before
swapping out. It is broken in that it releases the lock after obtaining
it. If it does grab the lock, it needs to hand it off to swapout()
instead of releasing it. This can be revisisted when the VM is locked
as this is a valid test to perform. It also causes a lock order reversal
for the time being, which is the immediate cause for temporarily
disabling it.
the process in question locked as soon as we find it and determine it to
be eligible until we actually kill it. To avoid deadlock, we don't block
on the process lock but skip any process that is already locked during our
search.
- Don't hold Giant in the swapper daemon while we walk the list of
processes looking for a process to swap back in.
- Don't bother grabbing the sched_lock while checking a process' sleep
time in swapout_procs() to ensure that a process has been idle for at
least swap_idle_threshold2 before swapping it out. If we lose the race
we just let a process stay in memory until the next call of
swapout_procs().
- Remove some unneeded spl's, sched_lock does all the locking needed in
this case.
other "system" header files.
Also help the deprecation of lockmgr.h by making it a sub-include of
sys/lock.h and removing sys/lockmgr.h form kernel .c files.
Sort sys/*.h includes where possible in affected files.
OK'ed by: bde (with reservations)
The zone allocator's locks should be leaflocks, meaning that they
should never be held when entering into another subsystem, however
the sysctl grabs the zone global mutex and individual zone mutexes
while holding the lock it calls SYSCTL_OUT which recurses into the
VM subsystem in order to wire user memory to do a safe copy. This
can block and cause lock order reversals.
To fix this:
lock zone global.
get a count of the number of zones.
unlock global.
allocate temporary storage.
format and SYSCTL_OUT the banner.
lock global.
traverse list.
make sure we haven't looped more than the initial count taken
to avoid overflowing the allocated buffer.
lock each nodes.
read values and format into buffer.
unlock individual node.
unlock global.
format and SYSCTL_OUT the rest of the data.
free storage.
return.
Other problems included not checking for errors when doing sysctl out
of the column header. Fixed.
Inconsistant termination of the copied string. Fixed.
Objected to by: des (for not using sbuf)
Since the output is not variable length and I'm actually over
allocating signifigantly and I'd like to get this fixed now, I'll
work on the sbuf convertion at a later date. I would not object
to someone else taking it upon themselves to convert it to sbuf.
I hold no MAINTIANER rights to this code (for now).
Protect pager object list manipulation with a mutex.
It doesn't look possible to combine them under a single sx lock because
creation may block and we can't have the object list manipulation block
on anything other than a mutex because of interrupt requests.
VOP_BWRITE() was a hack which made it possible for NFS client
side to use struct buf with non-bio backing.
This patch takes a more general approach and adds a bp->b_op
vector where more methods can be added.
The success of this patch depends on bp->b_op being initialized
all relevant places for some value of "relevant" which is not
easy to determine. For now the buffers have grown a b_magic
element which will make such issues a tiny bit easier to debug.
programs. There is a case during a fork() which can cause a deadlock.
From Tor -
The workaround that consists of setting a flag in the vm map that
indicates that a fork is in progress and using that mark in the page
fault handling to force a revalidation failure. That change will only
affect (pessimize) page fault handling during fork for threaded
(linuxthreads style) applications and applications using aio_*().
Submited by: tegge
call is correct, but it interferes with the massive hack called
vm_map_growstack(). The call will be returned after our stack handling
code is fixed.
Reported by: tegge
reference count was transferred to the new object, but both the
new and the old map entries had pointers to the new object.
Correct this by transferring the second reference.
This fixes a panic that can occur when mmap(2) is used with the
MAP_INHERIT flag.
PR: i386/25603
Reviewed by: dillon, alc
supported architectures such as the alpha. This allows us to save
on kernel virtual address space, TLB entries, and (on the ia64) VHPT
entries. pmap_map() now modifies the passed in virtual address on
architectures that do not support direct-mapped segments to point to
the next available virtual address. It also returns the actual
address that the request was mapped to.
- On the IA64 don't use a special zone of PV entries needed for early
calls to pmap_kenter() during pmap_init(). This gets us in trouble
because we end up trying to use the zone allocator before it is
initialized. Instead, with the pmap_map() change, the number of needed
PV entries is small enough that we can get by with a static pool that is
used until pmap_init() is complete.
Submitted by: dfr
Debugging help: peter
Tested by: me
of memory, rather than from the start.
This fixes problems allocating bouncebuffers on alphas where there is only
1 chunk of memory (unlike PCs where there is generally at least one small
chunk and a large chunk). Having 1 chunk had been fatal, because these
structures take over 13MB on a machine with 1GB of ram. This doesn't leave
much room for other structures and bounce buffers if they're at the front.
Reviewed by: dfr, anderson@cs.duke.edu, silence on -arch
Tested by: Yoriaki FUJIMORI <fujimori@grafin.fujimori.cache.waseda.ac.jp>
make sure that PG_NOSYNC is properly set. Previously we only set it
for a write-fault, but this can occur on a read-fault too.
(will be MFCd prior to 4.3 freeze)
this information via the vm.nswapdev sysctl (number of swap areas)
and vm.swapdevX nodes (where X is the device), which contain the MIBs
dev, blocks, used, and flags. These changes are required to allow
top and other userland swap-monitoring utilities to run without
setgid kmem.
Submitted by: Thomas Moestl <tmoestl@gmx.net>
Reviewed by: freebsd-audit
- All processes go into the same array of queues, with different
scheduling classes using different portions of the array. This
allows user processes to have their priorities propogated up into
interrupt thread range if need be.
- I chose 64 run queues as an arbitrary number that is greater than
32. We used to have 4 separate arrays of 32 queues each, so this
may not be optimal. The new run queue code was written with this
in mind; changing the number of run queues only requires changing
constants in runq.h and adjusting the priority levels.
- The new run queue code takes the run queue as a parameter. This
is intended to be used to create per-cpu run queues. Implement
wrappers for compatibility with the old interface which pass in
the global run queue structure.
- Group the priority level, user priority, native priority (before
propogation) and the scheduling class into a struct priority.
- Change any hard coded priority levels that I found to use
symbolic constants (TTIPRI and TTOPRI).
- Remove the curpriority global variable and use that of curproc.
This was used to detect when a process' priority had lowered and
it should yield. We now effectively yield on every interrupt.
- Activate propogate_priority(). It should now have the desired
effect without needing to also propogate the scheduling class.
- Temporarily comment out the call to vm_page_zero_idle() in the
idle loop. It interfered with propogate_priority() because
the idle process needed to do a non-blocking acquire of Giant
and then other processes would try to propogate their priority
onto it. The idle process should not do anything except idle.
vm_page_zero_idle() will return in the form of an idle priority
kernel thread which is woken up at apprioriate times by the vm
system.
- Update struct kinfo_proc to the new priority interface. Deliberately
change its size by adjusting the spare fields. It remained the same
size, but the layout has changed, so userland processes that use it
would parse the data incorrectly. The size constraint should really
be changed to an arbitrary version number. Also add a debug.sizeof
sysctl node for struct kinfo_proc.
mtx_enter(lock, type) becomes:
mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks)
mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized)
similarily, for releasing a lock, we now have:
mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN.
We change the caller interface for the two different types of locks
because the semantics are entirely different for each case, and this
makes it explicitly clear and, at the same time, it rids us of the
extra `type' argument.
The enter->lock and exit->unlock change has been made with the idea
that we're "locking data" and not "entering locked code" in mind.
Further, remove all additional "flags" previously passed to the
lock acquire/release routines with the exception of two:
MTX_QUIET and MTX_NOSWITCH
The functionality of these flags is preserved and they can be passed
to the lock/unlock routines by calling the corresponding wrappers:
mtx_{lock, unlock}_flags(lock, flag(s)) and
mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN
locks, respectively.
Re-inline some lock acq/rel code; in the sleep lock case, we only
inline the _obtain_lock()s in order to ensure that the inlined code
fits into a cache line. In the spin lock case, we inline recursion and
actually only perform a function call if we need to spin. This change
has been made with the idea that we generally tend to avoid spin locks
and that also the spin locks that we do have and are heavily used
(i.e. sched_lock) do recurse, and therefore in an effort to reduce
function call overhead for some architectures (such as alpha), we
inline recursion for this case.
Create a new malloc type for the witness code and retire from using
the M_DEV type. The new type is called M_WITNESS and is only declared
if WITNESS is enabled.
Begin cleaning up some machdep/mutex.h code - specifically updated the
"optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN
and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently
need those.
Finally, caught up to the interface changes in all sys code.
Contributors: jake, jhb, jasone (in no particular order)
by myself. It solves a serious vm_map corruption problem that can occur
with the buffer cache when block sizes > 64K are used. This code has been
heavily tested in -stable but only tested somewhat on -current. An MFC
will occur in a few days. My additions include the vm_map_simplify_entry()
and minor buffer cache boundry case fix.
Make the buffer cache use a system map for buffer cache KVM rather then a
normal map.
Ensure that VM objects are not allocated for system maps. There were cases
where a buffer map could wind up with a backing VM object -- normally
harmless, but this could also result in the buffer cache blocking in places
where it assumes no blocking will occur, possibly resulting in corrupted
maps.
Fix a minor boundry case in the buffer cache size limit is reached that
could result in non-optimal code.
Add vm_map_simplify_entry() calls to prevent 'creeping proliferation'
of vm_map_entry's in the buffer cache's vm_map. Previously only a simple
linear optimization was made. (The buffer vm_map typically has only a
handful of vm_map_entry's. This stabilizes it at that level permanently).
PR: 20609
Submitted by: (Tor Egge) tegge
- Proc locking in a few places.
- faultin() now must be called with the proc lock held.
- Split up swappable() into a couple of tests so that it can be locke in
swapout_procs().
- Use queue macros.
- replace the simplelock in struct vm_zone with a mutex.
- use a proper SLIST rather than a hand-rolled job for the zone list.
- add a subsystem lock that protects the zone list and the statistics
counters.
- merge _zalloc() into zalloc() and _zfree() into zfree(), and
move them below _zget() so there's no need for a prototype.
- add two initialization functions: one which initializes the
subsystem mutex and the zone list, and one that currently doesn't
do anything.
- zap zerror(); use KASSERTs instead.
- dike out half of sysctl_vm_zone(), which was mostly trying to do
manually what the snprintf() call could do better.
Reviewed by: jhb, jasone
in 4.2-REL which I ripped out in -stable and -current when implementing the
low-memory handling solution. However, maxlaunder turns out to be the saving
grace in certain very heavily loaded systems (e.g. newsreader box). The new
algorithm limits the number of pages laundered in the first pageout daemon
pass. If that is not sufficient then suceessive will be run without any
limit.
Write I/O is now pipelined using two sysctls, vfs.lorunningspace and
vfs.hirunningspace. This prevents excessive buffered writes in the
disk queues which cause long (multi-second) delays for reads. It leads
to more stable (less jerky) and generally faster I/O streaming to disk
by allowing required read ops (e.g. for indirect blocks and such) to occur
without interrupting the write stream, amoung other things.
NOTE: eventually, filesystem write I/O pipelining needs to be done on a
per-device basis. At the moment it is globalized.
The fix works by reverting the ordering of free memory so that the
chances of contig_malloc() succeeding increases.
PR: 23291
Submitted by: Andrew Atrens <atrens@nortel.ca>
struct swblock entries by dividing the number of the entries by 2
until the swap metadata fits.
- Reject swapon(2) upon failure of swap_zone allocation.
This is just a temporary fix. Better solutions include:
(suggested by: dillon)
o reserving swap in SWAP_META_PAGES chunks, and
o swapping the swblock structures themselves.
Reviewed by: alfred, dillon
of explicit calls to lockmgr. Also provides macros for the flags
pased to specify shared, exclusive or release which map to the
lockmgr flags. This is so that the use of lockmgr can be easily
replaced with optimized reader-writer locks.
- Add some locking that I missed the first time.
cases with file fragments and read-write mmap's can lead to a situation
where a VM page has odd dirty bits, e.g. 0xFC - due to being dirtied by
an mmap and only the fragment (representing a non-page-aligned end of
file) synced via a filesystem buffer. A correct solution that
guarentees consistent m->dirty for the file EOF case is being
worked on. In the mean time we can't be so conservative in the
KASSERT.
Backout the previous delta (rev 1.4), it didn't make any difference.
If the requested handle is NULL then don't add it to the list of
objects, to be found by handle.
The problem is that when asking for a NULL handle you are implying
you want a new object. Because objects with NULL handles were
being added to the list, any further requests for phys backed
objects with NULL handles would return a reference to the initial
NULL handle object after finding it on the list.
Basically one couldn't have more than one phys backed object without
a handle in the entire system without this fix. If you did more
than one shared memory allocation using the phys pager it would
give you your initial allocation again.
Removed most of the hacks that were trying to deal with low-memory
situations prior to now.
The new code is based on the concept that I/O must be able to function in
a low memory situation. All major modules related to I/O (except
networking) have been adjusted to allow allocation out of the system
reserve memory pool. These modules now detect a low memory situation but
rather then block they instead continue to operate, then return resources
to the memory pool instead of cache them or leave them wired.
Code has been added to stall in a low-memory situation prior to a vnode
being locked.
Thus situations where a process blocks in a low-memory condition while
holding a locked vnode have been reduced to near nothing. Not only will
I/O continue to operate, but many prior deadlock conditions simply no
longer exist.
Implement a number of VFS/BIO fixes
(found by Ian): in biodone(), bogus-page replacement code, the loop
was not properly incrementing loop variables prior to a continue
statement. We do not believe this code can be hit anyway but we
aren't taking any chances. We'll turn the whole section into a
panic (as it already is in brelse()) after the release is rolled.
In biodone(), the foff calculation was incorrectly
clamped to the iosize, causing the wrong foff to be calculated
for pages in the case of an I/O error or biodone() called without
initiating I/O. The problem always caused a panic before. Now it
doesn't. The problem is mainly an issue with NFS.
Fixed casts for ~PAGE_MASK. This code worked properly before only
because the calculations use signed arithmatic. Better to properly
extend PAGE_MASK first before inverting it for the 64 bit masking
op.
In brelse(), the bogus_page fixup code was improperly throwing
away the original contents of 'm' when it did the j-loop to
fix the bogus pages. The result was that it would potentially
invalidate parts of the *WRONG* page(!), leading to corruption.
There may still be cases where a background bitmap write is
being duplicated, causing potential corruption. We have identified
a potentially serious bug related to this but the fix is still TBD.
So instead this patch contains a KASSERT to detect the problem
and panic the machine rather then continue to corrupt the filesystem.
The problem does not occur very often.. it is very hard to
reproduce, and it may or may not be the cause of the corruption
people have reported.
Review by: (VFS/BIO: mckusick, Ian Dowse <iedowse@maths.tcd.ie>)
Testing by: (VM/Deadlock) Paul Saab <ps@yahoo-inc.com>
